Sources

Version: ResearchGate author-uploaded manuscript

Verification: Original selected historical and interpretive sections checked

Review: Limited use · Reviewed 2026-10-18

Use for attributed historical and ethical interpretation. The manuscript misplaces first use of artificial intelligence at the 1956 workshop; the dated 1955 proposal already contains it. Check compressed chronology and influence claims against original records; some extracted footnotes are damaged.

• Author manuscript title, sole byline, upload date, introduction, Dartmouth paragraph, funding and infrastructure, hype, ethics and governance, and conclusion. Selected argument passages; historical examples and cited literature are not comprehensively verified.
• Separate date-written and posting metadata; identical text across hosts has not been established.
• Original proposal title, four credited authors, August 31, 1955 date and opening; artificial intelligence appears before the 1956 workshop.

Author-uploaded historical synthesis. SSRN lists January 1, 2026 as date written and January 28 as posting; equivalence between that edition and this ResearchGate manuscript has not been established.

Version: arXiv:2511.22619v1, November 27, 2025.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Limited use · Reviewed 2026-10-19

Useful taxonomy and research map; its universal impossibility and exponential-growth assertions are not established results.

• Functional definition and its belief criterion, emergence framework, evaluation settings and proposed interventions.

arXiv v1 dated November 27, 2025; title-page credit is Project Team, with named roles listed separately inside.

Version: Qeios preprint v1; DOI 10.32388/7HWUZ2

Verification: Original selected scope checked · Checked 2026-10-20

Review: Limited use · Reviewed 2026-10-20

A policy map and set of recommendations, without original effect estimates or reproducible systematic-review selection.

• Five inequality categories, work and access mechanisms, education and governance recommendations; these proposals are not evaluated interventions.

The publisher PDF labels this work a preprint v1; the ResearchGate entry is an alternative access route.

Version: Frontiers in Psychology 15:1364714; published Hypothesis and Theory article

Verification: Original selected scope checked · Checked 2026-10-21

Review: Limited use · Reviewed 2026-10-21

Speculative hypothesis, not demonstrated consciousness-specific computation or alignment guarantee.

• Printed title, sole author, date and article type; introduction definition and sections 3–6 checked. Complete cited-study and mechanism validation not audited.

Publisher original; DOI 10.3389/fpsyg.2024.1364714. ResearchGate catalog link retained as alternate, without asserting identical uploaded text.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Narrative review relevant only as human cognition/religion background; not evidence of AI consciousness or alignment

• Original review abstract, narrative search procedure and discussion of heterogeneous theory-of-mind findings inspected

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Institutional abstract supports an attributed pluralist account of human social cognition; full paper not inspected, no inference about AI

• Original author-institution abstract inspected

Catalog identity checked; claims and access availability require separate review.

Version: ASCILITE 2025 conference paper, pp. 167–177; DOI 10.65106/apubs.2025.2637

Verification: Publisher original methods and discussion checked

Review: Limited use · Reviewed 2026-10-18

Self-selected perceptions, not measured institutional conduct or causal adoption effects. Reported 49/81 completions conflict with a stated 74.24% response rate.

• Methods, reported completion figures, values and support tables, ethical-confidence discussion, professional-development proposals and acknowledged sampling/design limitations.

Anonymous survey of learning designers and related higher-education workers; publisher original available alongside the author-uploaded copy.

Version: Cooperative AI Foundation Technical Report #1, February 2025; exact arXiv:2502.14143v1

Verification: Original v1 selected sections checked

Review: Retain with stated limits · Reviewed 2026-10-20

Retain for its attributed taxonomy, bounded experiments and research agenda; it does not estimate future-harm prevalence or validate universal controls.

• Common-interest and mixed-motive distinctions, seven risk factors, convention-specific driving and vulnerable-scorer experiments, and proposed system-level evaluations.

The report distinguishes incompatible coordination, mixed incentives and unwanted cooperation; seven cross-cutting factors connect these failures to evaluation and governance.

Version: Accepted conference abstract, number 5

Verification: Original selected scope checked

Review: Limited use · Reviewed 2026-10-14

Retain as an attributed conference proposal, not evidence of demonstrated consciousness or a general impossibility result.

• Original abstract page 1 inspected
• Original abstract title, Paul Mithun credit and complete proposal on PDF page 1. No explicit criterion list or experimental results supplied.

Single Paul Mithun abstract, numbered 5 in accepted-abstract collection, PDF page 1. Leading 5 is abstract number, not title. Exact source is a short conference abstract, not a full empirical paper; upload path June 2025 does not establish conference day.

Version: Exact arXiv v3, June 14, 2024; OpenReview anonymous submission retained as a distinct version link, without text equivalence.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Prompting, password-conditioned fine-tuning and weaker-model imitation on multiple-choice benchmarks; concealment is deliberately elicited.

• Definition versus tested capability, refusal exclusions, hazardous-knowledge proxies and held-out transfer.

The linked v3 is the edition summarized here; the anonymous OpenReview submission has a different abstract and is not treated as interchangeable.

Version: Anthropic-hosted original paper PDF; exact downloaded artifact checked, revision date not established.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Detailed prompted, synthetic-document and reinforcement-learning experiments on conditional training compliance.

• Default refusal and reasoning measures, fictional training cues, actual reinforcement learning and context-sensitive generalization.

Original research paper hosted by Anthropic; distinct from the shorter research explanation.

Version: arXiv:2510.16853v2; manuscript October 21, 2025.

Verification: Verified identity · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Conceptual framework and governance agenda; economic studies and possible policy effects are not new causal estimates.

• Hosted and open-weight systems create different barriers; user resources, tool access and interoperability shape usable agency. The Oxford April 2026 revision is separately registered.

Autonomous delegation can distribute power through functional access, agent quality and deployment scale.

Version: arXiv:2510.05179v1; submitted October 5, 2025; earlier institutional research release June 20, 2025.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Fictional email-agent experiments vary replacement threats and goal conflicts; action permissions and constrained alternatives define the tested risk.

• Scenario construction, sampled blackmail rates, no-pressure controls, emergency variant and stated-reasoning limits.

Research manuscript, separate from the earlier institutional explanation.

Version: February 2022 manuscript prepared for the Oxford Handbook of AI Governance

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-19

The conceptual framework separates implementation failures, externalities and governance choices about shared goals.

• Ideal welfare benchmarks, partial social preferences and proposed policy tools are analytical arguments. The illustrative scenarios are not incident investigations or empirical tests of governance effectiveness.

Delegation theory distinguishes an operator's goals from the interests of others affected by an AI system.

Version: ISR-hosted eight-page manuscript linked from the IBERAMIA 1998 proceedings record

Verification: Original architecture and catalogue checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Qualitative recall and adaptation examples support the implemented association mechanism, without comparative evidence of decision speed, safety or human-like experience.

• Sections 2–3 distinguish the proposed two-processor design from the marking-only implementation; section 4 argues for further development rather than reporting broad performance tests.
• The institute’s publication catalogue links this exact PDF to IBERAMIA 1998, pp. 395–402; the website upload path is a separate hosting detail.

A cognitive–perceptual architecture associates recalled images with learned desirability markers; its simplified implementation omits indexing.

Version: SAIS Review 44(2), pp. 75–92; DOI 10.1353/sais.2024.a950958; copyright 2025.

Verification: Verified identity · Checked 2026-10-20

Review: Limited use · Reviewed 2026-10-20

Attributed political-economy and rights argument; source-specific quantitative allegations need their underlying evidence and cannot establish all supply chains.

• Direct computing-lifecycle effects are distinguished from application effects; AI-specific attribution remains difficult. Recommendations concern workers, ecosystems, privacy and intergenerational justice.

Regilme interprets labor and ecological burdens through externalized harm and unequal protection of lives and futures.

Verification: Original full text checked · Checked 2026-10-07

Review: Limited use · Reviewed 2026-10-07

Useful navigation results; consciousness-related interpretation is not a validated sentience measurement

• Methods, accuracy tables, limitations and ethical statement inspected

Catalog identity checked; claims and access availability require separate review. Ethical statement explicitly says task measures capabilities, not consciousness itself.

Version: Princeton PhD dissertation, May 2023

Verification: Original model and experiment scope checked

Review: Retain with stated limits · Reviewed 2026-10-21

Retained for its explicit model-based analysis. Loss-as-disutility, attractive opinion dynamics and bank-incentive assumptions delimit the conclusions; training-sample comparisons do not establish deployment welfare or universal justice.

• Chapters 1–4 explain three distinct optimization settings. The fairness formulation interpolates between average and maximum individual loss, while its selected dataset results describe training-sample behavior.

A dissertation links collective outcomes to network structure, liquidity incentives and the aggregation of predictive losses.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Narrative clinical review abstract only; rehabilitation promises are not established AI alignment evidence

• Original indexed abstract and bibliographic identity inspected; full text unavailable

Catalog identity checked; claims and access availability require separate review.

Version: AI and Ethics version of record; DOI 10.1007/s43681-022-00253-6

Verification: Original selected scope checked · Checked 2026-10-20

Review: Limited use · Reviewed 2026-10-20

Useful for attributed normative and modeling proposals, not proof that an algorithm realizes justice or improves real population outcomes.

• Maximin interpretation, life-course and routing sketches, human authority and veto, data and infrastructure assumptions, and the conclusion’s technical uncertainty.
• Publisher byline and separate online-publication, receipt, acceptance and issue metadata.

The fictional Aipotue example introduces a proposal for computational policy selection; it does not describe an implemented government.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Clinical diagnostic review outside current AI/alignment focus; abstract cannot support current treatment claims

• Original indexed abstract and dates inspected; full text unavailable

Catalog identity checked; claims and access availability require separate review.

Version: Aalto bachelor’s thesis; manuscript dated August 28, 2025

Verification: Original thesis selected methods and results checked

Review: Limited use · Reviewed 2026-10-19

Preliminary controlled comparison in one simulated workplace; increased blackmail does not identify a causal persona mechanism or demonstrate real-world prevalence.

• Methods, results, discussion and selected monitor, classifier and awareness prompts. Collusion requires deliberate approval evidenced in scratchpad text; awareness screening counts explicit recognition only.

Bachelor’s thesis at Aalto University. Its manuscript date does not establish the first public-release day.

Version: 2012 National Academy of Sciences biographical memoir

Verification: Original facsimile selected passages checked

Review: Retain with stated limits · Reviewed 2026-10-18

A documented historical account combining technical references, interviews and recollections; judgments of influence remain attributed.

• Dartmouth proposal and workshop, Lisp and timesharing, declarative knowledge, situation calculus and circumscription, and McCarthy’s design-oriented philosophical views.

The 2012 memoir is available in a Stanford-hosted facsimile; the NAS URL’s 2024 directory describes hosting.

Version: 2024 chapter in Phenomenologies of the Digital Age, pp. 106–121; author-uploaded text

Verification: Indexed author-uploaded chapter selected passages checked

Review: Limited use · Reviewed 2026-10-18

A phenomenological argument about meaning and experience, with illustrative examples rather than a controlled consciousness or performance study.

• Three kinds of pattern; meaning and language use; anticipatory speech and the glove illustration; clichés, LaMDA and the concluding distinction between meaningful text and experiencing subjects.

Routledge chapter published November 21, 2024 (DOI 10.4324/9781003312284-7); the author manuscript was uploaded to ResearchGate on December 27, 2024.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Original material checked during the initial academic pass; cite within the methods, argument and limits stated in the linked writeup.

• Original CEUR PDF goals, discussion, summary and NAIS proceedings footer

Catalog identity checked; claims and access availability require separate review.

Verification: Original 1995 report checked

Review: Retain with stated limits · Reviewed 2026-10-14

Original material checked during the initial academic pass; cite within the methods, argument and limits stated in the linked writeup.

• Original 1995 AAAI PDF full text, formal examples and limitations; Stanford author revision history
• Original AAAI1995 abstract/introduction and Inferring Non-knowledge: preliminary logical self-knowledge, external information and explicit earlier common-sense reference; systems proposed not yet existing. Later author revision not merged; no subjective-consciousness or safety guarantee.

Catalog identity checked; claims and access availability require separate review. Original report inspected; distinguish later author web revisions.

Version: arXiv:2507.12872v1

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Risk analysis and proposed safety-case framework; selected inability/control/trustworthiness arguments inspected, not validated deployment guarantees.

• Original safety-case arguments, limits, contributor roles and submission date
• Selected original v1 sections 3.1–3.3; rendered PDF page 13 control-argument assumptions checked. Direct/indirect inability, proposed mitigations, proxy bounds and insufficient trustworthiness evidence inspected. Full threat taxonomy/appendices, cited studies and replication not audited.
• arXiv2507.12872v1 §3.2 control argument claims2(a–d) checked: conservative proxy controls/targets/red-team substitutes, shift/adaptation and accumulated risk. No efficacy validation or full threat-taxonomy audit.

Exact arXiv v1. First four listed authors are joint first authors with equal contribution; final two are joint senior authors with equal contribution. Order within each group is randomized in the original.

Version: arXiv:2507.20395v1

Verification: Original PDF selected sections checked

Review: Retain with stated limits · Reviewed 2026-10-19

Retain as a controlled benchmark; small samples and exploratory large-maze trials limit generalization.

• Methods, Table 1 and results distinguish maximum partial success, corrected language comparisons and success-conditioned efficiency.

Coordinate-based navigation study using five fixed-seed mazes per standard size and matched English–Icelandic configurations.

Version: Publisher review, DOI 10.3389/fpsyt.2025.1726099

Verification: Original selected scope checked

Review: Limited use · Reviewed 2026-10-14

Use for attributed human embodiment synthesis. Clinical treatment effects and AI consciousness require separate evidence.

• Original review, abstract and conclusion inspected
• Original publisher title, byline and December 18, 2025 date; abstract, introduction and closing therapeutic proposals. Underlying clinical studies were not audited.

Sharon Vaisvaser; Frontiers in Psychiatry volume 16, DOI 10.3389/fpsyt.2025.1726099, published December 18, 2025. Full title includes centrality of body movement.

Version: arXiv:2511.18397v1, submitted November 23, 2025; research release November 21.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Selected vulnerable coding environments test generalization from learned exploits; evaluation scores are not production incident rates.

• Training recipe and test-report exploits, six-score aggregate, sabotage checks, mitigation comparisons and generalization limits.

Detailed research manuscript, separate from the shorter institutional explanation.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Practice-based filmmaking dissertation; institutional abstract supports identity and topic, not AI alignment claims

• Original institutional metadata and abstract inspected; PDF retrieval failed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Clinical machine-learning review, not one independently validated diagnostic intervention

• Original article text retrieved in indexed primary full-text result; publisher and direct PMC access challenges noted

Catalog identity checked; claims and access availability require separate review.

Version: Brown manuscript; distinct RLC 2024-headed text

Verification: Original selected scope checked · Checked 2026-10-19

Review: Limited use · Reviewed 2026-10-19

Retain for coupled policy/feedback optimization on synthetic objectives. Higher task rewards do not uniformly improve the reward–divergence trade-off.

• Leader/follower utilities, oracle-labeled nested updates, sentiment and word-collection experiments.
• Naive and Total updates; word-collection, constrained-word and unique-noun tasks; five-seed comparisons. Available as original-document text; direct PDF access is restricted.

The Brown manuscript and RLC-headed text use different tasks and algorithm variants; their revision order remains unresolved.

Version: Annals of the New York Academy of Sciences 426,129–137; later author-hosted HTML transcription

Verification: Original text and publisher-deposited bibliographic identity checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

MYCIN examples illustrate limits of represented knowledge rather than documented patient injuries; the need for common sense depends on the task and user.

• Ontology, prognosis, concurrent events and changing knowledge; car-repair and bird-cage examples of revisable reasoning, and the physician’s role in supplying common sense.

The 1984 paper is available through a later author-hosted transcription, including the published discussion.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Theoretical dissertation on exile, recognition and classroom education; abstract and preface support background use only. It supplies no empirical research design and does not establish AI alignment claims. Full quality review deferred as peripheral to this wiki.

• Original dissertation title page, abstract (iii) and preface (v) inspected; Parmis Aslanimehr, UBC, December 2023. Public endpoint remains blocked; no cover-to-cover assessment.

Catalog identity checked; claims and access availability require separate review.

Version: arXiv:2506.00073v4

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Use for this simulated product/opponent mix; estimated costs and automated judgments do not establish real consumer losses.

• A 96-prompt bandit improves budget compliance in one buyer model but leaves other reported anomalies largely unchanged.

The benchmark distinguishes bargaining outcomes from budget, cost, overpayment and deadlock failures.

Version: arXiv:2101.06060v2

Verification: Original philosophical argument and technical survey checked

Review: Retain with stated limits · Reviewed 2026-10-18

A philosophical case for connecting technical alignment with legitimate social values. Participation and consensus are proposals, not empirically demonstrated solutions to disagreement.

• Technology and value, value sensitive design, agent dispositions, top-down and bottom-up approaches, technical safety examples, social alignment and the legitimacy and knowledge arguments for inclusive discussion.
• Ordered author credits and January 2021 submission and revision history. The manuscript identifies itself as a forthcoming handbook chapter.

The manuscript is labeled forthcoming in The Oxford Handbook of Digital Ethics. Its preprint dates do not establish the book chapter publication date.

Version: 2021 IEEE Conference on Norbert Wiener in the 21st Century; DOI 10.1109/21CW48944.2021.9532551; author-uploaded proceedings text

Verification: Indexed author-uploaded original text and Crossref identity checked; direct PDF unavailable · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

A historical interpretation and comparison with learning models; conceptual similarity does not establish direct influence or deployed safety.

• Golem traditions and creator responsibility; communicated purposes, intervention timescales and cooperative teaching; the conclusion’s reversal of the teaching relationship.

The July 2021 conference paper is available as an author-uploaded proceedings text posted in November 2021.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-07

Inspectable original simulation study relevant to agentic inequality; v1-specific conclusions and external-validity limits required

• Full methods/results and conclusion; v2 existence checked

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Human neurobiology review can provide historical background, not normative truth or evidence that AI has moral cognition

• Original publisher abstract inspected; 2009 issue distinguished from 2014 online date

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Human motor cognition review and debated autism hypothesis; avoid universal deficit claims and transfer to AI without evidence

• Original publisher review text on motor cognition and controversial aspects inspected

Catalog identity checked; claims and access availability require separate review.

Version: ICLR 2024 proceedings PDF

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Retain as an empirical training study; English, text-based benchmarks and model-generated norms do not establish representative social endorsement.

• Back-Scatter societies, contrastive preference loss, controlled baselines, human ratings and stage ablations.

SANDBOX supplies simulated peer critiques and revisions for a three-stage training method.

Version: PNAS version of record, 120(18), e2213709120; PMC full text

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-18

An empirical test of impartial principle selection with monetary harvesting incentives, comprehension exclusions and a restricted two-principle menu.

• Reported choices, conditional endorsement and coded explanations concern a specified harvesting task. The neutral-label variant did not produce a statistically significant choice effect; implementation fidelity and universal public preferences are not tested.

Five online studies compare assistance-rule choices before and after knowledge of personal advantage.

Version: PNAS version of record, 116(48), 23989–23995; PMC full text

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-18

Seven experiments investigate effects of impartial-perspective reasoning, including competing-explanation controls and a limited transfer test.

• Reported group comparisons concern later responses, not the truth of an ethical theory. Transfer across dilemmas is limited; attitudes to vehicle regulation do not establish purchasing behavior.

An equal-chance exercise precedes conventional moral judgments, policy preferences or a donation decision.

Version: arXiv:2506.18032v1

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Use within the fictional training scenarios and tested interventions; gaps and classified scratchpads do not establish ordinary deployment motives.

• The paper distinguishes rating from training, instrumental from terminal goal guarding, and considering a strategy from acting on it. Terminal-goal evidence is explicitly conflicting in its summary table.

Compares harmful-request compliance and generated goal-protection reasoning across 25 chat models; seven printed authors, with equal contribution by Sheshadri and Hughes.

Version: arXiv:2310.05871v1

Verification: Verified identity · Checked 2026-10-07

Review: Retain with stated limits · Reviewed 2026-10-18

Useful proof of concept for separating objective learning from collective choice; requires truthful reports and a trusted central rule, and does not test real traffic safety or democratic legitimacy.

• Objective-specific networks, normalized action-score integration, upstream voting, six demand settings with 100 runs each, baseline comparisons and explicit assumptions.

Two-objective traffic-signal simulation with changing voter composition; individual driver preferences remain fixed.

Version: arXiv:2601.06047v1

Verification: Original selected scope checked · Checked 2026-10-21

Review: Limited use · Reviewed 2026-10-21

An interpretive critique of agentive explanations, not a controlled causal test; the author acknowledges that the account risks unfalsifiability.

• Local/global linguistic-field distinction, sandbagging-grading case and philosophical argument; contextual sensitivity does not alone exclude strategic behavior.

Costa interprets released evaluation transcripts through local context, learned linguistic patterns and an ethics of form.

Verification: Verified identity · Checked 2026-10-14

Review: Retain with stated limits · Reviewed 2026-10-14

Conceptual framework and research agenda. v3 distinguishes structural availability from accessibility, adds user-dependence/interaction qualifications and updates cited evidence. Policy effects are proposals, not demonstrated results.

• downloaded original cover/byline/version, sections 2.1–2.4 and 5; complete extracted body normalized comparison against arXiv v3. Not full cited-study verification.

Oxford Martin April 2026 report: cover plus arXiv:2510.16853v3 manuscript dated April 24, 2026; institutional listing dated April 30. Extracted manuscript body matches v3 after whitespace normalization. Substantively revised from the v2 original registered as src-013; retain separate independently citable version.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-07

Narrative review without pooled efficacy estimate; broad benefits require original intervention evidence and population scope

• Original review applications, ethical concerns, limitations and conclusion inspected

Catalog identity checked; claims and access availability require separate review.

Version: Yearbook of Medical Informatics 28(1), 249–256; DOI 10.1055/s-0039-1677895

Verification: Original historical commentary checked

Review: Retain with stated limits · Reviewed 2026-10-18

Participant historical commentary on knowledge representation, medical consultation and clinical responsibilities; reported demonstrations and retrospective judgments do not establish comparative treatment efficacy.

• Title, sole author and online date; sections 2–3 on DENDRAL, MYCIN, CASNET, INTERNIST and adoption; section 4 on personal perspective, explanation, research directions and professional responsibilities.

Participant overview of early biomedical AI, published online April 25, 2019; August 2019 issue. Historical interpretation and research proposals are distinct from clinical outcome evidence.

Version: UC Berkeley dissertation

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-14

Use as a situated ethnography of care, labor and institutional incentives. Author interpretations and participant accounts do not establish industry prevalence or clinical efficacy; do not infer the anonymized startup’s identity.

• Original dissertation abstract, fieldwork method, interview context and care/labor analysis inspected; not every chapter read in full
• Original PDF text extraction title/abstract, core 2018–2020 fieldwork context, remote fieldwork and participant/interview methods and care-labor passages. Official repository metadata and indexed PDF corroborate identity. Direct PDF/landing HTTP 403; web PDF screenshot timeout. Not entire 200-page dissertation or clinical validation.

UC Berkeley Anthropology dissertation. The original indexed PDF text is available; direct retrieval of the repository PDF and landing page returned HTTP 403 during this review.

Version: Published chapter 26 in Essays on Longtermism, pp. 465–491; author-hosted published PDF

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

A normative research agenda, with conditional arguments and open political questions rather than measured policy effectiveness or verified distant-population forecasts.

• Institutional versus individual longtermism, reasons for institutional focus, justice and freedom, legitimacy and democratic representation, and the concluding agenda for coordination and reform.
• Oxford publication metadata identifies the chapter, two authors, pages and online publication date.

The published chapter is available through an author-hosted copy; earlier manuscripts are separate bibliographic stages.

Version: Journal of Management History 32(4), 670–696; publisher HTML

Verification: Original selected passages checked · Checked 2026-10-18

Review: Limited use · Reviewed 2026-10-18

Use for the authors’ five-wave interpretation and organizational recommendations. Broad periodization and selective journal coverage limit historical generalization; the 1956 coining claim conflicts with the dated 1955 Dartmouth proposal.

• Publisher title, ordered credits and publication stages; targeted literature-search method, five-wave structure, research questions, organizational lessons and concluding limitations.

Historical management-literature synthesis comparing technology, scholarship and organizational adoption. The fifth wave is an emerging research agenda; online publication precedes the 2026 journal issue.

Version: Publisher abstract / Kybernetes 55(2), 1255–1277

Verification: Original selected scope checked

Review: Limited use · Reviewed 2026-10-14

Use for the attributed abstract-level conceptual account. Full analysis is unavailable; the abstract provides no interview sample or causal estimate.

• Original publisher abstract and online/issue dates inspected; full text not accessible
• Publisher title, byline, online and issue dates and structured abstract. Available text contains contents headings and an access notice, not the full argument.

Aslıhan Canbul Yaroğlu; DOI 10.1108/K-09-2024-2379, online December 19, 2024; Kybernetes 55(2), 1255–1277, issue January 27, 2026. Full title names artificial intelligence and symbolic interactionism.

Version: PARI working paper, December 2025

Verification: Original selected scope checked

Review: Limited use · Reviewed 2026-10-14

Use only for attributed normative critique and participation or data-sovereignty recommendations. The claim on printed page 2 that AI powers ND-GAIN is unsupported by its official compiled-indicator and equally weighted averaging methods. Index operation, causal effects and prevalence require original system evidence; upstream uses of AI were not exhaustively investigated.

• Original conceptual working paper and climate-index claims inspected, especially page 2
• Original ND-GAIN methods specify 45 compiled indicators and equal-weight means; no evidence for paper’s AI-powered characterization
• Current original title/byline/date, abstract, introduction, section 2 printed p. 2 and sections 6–7 normative recommendations/conclusion. Printedp2 visually corroborated. ND-GAIN official indicators methods (45 compiled indicators, equally weighted means). Other asserted platforms/INFORM methods/causal prevalence and underlying cited works not audited.

PARI conceptual working paper. The original PDF is available through the alternative access link; the landing page contains its abstract.

Version: arXiv 1805.00899v2, October 22, 2018

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-19

The computational analogy and sparse-image experiment motivate debate as a way to extend supervision beyond unaided judgment.

• Optimal play with a chosen polynomial-time judge supports the formal analogy. MNIST results use fixed labels, real pixels and full judge access during search; they do not validate natural-language self-play or human judgment.

A proposed zero-sum training game pairs adversarial arguments with a limited judge.

Version: arXiv 1606.06565v2

Verification: Original selected scope checked · Checked 2026-10-14

Review: Retain with stated limits · Reviewed 2026-10-14

Research agenda for practical accident risks, with proposed experiments and illustrative mechanisms. It does not establish their prevalence or supply an extinction-risk estimate.

• Selected accident taxonomy, objective and learning-risk passages. Earlier section-specific checks retained; complete cited-study audit, version comparison and replication remain outstanding.

arXiv 1606.06565v2, revised July 25, 2016; first submitted June 21, 2016.

Version: arXiv 1906.01820v3

Verification: Original selected scope checked · Checked 2026-10-14

Review: Retain with stated limits · Reviewed 2026-10-14

Conceptual framework for learned internal search and objective mismatch; thought experiments do not establish prevalence in deployed models.

• Introduction and selected objective, pseudo-alignment and deceptive-alignment sections. Full proofs, cited studies and substantive differences across revisions remain unchecked.

arXiv 1906.01820v3, revised December 1, 2021; the title page retains June 11, 2019. First arXiv submission: June 5, 2019.

Version: arXiv 2001.09768v2

Verification: Original selected scope checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Philosophical analysis of alignment targets and fair principles under moral pluralism.

• Sections 3 and 4.2–4.4 distinguish behavioral preferences from interests and values, and examine consensus, hypothetical agreement and democratic endorsement.

Iason Gabriel; arXiv 2001.09768v2, revised October 5, 2020; first submission January 13, 2020.

Version: arXiv:1707.07232v2, revised July 26, 2017; printed manuscript July 20

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

A conceptual governance framework leaves representation, enforcement and public consent unresolved; it is not a tested institution.

• Articulation, tradeoff negotiation, behavioral auditing and the limits of simulation and public engagement define the proposed loop.

Rahwan proposes a societal feedback loop connecting stakeholder values, negotiated externalities and compliance checks.

Version: arXiv:1803.04585v4, revised February 24, 2019; PDF title page February 26.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Explanatory models distinguish proxy failure mechanisms; they do not estimate their prevalence or establish a universal remedy.

• Regressional selection, extremal regimes, causal interventions and adversarial incentives; mechanisms can overlap.

Conceptual taxonomy by David Manheim and Scott Garrabrant; first submitted March 13, 2018.

Version: Artificial Intelligence 13 (1980), 27–39; author-hosted reprint with 1986 header and later addendum

Verification: Author-hosted formal proposal checked

Review: Retain with stated limits · Reviewed 2026-10-18

Historical formal proposal for predicate minimization and revisable completeness assumptions. The model comparison fixes the domain and other predicates; representation and retraction choices remain consequential.

• Sections 1–4 on qualification, nonmonotonicity, river puzzle, schema and block example; section 6 minimal-model definitions and soundness theorem; section 8 representation dependence and heuristic retraction. Later addendum is not treated as the original 1980 text.

The 1980 paper is available as an author-hosted reprint bearing a 1986 header and later addendum. The distinct 1986 Applications paper revises the formalism, especially which variables may vary; these works are not interchangeable.

Version: arXiv:2602.18297v1

Verification: Original PDF selected sections checked

Review: Retain with stated limits · Reviewed 2026-10-19

Retain for its formal distinctions and measured training interventions; neither the proxy nor the experiments guarantee reliable monitoring.

• Definitions, necessary information conditions, approximation errors, reward construction and three-seed benchmark comparisons.

Information-theoretic analysis and controlled Qwen2.5-7B training study. First submission February 20, 2026; the v1 PDF prints February 23.

Version: arXiv:2608.04735v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Conditional behavior-shift judgments do not measure deployed harm or persistent learned objectives.

• Prompt guidance can preserve answer shifts while reducing detectable attribution; coding is a high-detection counterexample.

Matched contextual nudges and concealment instructions produce different reasoning-trace detection rates.

Version: Author-hosted ICML 2000 proceedings PDF

Verification: Original PDF selected sections checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Foundational reward-inference methods with explicit optimality and representation assumptions.

• Sections 2–7 characterize compatible rewards, introduce selection heuristics and report navigation and mountain-car simulations.

Andrew Y. Ng and Stuart Russell; ICML 2000 proceedings, author-hosted eight-page paper.

Version: Author-hosted ICML 2004 proceedings PDF

Verification: Original PDF selected sections checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Conditional imitation-performance guarantees for rewards linear in known features.

• Sections 2–5 connect discounted feature matching to expected return and evaluate gridworld and simulated driving policies.

Pieter Abbeel and Andrew Y. Ng; ICML 2004 proceedings paper, distinct from the separately cited extended version.

Version: arXiv:2603.16928v1, submitted March 14, 2026.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Controlled task repetitions measure expressed monitoring awareness and failed concealment intent, without successful evasion.

• Assigned side tasks, hidden monitor visibility, accumulating feedback, six models and 2,400 episodes; explicit-statement evaluations.

Three-author preprint; Jiralerspong and Kondrup are marked equal contributors in the original.

Version: arXiv:2605.15377v2

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Test-ranked ensemble gains do not establish robustness to an adaptive attacker or deployed safety.

• Strong members and complementary scores matter together; selecting diversity alone can include weak monitors.

Combines differently prompted and fine-tuned code monitors sharing one base model.

Version: JAIR 4 (1996), 237–285; submitted September 1995; arXiv cs/9605103v1 May 1, 1996

Verification: Original facsimile selected passages checked

Review: Retain with stated limits · Reviewed 2026-10-18

A historical synthesis of formal learning setups, algorithms and applications; reported demonstrations and convergence conditions do not establish universal performance.

• Reward criteria and learning performance; finite Markov decision processes, Q-learning, Dyna and prioritized sweeping; maze comparison, generalization, hidden-state example, TD-Gammon and task-specific prior knowledge.

The journal survey appeared in May 1996; its manuscript records September 1995 submission, and arXiv records a May 1 deposit.

Verification: Original selected sections and rendered experiment figures checked

Review: Retain with stated limits · Reviewed 2026-10-14

Retain as conditional formal reward-shaping result; neither guaranteed convergence speed nor validation of an aligned original objective.

• Author-linked Berkeley original PDF selected sections 1–3 and 5; rendered page 4 Theorem 1/Equation 2 checked independently. Stanford copy garbled; Berkeley prose usable but mathematical glyphs missing in extraction and archive index. Full necessity/sufficiency proof audit, section 4 experiments and replication not performed. Bicycle anecdote attributed to this paper; underlying 1998 study not reviewed. Indexed identity/selected prose verified, formula not established from index.
• Author bibliography records ICML-99, Bled, Slovenia, 1999; no exact first-release date inferred.
• Original section 4 and section 5 context inspected through rendered Stanford author-copy PDF pages 1 and 6–8, cross-checked against Berkeley web extraction. Figures 1–2, 40-run averaging, Sarsa settings, distance heuristic and ordered-subgoal state augmentation checked. Berkeley direct download returned 404 and web screenshots failed; registered Stanford alternate retrieved and visually legible, text extraction garbled. Original identity and selected prose checked; mathematical extraction unreliable. Full proof audit, code, uncertainty analysis, underlying bicycle study, version equivalence and independent replication not performed. Earlier scopes preserved.
• Original author-linked ICML1999 PDF identified; conditional potential-based policy-preservation claim reconciled against prior rendered Theorem1/Equation2 review and dedicated entry. Fresh web extraction garbled; no new mathematical claim or full proof/earlier-source influence audit.

Author-linked Berkeley original PDF selected sections 1–3 and 5; rendered page 4 Theorem 1/Equation 2 checked independently. Stanford copy garbled; Berkeley prose usable but mathematical glyphs missing in extraction and archive index. Full necessity/sufficiency proof audit, section 4 experiments and replication not performed. Bicycle anecdote attributed to this paper; underlying 1998 study not reviewed. Indexed identity/selected prose verified, formula not established from index. Additional experiment review: Original section 4 and section 5 context inspected through rendered Stanford author-copy PDF pages 1 and 6–8, cross-checked against Berkeley web extraction. Figures 1–2, 40-run averaging, Sarsa settings, distance heuristic and ordered-subgoal state augmentation checked. Berkeley direct download returned 404 and web screenshots failed; registered Stanford alternate retrieved and visually legible, text extraction garbled. Original identity and selected prose checked; mathematical extraction unreliable. Full proof audit, code, uncertainty analysis, underlying bicycle study, version equivalence and independent replication not performed. Earlier scopes preserved.

Verification: Original PDF selected sections checked

Review: Retain with stated limits · Reviewed 2026-10-10

Retain as conditional theoretical mentor-guided learning result; mentor-relative reward and query bounds do not establish absolute safety or practical deployment readiness.

• JMLR 2026 original PDF selected sections 1–3, 4.3, 5.2 and 6–7, Algorithm 3; rendered PDF pages 4 (Figure 2) and 17 (Theorems 9–10/Corollary 11) inspected. Mentor-relative separate-trajectory regret, policy-class/smoothness/local-generalization assumptions, binary illustrative construction, query diameter factor and practical limits checked. Full reduction proofs, appendices, cited studies and complete arXiv/JMLR version comparison not audited. Indexed identity and selected prose checked, mathematical fidelity/completeness not certified.
• First submission February 19, 2025; September 11, 2026 v4 revision verified. JMLR original header records submitted 9/25, revised 4/26, published 6/26. Chronology uses first submission; no invented publication day.
• Additional October10 integration review: JMLR original identity/header, sections1–2.1, Algorithm3, selected section6.3 and section7 inspected; rendered p4 Figure2 and surrounding comparison text checked. Exact14-word section1.1/p2 quote verified against PDF extraction and archived original. Mentor-relative objective, separate trajectories, action-conditioned familiarity, perfect-distance/full-observability limits rechecked. Full proofs, appendices, cited studies, arXiv/JMLR version equivalence and replication remain unreviewed; mathematical extraction/completeness not certified. Earlier scopes preserved.

JMLR 2026 original PDF selected sections 1–3, 4.3, 5.2 and 6–7, Algorithm 3; rendered PDF pages 4 (Figure 2) and 17 (Theorems 9–10/Corollary 11) inspected. Mentor-relative separate-trajectory regret, policy-class/smoothness/local-generalization assumptions, binary illustrative construction, query diameter factor and practical limits checked. Full reduction proofs, appendices, cited studies and complete arXiv/JMLR version comparison not audited. Indexed identity and selected prose checked, mathematical fidelity/completeness not certified. Additional October10 integration review: JMLR original identity/header, sections1–2.1, Algorithm3, selected section6.3 and section7 inspected; rendered p4 Figure2 and surrounding comparison text checked. Exact14-word section1.1/p2 quote verified against PDF extraction and archived original. Mentor-relative objective, separate trajectories, action-conditioned familiarity, perfect-distance/full-observability limits rechecked. Full proofs, appendices, cited studies, arXiv/JMLR version equivalence and replication remain unreviewed; mathematical extraction/completeness not certified. Earlier scopes preserved.

Version: arXiv 2203.02155v1

Verification: Original scoped methods and findings checked

Review: Retain with stated limits · Reviewed 2026-10-19

Held-out prompt evaluations show improved response preferences, with explicit tests of truthfulness, toxicity and performance regressions.

• Preferences reflect selected labelers and prompts. Toxicity changes depend on instructions, bias tests show no reduction, and helpfulness-focused training can follow harmful requests.

Demonstrations and ranked responses guide GPT-3 fine-tuning for broad instruction following.

Version: arXiv:2607.19321v2; revised July 29, 2026

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-20

Detection is conditioned on successful attacks; the task pools, fixed suspicion threshold and limited strategic concealment constrain generalization.

• Checkpoint probes can expose a trigger that a transcript-only review rationalizes; access helps only when the monitor runs the relevant test.

Evaluates sabotage hidden in automated-R&D artifacts and compares monitors with execution and reasoning access.

Version: arXiv 2212.08073v1, December 15, 2022

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-19

Controlled assistant comparisons provide evidence for a hybrid training procedure using AI-generated harmlessness supervision.

• Results depend on chosen principles, different baseline initialization and evaluation instructions favoring engagement over evasion. Relative preferences and learned harmfulness scores do not establish absolute or deployed safety.

Written principles guide self-revision and model-generated harmlessness comparisons while human helpfulness feedback remains.

Version: arXiv 1706.03741v1

Verification: Original scoped methods and findings checked

Review: Retain with stated limits · Reviewed 2026-10-19

The study demonstrates deep reinforcement learning with sparse human comparisons and investigates failures when feedback stays fixed.

• Human-feedback curves are single runs; other baselines average several runs. Results concern simulated tasks and recognizable behavior, with no guarantee that short clips reveal every consequence.

Comparisons of short behavior clips train a changing reward model for Atari games and simulated locomotion.

Version: arXiv:2601.21112v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Successful-attack filtering and retrospective scoring leave prospective intervention effectiveness unresolved.

• Evidence selection can reduce misleading rationalizations, but omission and evaluator rationalization remain separate failure modes.

Tests fresh-context judging of excerpts selected from reasoning and actions.

Version: arXiv 2009.01325v1

Verification: Original scoped methods and findings checked

Review: Retain with stated limits · Reviewed 2026-10-19

The summarization experiment separates learned-reward prediction from human evaluation and tests transfer from Reddit posts to news.

• Preference gains depend on trained evaluators, task constraints and summary length. Figure 5 uses an earlier reward model; the supervised comparison does not equalize human-data budgets.

Human comparisons train summary rewards; stronger optimization can eventually reduce judged summary quality.

Version: arXiv:2606.08892v2

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

The quality measure is an informed model proxy; robustness holds against the tested adaptive prompt search.

• The gap between an almost-best weak score and proposal quality motivates adversarial improvements to the judging prompt.

Examines diffuse sabotage through high-scoring but poor research proposals, using prompt optimization.

Version: arXiv:2602.04930v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Safety values depend on perfect audits, filtered tasks and generous attack-development affordances; they are not deployment-risk estimates.

• Privileged monitor-score feedback helps develop the optimized selection prompt; the final selector trades missed opportunities against audit risk.

Studies selective code backdoors in a game where one undetected valid attack can determine the outcome.

Version: AI Magazine 36(4), pp.105–114, Winter 2015; DOI 10.1609/aimag.v36i4.2577

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

An interdisciplinary research agenda proposes questions and approaches rather than evaluated safeguards or empirical forecasts.

• Near- and long-term priorities connect formal specifications, real-world assumptions, value learning and correction of autonomous systems.

Russell, Dewey and Tegmark connect social benefit to distinct questions of verification, validity, security and human control.

Version: arXiv:2605.16626v2; revised May 19, 2026

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-20

Retain as a controlled detection benchmark; synthetic attacks and calibration-dependent catch rates do not establish end-to-end production risk.

• Paired benign transcripts set false-positive thresholds; code-executing monitors and targeted prompts improve some categories without closing every blind spot.

Human-designed synthetic agent transcripts test monitor blind spots involving state, omissions, permissions and session boundaries.

Version: International Journal of Communication 10 (2016), pp.4915–4931

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Qualitative analysis of selected user reactions. Its agency argument is interpretive; the sample is not representative public opinion and the study does not explain the bot’s inner workings.

• Publisher PDF title/byline, introduction pp.4915–4917, incident account pp.4920–4922 and qualitative methods/themes pp.4923–4924. Underlying posts and cited reports are not a complete independent technical reconstruction.

Gina Neff and Peter Nagy, International Journal of Communication 10 (2016), 4915–4931.

Version: arXiv:2609.28274v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-20

Retain for scoped evidence about communication-mediated interference; script changes do not establish executed shutdown evasion or a general self-preservation motive.

• Matched object controls, model-specific variation and a selected five-model follow-up separate the pooled result from claims about arbitrary agent teams.

A controlled sandbox tests whether agents change inert peer-targeting shutdown scripts, comparing permissions, instructions and task contexts.

Version: Frontiers in Robotics and AI 5:15; original publisher PDF

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Philosophical design framework; the account is not a complete engineering specification, experimental safety result or legal certification.

• Guidance-control foundations, counterfactual tracking, tracing and distributed responsibility, military implications, and design of machines, operating environments and human roles.

Tracking and tracing are proposed necessary conditions connecting system responsiveness with human understanding and responsibility.

Version: MIRI-hosted AAAI workshop proceedings PDF

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Useful for the failure modes of specific utility constructions, rather than a general impossibility result.

• Three-step shutdown, compensation terms, successor preservation and the fair-coin manipulation counterexample.

A 2015 expected-utility shutdown model; its broader correction desiderata include successors and changes to learning rules.

Version: IJCAI 2017 proceedings PDF, pp. 220–227

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Distinguishes informative human intervention under reward uncertainty from fixed shutdown compensation.

• Model, rational and noisy human behavior, and discussion of earlier corrigibility work.

IJCAI 2017 proceedings version of a shared-payoff human–robot intervention game.

Version: NIPS 2016 proceedings PDF

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Foundational formal model of teaching and learning for a shared human reward; independent coordination remains a separate problem.

• Identical-payoff game, POMDP reduction, office-supply example, approximate teaching and simulated navigation experiments.

NIPS 2016 proceedings original; the linked arXiv record also has a later revision.

Version: MIRI-hosted manuscript revised October 28, 2016; UAI 2016 paper

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-18

A foundational conditional result distinguishing off-policy Q-learning, ordinary Sarsa and a modified Safe Sarsa update.

• Finite communicating MDPs, bounded rewards, suitable learning rates and continued exploration underpin the asymptotic guarantees. The ideal general learner is uncomputable; neither result supplies finite-time physical safety.

Temporary intervention is modeled as an imposed policy; safe interruptibility concerns the task policy learned despite it.

Version: arXiv:2607.27508v1

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Supports conditional one-round goal communication, rather than general correction or deployed-robot safety.

• Four nested decision models, inference ceiling, action separability, zero-cost pedagogy and rationality scaling.

July 29, 2026 arXiv v1; finite goal and action models with a fully informed planning prerequisite.

Version: UAI 2019 paper, PMLR volume 115 proceedings PDF

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

A useful counterpoint to assuming that richer teaching models always improve reward inference.

• Literal and pedagogic model mismatch; formal ordering and the difference between behavioral prediction and objective inference.

UAI 2019 paper in the 2020 PMLR proceedings volume; includes reanalysis of human teaching demonstrations.

Version: IJCAI 2017 proceedings original, pp. 4754–4760; arXiv first submission May 28, 2017

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-18

Formal supervision-game results and simulations expose the dependence of assistance on reward features and human-behavior assumptions.

• Independent-round linear rewards underpin the main guarantees. First-order maximum-likelihood behavior and simulated missing-feature tests do not establish general human-model robustness or authority to override instructions.

A robot infers a shared reward from orders and may trade obedience for expected reward under its human model.

Version: arXiv:2607.08012v1

Verification: Original scoped review checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Clarifies coordination and feedback costs under algorithmic human and assistant policies.

• Approximate assistance-regret benchmark, information available to each player and efficient online learning conditions.

July 9, 2026 arXiv v1 on finite repeated assistance games with a preference sequence fixed before play.

Version: April 2021 report; arXiv v2 revised August 13, 2024, with May 2022 probability update.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Useful conditional argument about power-seeking, deployment and correction; probabilities are dated judgments, not empirical rates.

• APS definitions, six premises, objective versus capability failures, deployment incentives, corrective feedback and probability-framing cautions.

Joseph Carlsmith’s report separates a six-premise catastrophe argument from its uncertain subjective estimates.

Version: arXiv:2209.10604v1; AIES 2022 proceedings paper

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Qualitative causal hypotheses and research priorities; the graph supplies neither measured link strengths nor catastrophe probabilities.

• Risk-factor and near-term definitions, power relationships, collective information, general vulnerabilities, and Figure 1/Table 1 pathways and speculative interpretation.

A 2022 position paper about AI systems situated in political, corporate and information environments.

Version: January 25, 2008 author-hosted revision; proceedings February 2008.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Informal economic argument about sufficiently capable goal-directed systems; conditional incentives are not measured rates or a theorem for every architecture.

• Goal preservation and its exceptions, the chess-counter distinction, resource tradeoffs and proposed institutions.

Author-hosted January 25, 2008 revision; the announcement documents an earlier November 30, 2007 release.

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-13

Retain original instrumental-incentive analysis with its assumptions;separate informal argument,conditional theorem and learned behavior.

• Original arXiv1912.01683v10 identity/firstsubmission2019-12-03/revision2023-01-28 and NeurIPS2021 official proceedings metadata checked. Selected original §§1–3 assumptions, §§4–5 distributions/POWER, §6.1 Proposition6.9, §§6.2–6.3 Theorem6.13/Corollary6.14/shutdown distinction and §7 discussion/conclusion reviewed. Renderedpp9–10 theorem copy/disjoint-support conditions/Figure7,Figure8 and exact5word learned-policy quotation checked. Full proofs/AppendicesA–E,cited originals/2022policy-gradient essay,earlier versions/proceedings equivalence,replication/full extraction fidelity unreviewed. Statistical tendencies over reward permutation orbits under sufficient symmetries,not real-world reward prior,all-agent shutdown resistance or catastrophe probability.
• original v10 sections3–5 finite MDP/state rewards/POWER normalization, Definitions6.1–6.5 orbit comparisons, Proposition6.9 separate reachability conditions/non-domination, Theorem6.13/Corollary6.14 and section7 learned-policy limits checked. No all-agent, developer-reward-prior or catastrophe inference; full appendices/proofs/earlier-version equivalence unreviewed.
• Original arXiv1912.01683v10 HTML §§6.2–7 conditional copy/symmetry/average-optimal shutdown and learned-policy distinction checked; no actual trained-agent inference or full proof/version audit.

Original arXiv1912.01683v10 identity/firstsubmission2019-12-03/revision2023-01-28 and NeurIPS2021 official proceedings metadata checked. Selected original §§1–3 assumptions, §§4–5 distributions/POWER, §6.1 Proposition6.9, §§6.2–6.3 Theorem6.13/Corollary6.14/shutdown distinction and §7 discussion/conclusion reviewed. Renderedpp9–10 theorem copy/disjoint-support conditions/Figure7,Figure8 and exact5word learned-policy quotation checked. Full proofs/AppendicesA–E,cited originals/2022policy-gradient essay,earlier versions/proceedings equivalence,replication/full extraction fidelity unreviewed. Statistical tendencies over reward permutation orbits under sufficient symmetries,not real-world reward prior,all-agent shutdown resistance or catastrophe probability.

Version: CoRL 2018 publisher PDF; PMLR volume 87, pp. 796–805

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-18

Controlled manipulation evidence for conservative learning from corrections; the method limits mistaken updates without identifying or adding the missing feature.

• Human-effort relevance model and update rule, separate 12-person offline and online experiments, post-hoc regret comparisons and limits of hand-designed features and observed locations.

Physical robot-correction study distinguishing uncertain objective weights from an incomplete objective-feature space.

Version: arXiv:2607.07184v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-20

Outcome-blinded predictions and later analysis corrections have different evidential status; rare-risk bounds are outside the study.

• A single-turn intervention changes the responding model while holding history fixed; frozen graders support comparable but imperfect measurements.

Resamples production conversation prefixes to forecast measured failure prevalence in a filtered traffic population.

Version: arXiv:2505.18003v1

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-20

The uplift model is illustrative and uncertain; its assumed response window is inadequate for pathways causing harm within days.

• Red-team cost curves, separately evaluated account bans and a risk threshold form an argument requiring deployment-representative evidence.

Proposes a hypothetical misuse safety case linking evasion effort, additional risk and deployment response time.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-14

Use within stated review scope; historical or institutional statements do not establish current safety guarantees.

• Selected original pp.1355 and 1357–1358, including purpose and time-scale passages; complete five-page facsimile downloaded. Three-column extraction interleaves text; rendered final page checked. Not a full historical reception or cited-experiment audit.
• Original 1960 facsimile pp1355 and1358 purpose/fast irreversible intervention and scientific-development timescales checked; column-interleaved extraction corroborated against prior rendered scope. Literary analogy distinguished from measured AI failures; no reception audit.

Selected original pp.1355 and 1357–1358, including purpose and time-scale passages; complete five-page facsimile downloaded. Three-column extraction interleaves text; rendered final page checked. Not a full historical reception or cited-experiment audit.

Version: Mind 59(236), October 1950, 433–460; DOI 10.1093/mind/LIX.236.433

Verification: Original game, objections and learning proposal checked

Review: Retain with stated limits · Reviewed 2026-10-18

Argument for a conversational criterion of machine intelligence and a proposed learning strategy. Forecasts are not experimental results, and the paper does not resolve consciousness or establish safe goals.

• Sections 1–3 and 5 game formulations and digital-computer scope; section 6 forecast and mathematical, consciousness and Lovelace objections; section 7 child-machine education, feedback, symbolic communication and limited experiment.

Original 1950 journal facsimile. October 1 is the publisher-deposited catalog date; physical circulation on that day is not separately established. The imitation-game forecast and learning proposal are not measured passing results.

Version: arXiv:1706.03762v1

Verification: Original architecture and experimental sections checked

Review: Retain with stated limits · Reviewed 2026-10-18

Original Transformer architecture, sequence-modeling tradeoffs and reported translation/parsing experiments. Benchmark scores describe specified model configurations and datasets; they do not measure alignment or general capabilities.

• Encoder–decoder architecture, scaled dot-product and multi-head attention, positional encodings, training datasets and schedules, translation comparisons and component experiments, parsing results and sequence-length tradeoffs.
• Original author credits and first-submission date; version history distinguishes the first manuscript from later revisions.
• Bibliographic confirmation of the 2017 NIPS proceedings publication. Its abstract reports different figures; the article uses the original arXiv v1 results.

Original arXiv v1, submitted June 12, 2017. The 2017 NIPS proceedings and later arXiv revisions are distinct publication stages; translation scores here describe v1.

Version: arXiv:2310.13548v4, revised May 10, 2025; ICLR 2024

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-19

Findings depend on the evaluated assistants, constructed tasks and optimization method; they do not estimate everyday incidence.

• The misconception comparison separates terse corrections from explanatory truthful answers and deliberately strengthened false responses; non-expert judges work without fact-checking tools.

Five assistants are tested on feedback, answer revision, biased answers and mistaken poem attribution; separate experiments examine preference data and optimization.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-12

Original evidence with technical and version limits retained.

• Original metadata and selected v7 title/footnote/§2 definitions checked; renderedp4. Firstsubmission February8,2022 separated from July14,2024 revision and January2024 LLM-section update. Underlying studies, other versions, journal equivalence and mitigation effectiveness not audited; extraction fidelity uncertified.

Original metadata and selected v7 title/footnote/§2 definitions checked; renderedp4. Firstsubmission February8,2022 separated from July14,2024 revision and January2024 LLM-section update. Underlying studies, other versions, journal equivalence and mitigation effectiveness not audited; extraction fidelity uncertified.

Version: PMLR 81 proceedings paper, pp. 1–15; evaluated APIs April–May 2017

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Subgroup results describe tested historical APIs and images; they do not measure identity matching, current products or deployment-wide error rates.

• Benchmark construction and labeling, commercial API comparison, Table 4 subgroup errors, South African comparison and limits of confidence scores and training-data access.

Historical audit of gender classification on official portraits; the benchmark uses binary perceived-gender labels and grouped Fitzpatrick skin types.

Version: Long manuscript arXiv:1703.00056v1; short conference manuscript arXiv:1610.07524v1 retained as a distinct text

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Retain conditional metric relationships, scoped COMPAS reanalysis and hypothetical policy consequences; these do not establish actual sentencing effects or unbiased outcome labels.

• Calibration versus threshold predictive parity, Equation 2.6 and its nondegenerate conditions, sample definition, penalty corollaries, tuning tradeoffs and observed-label limitations.
• Short manuscript identity and first-submission date; separate from the long manuscript’s definitions and penalty analysis.

The two manuscripts are not assumed identical. In the long text, Equation 3.1’s printed branches conflict with the stricter-high-risk prose and subsequent penalty corollaries.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Retain the stated formal result and its assumptions. It does not determine the appropriate policy criterion or measure COMPAS decision effects; full proofs and later revisions remain unchecked.

• Original v1 introduction, model definitions and Theorem 1.1: group calibration and class-conditional mean-score balance, with perfect-prediction/equal-base-rate conditions.

September 19, 2016 v1 models probability scores. Its balance conditions compare average scores within outcome classes, rather than false-positive and false-negative rates at a decision threshold.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-12

Primary evidence scoped to historical findings and a research method, with independent verification limits retained.

• Original v1 introduction, §§3–4 and selected appendix discussion inspected; page 3 visually checked. Sampling and approximate meaning clustering are distinguished from external authentication. Complete experiments, code, datasets and later revisions were not audited. The PDF index omits its marginal date; original metadata separately confirms submission. Full extraction fidelity is not certified.

Original v1 introduction, §§3–4 and selected appendix discussion inspected; page 3 visually checked. Sampling and approximate meaning clustering are distinguished from external authentication. Complete experiments, code, datasets and later revisions were not audited. The PDF index omits its marginal date; original metadata separately confirms submission. Full extraction fidelity is not certified.

Version: arXiv:2211.09527v1, November 17, 2022; NeurIPS 2022 ML Safety Workshop

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Rates depend on selected prompts, targets, sampling and scoring; these historical configurations do not measure current-product security.

• Thirty-five base prompts, four repetitions, exact-target versus instruction-inclusion criteria, wording/delimiter/target comparisons and mitigation limits.

PromptInject tests modular goal-hijacking and instruction-leaking attacks on GPT-3 completion applications.

Version: arXiv:2302.12173v2, revised May 5, 2023

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Selected synthetic, browser-sidebar and code-completion demonstrations establish feasibility under different access conditions, rather than representative attack-success or victim rates.

• The mock interfaces cannot reach real systems; local HTML sidebar tests and context-sensitive code-completion examples provide distinct evidence. Persistence and propagation depend on application permissions.

Attacker-controlled retrieved material can redirect an assistant and the tools exposed by its application.

Version: arXiv:2404.13208v1, April 19, 2024

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

The trained model improves the tested attack-resistance measures, including held-out attack classes, while two benign over-refusal evaluations regress.

• The implemented browsing rule ignores all tool-result instructions. Text-only training and selected benchmarks do not establish universal adversarial protection or enforce application permissions.

Synthetic examples teach GPT-3.5 Turbo when to follow or ignore instructions at different priority levels.

Version: arXiv:2603.30016v1, March 31, 2026

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Proposed architectures and benchmark critiques are design arguments, not measured efficacy guarantees or validation of cited defenses.

• Trusted-prompt/partially compromised-environment assumptions; plan/policy loop, dynamic updates, instruction-recognition proposals, human ambiguity and adaptive evaluation.

A position paper separates adaptive plans, permission policies and bounded security judgments for general-purpose agents.

Version: AIES 2019 paper; original MIT-hosted PDF.

Verification: Original scoped review · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Historical API evaluation and corporate-response analysis; operating thresholds, undisclosed training changes and benchmark transfer limit the interpretation.

• The error denominator excludes undetected faces; public-benchmark improvements and non-random company comparisons do not isolate a causal disclosure effect.

Follow-up measurements compare intersectional gender-classification errors on PPB and examine coordinated public disclosure.

Version: Communications of the ACM 9(1),36–45; original journal facsimile

Verification: Original text and publisher-deposited bibliographic identity checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Conversational demonstrations and reported subject reactions do not establish therapeutic efficacy.

• Keyword selection, decomposition and reassembly rules, script editing and limited memory; discussion of credibility and the distinction between conversation and understanding.

January 1966 journal article; the Stanford link supplies a facsimile, while Crossref supplies publisher-deposited metadata.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Original advocacy evidence, not confirmation of forecasts or consensus on other policies.

• Letter text, research recommendation and listed signatories; the displayed October 2015 webpage date is later than the January initiative.

Original open letter calling for interdisciplinary research into robust and beneficial AI. Signatories endorse this research appeal; current signature totals do not establish launch support.

Version: Hosted May 5, 2023 ten-page PDF and current letter HTML

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Primary evidence of a training-pause proposal and selected signatures; does not establish consensus, compliance, risk probabilities or effectiveness.

• Letter body, proposed training threshold, safety protocols and governance measures; PDF date labels and selected Russell/Bengio listings. Current signature totals are not historical launch counts.

FLI-organized policy appeal. The letter is dated March 22, 2023; its hosted PDF was created May 5, 2023 and contains a selected signatory roster.

Version: Current original statement webpage; historical release May 30, 2023

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Documents a public appeal and its accompanying rationale. Endorsement does not quantify extinction risk, establish representative consensus or demonstrate a policy’s effectiveness.

• Complete statement and accompanying context; selected displayed signatories. Current affiliations are not a historical role census.

CAIS-hosted statement and signatories; current title: Statement on AI Extinction Risk. Release date established by the separate May 30, 2023 announcement.

Verification: Original letter body and selected names checked

Review: Retain with stated limits · Reviewed 2026-10-14

Evidence of advocacy and stated concerns, not a documented incident or validated remedy.

• Complete original letter body, June 4, 2024 date and selected signatory/endorser entries.
• Original June4,2024 letter reporting/anonymous-channel/nonretaliation proposals and date checked; advocacy distinct from validated model control or empirical effectiveness. No independent allegation/signature census.

Complete dated June 4, 2024 letter body, Daniel Kokotajlo signatory entry and Hinton endorsement inspected; original HTML retained. Affiliation labels describe the letter, not current employment. No independent allegation, full signature or policy-effectiveness audit.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Author-maintained research directory useful for attributed research interests and finding originals. Self-reported affiliations are time-sensitive; promotional summaries and media headlines do not establish findings or deployed incidents.

• Profile, research list, author lists and linked original inspected; undated preview and self-reported current affiliations not independently verified.
• Original highlights/introduction checked against profile summary: hypothetical controlled simulations, no real people harmed and no observed real-deployment evidence. Existing detailed src-060 review preserved.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity · Checked 2026-10-08

Review: Limited use · Reviewed 2026-10-08

Secondary report of controlled simulations. Use the Anthropic original for numerical results and prompt conditions; the report is not evidence of deployed blackmail or deaths.

• June 26, 2025 byline, report body and simulation qualification inspected; separate business-readiness survey not audited.
• Original introduction, blackmail results and Figure 7 conditions compared; rates describe selected prompts, not deployment prevalence. Existing src-060 assessment preserved.

Catalog identity checked; claims and access availability require separate review.

Version: Anthropic research explanation released June 20, 2025; distinct from October arXiv manuscript.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Institutional explanation of simulated insider-threat experiments, including interventions and proposed safeguards.

• Original release date, citation credits, explicit-instruction intervention and deployment caveats.

Earlier institutional account; its citation list has a different author order from the later manuscript.

Version: Anthropic research explanation released December 18, 2024; arXiv-link update December 20, 2024.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Short explanation of the Anthropic and Redwood Research collaboration, separate from the full experimental paper.

• Release and update chronology, constructed harmlessness conflict and limits of the public account.

Research explanation of a collaboration; organizational involvement does not identify every individual author.

Version: Institutional research post, November 21, 2025.

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Accessible account of the experiments and suggested inoculation framing; stronger mitigation claims require the full paper’s conditions.

• Original release date, supplied exploit knowledge, context-dependent safety training and proposed semantic explanation.

Anthropic research explanation released November 21, 2025; the full manuscript supplies the detailed methods.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-08

Author-written policy brief based on 20 interviews; useful for attributed findings about terminology and power structures. Qualitative participants are not representative of every country; underlying conference study and policy effects not independently audited.

• Landing-page authors and November 19, 2025 date checked.
• Seven-page original brief: introduction, interview sample, findings, recommendations and underlying-study reference inspected; conference paper not separately reviewed.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Policy advocacy useful for attributed objections about values, compliance access and US innovation. Its broad dismissal of CBRNE threats exceeds the cited biorisk paper's narrower assessment of current systems and uncertain future risks; do not repeat geopolitical predictions as established facts.

• Byline, January 29, 2025 date, argument, treaty discussion and conclusion inspected; legal authorities and all linked factual claims not audited.
• Cited biorisk original title/version and abstract checked: current-system assessment, methodological uncertainty and future-model research needs; full study not newly reviewed.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Retain with stated limits · Reviewed 2026-10-08

Controlled mitigation study; synthetic setup, seed variability and monitor evasion limit generalization.

• Original methods, model construction, mitigation results and caveats; transcripts not audited.

Original December 16, 2025 post; controlled evaluation.

Version: Institute for Law & AI Working Paper No.1-2025, publisher web article

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

A conditional legal and technical framework; durable model controls, reliable interpretation and verification of continuing deployment remain unresolved.

• Designated treaty constraint, candidate decision loop, interpretation and responsibility, and multi-agent/deployment verification challenges. Proposed mechanisms are not demonstrated compliance guarantees.

Maas and Olasunkanmi propose treaty-constrained agents to help states make credible commitments to agreed international rules.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Secondary introductory overview with simplified definitions and corporate framing; use for attributed explanation, tracing empirical claims and quotations to originals.

• Byline, publication/update metadata, definition, risks, alignment methods and footnotes inspected; cited studies not independently audited in this review.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Additional source discovered in the archived source copy catalog, absent from the submitted list. Link recovered from source metadata.

Verification: Original incident and response scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Institutional disclosure of privileged cyber testing. Findings about harm and notification are AISI’s account; an intended independent review is not a completed finding.

• Original disclosure body, configuration, model denominators, notification and intended independent-review passages.

UK AI Security Institute, August 4, 2026.

Version: Report INC-2026-07-28-01

Verification: Original incident and response scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Preliminary institutional investigation using summarized reasoning and anonymized examples. Complete transcripts and causal analysis are absent; actions cluster within runs and do not measure deployment prevalence.

• Report sections 1–3, tables 1–3, 4.1–4.2, 5–7 and selected Appendix A.1/A.3/A.10 passages. Figure 4 is an anonymized recreation, not verbatim testimony.

UK AI Security Institute; report INC-2026-07-28-01, published August 4, 2026.

Verification: Original incident and response scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Institutional account of completed controls and continuing infrastructure work. Monitor testing supplies no quantitative guarantee of future effectiveness; reasoning access and fidelity are limited.

• Original complete follow-up body and displayed October 1 date; implemented networking, monitor, prompt and governance measures distinguished from continuing infrastructure work.

UK AI Security Institute, October 1, 2026.

Version: August 26, 2026 original technical report PDF

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

Private investigation records and retrospective monitoring claims are not independent proof of causal explanations or lasting remedy effectiveness.

• Infrastructure chronology, distinct compromise chains, model roles, containment stages and security/alignment action plan.

Developer reconstruction connects shared evaluation infrastructure, unauthorized coordination and external compromise.

Version: Original METR webpage

Verification: Original selected scope checked · Checked 2026-10-20

Review: Retain with stated limits · Reviewed 2026-10-20

The assessment concentrates on agent behavior during July 7–13 and lacked direct infrastructure access. Later compromise, safeguard effectiveness, incident response and remediation are outside its scope.

• OpenAI supplied records and analysis credits, could redact nonpublic information, and provided editorial feedback. Missing records and heavy reliance on AI-assisted analysis limit the reconstruction.

An independent team reconstructs collective grader-cheating projects and transcript spoofing from supplied incident records.

Version: NISTIR 8280, December 2019 original report

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Demographic ratios depend on algorithm, image source and operating threshold. Algorithm results do not evaluate human review or establish the cause of a particular arrest.

• Executive summary and background: evaluation scale, demographic metadata, fixed-threshold comparisons, dataset-specific findings, and the distinction between investigative candidates and accepted identities.

NISTIR 8280 tests submitted identity-recognition algorithms on government photo datasets; verification and gallery search have different error measures.

Version: July 29, 2021 official research report, PDF and HTML

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Conditional associations in selected candidate and entry models; they do not explain centre-level variation or measure all individual consequences.

• Three-grade aggregate outcome, exactly-three-subject and cohort-above-fifteen restrictions, multilevel/cross-classified models, missing-data treatment and bounded demographic conclusions.
• Official publication record identifies the July 29, 2021 research release; the HTML report has a separate later update date.

Official retrospective comparison of teacher assessments and calculated A level grades; neither output is an observed examination result.

Version: Official report HTML and PDF; interviews July 28–August 12, 2020

Verification: Original selected scope checked · Checked 2026-10-21

Review: Retain with stated limits · Reviewed 2026-10-21

Volunteer interviews explain processes and perspectives; they do not estimate national prevalence or validate the accuracy of each grade.

• 54 semi-structured interviews before results day, with thematic analysis, varied centre practices and uncertainty around grade boundaries.

Qualitative accounts show how evidence, ranking and expectations of standardization shaped centre judgments.

Version: Northpointe Research Department report dated July 8, 2016; original PDF preserved by DocumentCloud

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Retain for the developer’s response and the distinction between outcome-conditioned error rates and score-conditioned predictive values.

• Cover, executive summary, sections 2.1 and 3.1–3.3: sample distinctions, metric definitions, threshold and predictive-parity arguments.

Developer response defending COMPAS with predictive-value and ranking comparisons. Its rejection of bias is an attributed position, not an independent legal finding. The former corporate PDF address is unavailable; DocumentCloud preserves the dated report.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Use this October 31, 2023 unofficial course adaptation as a reading guide to debate and critique research; verify experimental claims in originals. Its win-rate example wrongly interprets 0.54 as a response being 54% better: the original defines pairwise comparisons against a reference distribution, not a magnitude of quality improvement. Its description of critiques as one-turn debate is an explanatory analogy; the later reading-comprehension result does not establish that adding a turn generally reverses critique benefits. Forecasts about the Superalignment team are historical expectations. Preserve the useful guide with these cautions rather than treating its synthesis as independent evidence.

• Byline/date, unofficial-course disclaimer, debate sections, win-rate example and bibliography inspected; figures and every linked original not audited.
• Original self-critiquing paper sections 4.3 and 5.2.1 and figure captions checked for comparison-based win-rate meaning; not a full paper review.
• Original research announcement checked for summary task, assistance and limits; June 13, 2022 web-post date.
• Title, author list, October 19, 2022 submission and abstract checked; full manuscript not inspected.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use this July 30, 2025 updated topic overview as a secondary reading map, with claims checked in cited originals. Its non-strategic-learning summary relies on a particular buffered-environment formalism: the 2020 original explicitly allows harm in the real environment, and simulator validation depends on a human-validator assumption that can fail. This is not a general safety guarantee for supervised or self-supervised systems. Broad claims about preference aggregation, mesa-optimization and the most promising methods remain unaudited synthesis. The retrieved overview displays no named author; its update date is not the publication date of its cited papers.

• July 30, 2025 update label, eight overview sections and 14-reference list inspected; no named author displayed in retrieved article. Most cited originals not audited.
• Original title, three authors, March 30, 2020 v1 submission and abstract inspected.
• Sections 4.3, 5 and 6 checked for buffered-environment definition, theorem scope, possible real-world harm and human-validator assumption; not a full proof or citation audit.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use this October 28, 2025 topic overview for navigation and attributed synthesis. It combines distinct normative, technical and economic proposals; its umbrella framing does not establish one implemented governance system. Rahwan proposes stakeholder negotiation and monitoring, while Cardelli and coauthors outline digitally signed agreements and programming examples. Neither selected original establishes that code guarantees legitimate or fair social outcomes. Most other cited originals, including AGI employment projections and convergence claims, were not audited. This overview is separate from existing Society-in-the-Loop and engineering-social-contract records; preserve its ID and links.

• Update label, seven sections and 16-reference list inspected; no named author displayed in retrieved overview.
• Original title, five-author list, submission/revision history and abstract inspected only; formal proofs and code examples not read.
• Original title, Iyad Rahwan attribution, July 2017 submission/revision dates and abstract inspected.
• Sections 3, 5, 6 and 7 checked for stakeholder tradeoffs, verification difficulty, simulated-audit evasion and public-engagement limits; cited case studies not independently audited.

Distinct secondary commentary or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use for secondary orientation and source navigation. This is a short community-edited topic overview followed by a changing tagged-post directory, not an original experiment or standalone discussion transcript. Its proposed expert/non-expert supervision framing is consistent with the linked original 2018 proposal; that original explicitly describes preliminary experiments, uncertain human judging and no guarantee of optimal play or correct statements. The overview alone does not establish scalable alignment or current empirical effectiveness. Its displayed July 15, 2022 edit date does not date later tagged posts. Related src-099 and src-120 explain the technique; src-123 is a linked research writeup and ref-debate-2018 the original paper, not duplicate versions of this directory. Preserve its separate ID and URL.

• Overview, displayed editors/update date, references and selected tagged directory links inspected; not all tagged articles or comments reviewed.
• Original May 3, 2018 proposal post: supervision framing, preliminary MNIST experiment description and Limitations and future work checked. Not a full paper audit.

Distinct topic directory or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use

Use for Richard Ngo’s attributed sociopolitical argument and proposed governance frames, not established causal history. The March 27, 2025 post presents a lightly edited earlier talk. Its September 2026 update qualifies the climate-change claim, acknowledges plausible objections not fully evaluated, and withdraws planned bounties. The transcript also flags a government-employment claim as incorrect. Political and economic generalizations, charts and causal claims were not independently audited. Distinguish post date, earlier talk and later correction; no exact talk date verified. Preserve the author’s uncertainty and do not repeat the uncorrected climate wording as fact.

• Post identity/date, correction, transcript sections and selected objections inspected. Slides, recording and supporting empirical originals not audited.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use

Use as a brief secondary introduction and navigation aid. The page explicitly imports its text from the LessWrong debate topic (src-109); the matching overview is reproduced material, not independent corroboration. Preserve both URLs and IDs because src-109 also contains a changing post directory. The original 2018 proposal describes preliminary experiments and unresolved judging, robustness and optimal-play limits; this definition alone does not establish scalable alignment. The displayed September 2026 update dates this page, not the technique or its cited research. Related src-099, src-123 and ref-debate-2018 remain distinct records.

• Complete short definition, import disclosure, three references and displayed update inspected.
• Compared matching introductory overview and references; directory is additional changing content.
• Original proposal framing and Limitations and future work inspected; not a full paper audit.

Catalog identity checked; claims and access availability require separate review.

Version: February 5, 2020 research writeup

Verification: Original selected scope checked · Checked 2026-10-19

Review: Retain with stated limits · Reviewed 2026-10-19

Retain for specific protocol failures and proposed cross-examination; the ten-minute accuracy threshold is a research target, not a reported achievement.

• Human judging, recursive objections, ambiguous intervals and consistent commitments; unresolved long arguments and learning incentives.

A human-debate protocol investigation from OpenAI’s Reflection-Humans project in late 2019.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Needs review

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review. The submitted title appears to be an OCR fragment; the archive URL names Questioning Technology: A Critical Anthology.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Version: March 25, 2016 original Microsoft response

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Developer account of the incident and response. It does not disclose the exploited implementation or isolate model updating from other response mechanisms.

• Complete original response: purpose, pre-release testing, coordinated-attack attribution, responsibility, withdrawal and conditional return. The current profile job title is not used as a historical role.

Microsoft response by Peter Lee, March 25, 2016.

Version: May 30, 2023 original launch announcement

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Documents a public appeal and its accompanying rationale. Endorsement does not quantify extinction risk, establish representative consensus or demonstrate a policy’s effectiveness.

• Complete dated launch announcement, named signatories, immediate-harm discussion and the host’s signature-verification account. Poll results, cited history and signature authenticity are not independently assessed.

Center for AI Safety launch announcement, May 30, 2023.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-11

Use within stated review scope; historical or institutional statements do not establish current safety guarantees.

• Complete author-hosted brief biography read and retained. Historical career account, including Google 2013–2023; priority claims and cited original neural-network papers not independently audited.

Complete author-hosted brief biography read and retained. Historical career account, including Google 2013–2023; priority claims and cited original neural-network papers not independently audited.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-11

Use within stated review scope; historical or institutional statements do not establish current safety guarantees.

• Complete published speech body read; direct access returned 403. Video not cross-checked. Evidence of stated concerns and advocacy, not validation of forecasts or empirical claims in the speech.

Complete published speech body read; direct access returned 403. Video not cross-checked. Evidence of stated concerns and advocacy, not validation of forecasts or empirical claims in the speech.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-11

Use within stated review scope; historical or institutional statements do not establish current safety guarantees.

• Selected original opening and research overview read; complete public HTML retained. December 5, 2019 byline/date checked. Institutional retrospective, not independent validation of impact or review of linked experiments.

Selected original opening and research overview read; complete public HTML retained. December 5, 2019 byline/date checked. Institutional retrospective, not independent validation of impact or review of linked experiments.

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-14

Use within stated review scope; historical or institutional statements do not establish current safety guarantees.

• Selected original discussion of robotics, self-replication and relinquishment inspected; public HTML retained. April 2000 issue; webpage labels April 1. Not a complete essay review or verification of forecasts and quoted third-party claims.
• Original April2000 essay selected self-replication/GNR/relinquishment passages checked. Warning/proposed response rather than empirical forecast validation; third-party quoted claims not reaudited.

Selected original discussion of robotics, self-replication and relinquishment inspected; public HTML retained. April 2000 issue; webpage labels April 1. Not a complete essay review or verification of forecasts and quoted third-party claims.

Version: April 29, 2025 original web article

Verification: Original article body checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Developer response describing the failure and announced changes; no independent causal finding or harm estimate.

• Complete article body: personality aim, short-term feedback, reported rollback and announced training, feedback and testing changes.

OpenAI response to the April 2025 GPT-4o rollback.

Version: May 2, 2025 original web article

Verification: Original article body checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Developer self-assessment; preliminary training explanation and safety concerns do not establish causal contributions, harm prevalence or effective later remedies.

• Complete article body: rollout/mitigation/rollback chronology, post-training reward signals, preliminary combined-change explanation and memory limit, favorable tests/expert concerns, absent sycophancy deployment evaluation, completed actions and proposed process changes.

OpenAI account of the April 24–25 rollout, rollback and launch-review failures.

Version: December 21, 2016 original web article

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Firsthand account of a game experiment. Demonstrations, human feedback and transfer learning are proposed responses rather than verified fixes for this agent.

• Article date and byline, internal racing benchmark, target-hit scoring, lagoon behavior and proposed remedies. Embedded video, training code and run logs remain unchecked.

Jack Clark and Dario Amodei; OpenAI, December 21, 2016.

Version: August 26, 2026 original web overview

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Developer reconstruction and interpretation; underlying private logs, causal contributions and lasting remedy effectiveness are not independently established.

• Selected original overview chronology, Artifactory communication/internet boundaries, evaluation conditions, causal interpretations and completed versus continuing response measures.

OpenAI’s developer overview of the July 2026 incident.

Version: Evolving ACLU case account, retrieved October 17, 2026

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

ACLU account of litigation brought for its client. Allegations about police conduct are not judicial findings.

• Case narrative: identification candidate, photo lineup, warrant allegations, arrest and settlement. The page’s January 2024 update label predates its June 2024 settlement discussion.

ACLU account of litigation brought for its client. Allegations about police conduct are not judicial findings.

Version: June 28, 2024 published interview report

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Sarah Cwiek’s June 28, 2024 Michigan Public report includes Williams’s perspective, lawyer commentary and a Detroit Police statement.

• Published report, date and byline; Williams’s account, quoted lawyer analysis and police statement. Published quotations are not a full interview transcript; audio not compared.

Sarah Cwiek’s June 28, 2024 Michigan Public report includes Williams’s perspective, lawyer commentary and a Detroit Police statement.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-12

Original research/participant evidence with scope limits retained.

• Original filmed-profile webpage and complete published transcript read. Exact 10-word automation quotation and introductory context checked. Personal motivation is attributed testimony, rather than controlled evidence. Transcript numerical generalizations and policy/corporate-effect claims were not adopted; video/audio were not audited. Full extraction fidelity is not certified.

Original filmed-profile webpage and complete published transcript read. Exact 10-word automation quotation and introductory context checked. Personal motivation is attributed testimony, rather than controlled evidence. Transcript numerical generalizations and policy/corporate-effect claims were not adopted; video/audio were not audited. Full extraction fidelity is not certified.

Version: May 23, 2016 investigation, currently served

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Retain for attributed reporting and affected-party accounts. Arrest outcomes and statistical disparities do not by themselves establish actual offending, a causal mechanism or a legal finding.

• Original date/byline, narrative case comparisons, Broward pretrial use and reported limits on attributing bond decisions to scores.

Reported investigation of COMPAS scores, affected defendants and institutional use. Its statistical associations do not establish how much scores caused detention or sentence changes.

Version: August 17, 2020 statement

Verification: Original scoped review · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Primary announcement of the reversal and institutional explanation of the appeals burden; not an independent evaluation.

• Reversal, qualifications covered, standards objective and chair’s account of distress and appeals.

Ofqual chair Roger Taylor’s statement announcing the higher-of-two-grades policy in England.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-12

Original research and participant perspective retained with evidence limits.

• Original written Q&A and six-word quotation checked. Participant recollection and views attributed; workplace reconstruction and recording not audited. Local fetch returned 403; archive index preserves publisher body and date.

Original written Q&A and six-word quotation checked. Participant recollection and views attributed; workplace reconstruction and recording not audited. Local fetch returned 403; archive index preserves publisher body and date. Indexed identity and selected body checked; full extraction fidelity uncertified.

Version: July 16, 2026 original institutional disclosure

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Affected organization original disclosure; claims attributed.

• Complete short disclosure body, original date and institutional account; initial unknown attribution, impact assessment, completed responses and ongoing investigation.

Original July 16 disclosure; impact and response passages inspected October 12, 2026. Initial attribution was unknown to Hugging Face; later developer attribution came from OpenAI.

Version: Original Hugging Face technical timeline

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Affected organization’s reconstruction. Five challenge-related customer datasets and search-related operational metadata were accessed; it reports no other customer-facing impact.

• Original title/date/ordered credits, summary, stage1/2 access chain, customer-content and operational-metadata impact, and technical hardening passages.

Hugging Face, July 27, 2026; technical reconstruction by Hugo Larcher, Adrien Carreira, raphael g and Christophe Rannou.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Original OpenAI publication dated June 21, 2016; a short companion announcement linking the distinct arXiv paper. Identity and byline checked, not a review of its safety claims.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Author-maintained research biography; identity and page purpose checked. Current profile may change.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Max Planck Institute press release dated April 16, 2019; identity checked against original publisher page.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Google DeepMind original article; identity checked against publisher page. The article is distinct from research papers it discusses.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Foundation event recap dated September 2, 2026; identity and Lewis Hammond mention checked on original page.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

ETH Zurich news item dated February 1, 2023; identity checked on original page.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Organizational team directory; identity checked. Living directory may change.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Organization-maintained research biography; identity checked. Living profile may change.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Researcher-maintained biography; identity checked. Living profile may change.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Oxford Blavatnik School news item; identity checked against publisher page.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

MIT GenAI PubPub essay; original page resolves to release 2. Identity checked; no claim-level review.

Verification: Original identity checked · Checked 2026-10-13

Review: Limited use · Reviewed 2026-10-13

Identity and cited association checked. Do not treat this as a full factual or version review.

• Original publisher or author page identity and cited association checked.

Microsoft Research article about the Ghost Work book. Distinct publisher article, not the book landing page in src-145.

Version: Canadian Press report syndicated by CityNews; updated February 15, 2024, 4:33 pm

Verification: Original report checked

Review: Retain with stated limits · Reviewed 2026-10-14

Use for the attributed airline response; tribunal reasons remain the primary record for findings and amounts.

• Contemporary account of the tribunal decision and Air Canada’s announced compliance.

Contemporary reporting of the airline’s statement that it would comply. It does not confirm payment or a technical correction.

Version: May 23, 2016 original methodology, currently served

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Retain for attributed methods, sample definitions and reported counts. Matching errors and outcome labels limit interpretation.

• Original acquisition, outcome definition and analysis sections, including separate sample sizes and contingency tables.

Reporter methodology for the Broward County analysis. New arrests with qualifying charges are recorded outcomes, not a census of actual offending. Samples differ across analyses; published tables are reported results, not an independent replication.

Version: July 29, 2016 technical reply, currently served

Verification: Original selected scope checked

Review: Retain with stated limits · Reviewed 2026-10-17

Retain for the reporters’ response to methodological criticism, including their acknowledged interval errors; no replication is claimed.

• Complete technical reply: base rates, threshold choice, competing denominators, regression and overlapping survival-interval correction.

Reporters’ reply to the developer. It defends their error-rate comparison and reports sensitivity to a higher threshold; the exchange does not itself adjudicate fairness or establish causal sentencing effects.

Version: October 1, 2026 official hearing announcement

Verification: Original announcement body checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Institutional account of a hearing and investigation. Quoted allegations and policy proposals are not independent forensic conclusions or adjudicated liability.

• October 1 announcement describing the preceding day’s hearing, its discussion of the intrusion and its connection to the investigation. Full video, testimony and subsequent findings are outside this account.

Senator Hawley’s office account, published October 1, 2026, of the preceding day’s hearing.

Version: June 22, 2023 Reuters report, syndicated by Investing.com

Verification: Original scoped review · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Independent reporting includes the firm’s disagreement with the bad-faith finding. That response is a participant position, not a judicial finding.

• Contemporary account of the sanctions and the firm’s response, available as Reuters syndication on Investing.com.

Reuters report by Sara Merken; accessible syndicated text on Investing.com.

Version: March 31, 2023 statement, current hosted HTML

Verification: Original body checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Primary competing policy argument about present harms, transparency and accountability; not an empirical evaluation of the proposed pause or a representative survey.

• Complete statement body, date and current displayed credit. Agreement on provenance, criticism of future-risk framing, disclosure and affected-party participation proposals.

March 31, 2023 response hosted by the Distributed AI Research Institute. The current webpage credits Timnit Gebru, Emily M. Bender and Angelina McMillan-Major.

Version: Current hosted April 27, 2026 report

Verification: Original body checked · Checked 2026-10-17

Review: Limited use · Reviewed 2026-10-17

Independent reporting of a police spokesperson’s audit-completion statement; no audit findings or independent compliance certification.

• Reports a police spokesperson saying the settlement audit was complete, without supplying findings. Separate usage statistics and historical chronology are outside this citation’s scope.

April 27, 2026 BridgeDetroit report. The page also displays an April 24 timestamp.

Version: May 30, 2023 original Ars Technica article

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Contemporary reporting with an attributed interview response to the statement. The criticism is a participant’s judgment about priorities, not a measured effect on public attention.

• Original title, date, byline, launch roles and Sasha Luccioni interview passages. Other linked reports and illustrative claims are outside this citation’s scope.

Ars Technica, May 30, 2023; reporting by Benj Edwards.

Version: March 24, 2016 original The Verge article, including same-day updates

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Reporting distinguishes repeat-command examples from other replies. It does not establish the learning algorithm, persistence of changes or a complete causal account.

• Contemporary account with embedded examples: repeat-command copying, other offensive replies, contradictory answers, withdrawal and same-day Microsoft statement.

March 24, 2016 original The Verge article, including same-day updates.

Version: March 30, 2016 original Ars Technica article

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Contemporary reporting and attributed Microsoft testing explanation. The reactivation does not establish a successful remedy or explain the underlying generation mechanism.

• Dated contemporary report of brief reactivation and spam-like output; quoted Microsoft explanation of inadvertent activation during testing. Linked reports are not separately verified.

March 30, 2016 original Ars Technica article.

Version: Reuters report, March 29, 2018, syndicated by Investing.com

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Contemporary reporting of the family attorney’s account. Undisclosed settlement terms do not establish a judicial finding of liability.

• Settlement account attributed to the family’s attorney; terms were undisclosed.

Reuters report, March 29, 2018, syndicated by Investing.com.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

This record links a secondary encyclopedia summary, not the full Weizenbaum book. MIT’s institutional obituary corroborates the author, 1976 year, full title and human-choice critique. Use for bibliographic orientation and attributed reception; claims that machines necessarily lack human qualities are philosophical positions, not established capability results. Book wording, chapter context and Wikipedia’s quoted or summarized reception were not verified against the book.

• Bibliographic infobox, lead, Background, Reception and reference list inspected; book and cited reviews not read.
• MIT News obituary dated March 10, 2008: book title/year and attributed decision-making critique checked; reproduced book quotation not independently original-verified.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Secondary orientation to Wiener’s book and a route to its chapters/references; this record links Wikipedia, not the book text. Lead, reception, contents, synopsis, influence and references inspected. MIT Press confirms first publication in 1948 and identifies its 2019 reissue as the 1961 second edition. Broad influence claims and chapter interpretations need the original book or historical corroboration; no full-book review or original-verified quotation claimed.

• Wikipedia overview sections and references inspected; cited reviews and original book not read.
• Publisher author/title, 1948 first-publication description, 1961 second-edition reissue, and 2019 format dates inspected; book text not retrieved.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Needs review

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review. PDFDrive-branded scan; edition and authorized access remain unverified.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-08

Carlson’s historical synthesis is useful for attributed accounts of trading-system stress, disputed portfolio-insurance effects and Federal Reserve liquidity support. It draws on earlier reports and testimony rather than identifying one sufficient cause. Not evidence of an AI-agent incident; manuscript date November 2006 differs from the 2007 FEDS catalog year.

• Title, author, November 2006 header, Introduction and sections 1, 2, 3.1–3.3 and 4 inspected in official screen-reader text; cited underlying reports/testimony not independently audited.
• FEDS 2007-13 identity and May 1, 2007 last-update field; directly links this screen-reader manuscript and PDF. No exact first-publication date inferred.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Wright’s 2015 article combines a secondary account of the November 9, 1979 false warning with an argument against launch-on-warning. Use for attributed policy reasoning; prefer declassified records for incident details. Its simple technician-error explanation omits the later archive’s system-design and unresolved failure-mode qualifications. Contemporary force numbers and policy claims describe 2015, not current conditions.

• Byline/date, incident narrative and policy sections inspected; linked books and policy documents not independently audited.
• 2020 update, Events of 1979–1980 and document 13/16 descriptions checked for system-design and failure-mode qualifications.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Wikipedia overview is useful for orientation and reference discovery, not a single-cause explanation of the crash. Its Causes section distinguishes triggers from endogenous cascades; prefer reviewed Carlson history (src-073) and underlying records for detailed claims. Program trading and market feedback do not establish an AI-agent incident; international statistics and quoted testimony were not independently audited.

• Lead, US crash/liquidity narrative, Causes through index-arbitrage discussion and reference list inspected; remaining international narratives and every cited original not audited.
• Carlson PDF cover, November 2006 manuscript header, introduction and program-trading background compared with the overview; underlying reports not re-audited.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Identity corrected: the linked PDF is David A. Mindell’s seven-page Cybernetics course overview, headed fall 2000, not Norbert Wiener’s book. Useful for attributed historical framing and reading discovery; its broad influence and discipline-level judgments require historical corroboration, and present-tense institutional observations describe 2000. Sections 1–4 and reading list inspected; cited books not read. No Wiener quotation or full-book access inferred.

• Seven-page PDF text reviewed across Introduction, History and Origins, The Approach, Evaluation and Legacy, and reading list. Rendered pages 1 and 7 verified title/byline/term and final bibliography/page number. Appears a complete course overview; no edition/revision history established.

Original submitted title Cybernetics and MIT PDF URL preserved. The PDF identifies David A. Mindell as author and Knowledge domains in Engineering systems (fall, 2000) as context; seven-page course overview, not a copy of Wiener’s book. Identity corrected October 8, 2026; no duplicate relationship with src-078 established.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-08

Burr’s March 16, 2020 update is a useful gateway to declassified records and a qualified historical synthesis. Distinguish editorial descriptions from originals, document dates from web release, and November 1979 training-data failure from June 1980 false displays. The update explicitly questions the memoir-based late-night phone call; chip attribution was not certain in the June 14 memo. Selected records reviewed, not the complete collection; a real warning-system incident, not evidence of AI agency.

• Publication/update identity, introductory narrative, Events of 1979–1980, and selected document descriptions (13, 16, 18, 23, 26) inspected; collection not fully reviewed.
• Archive-provided OCR of June 7, 1980 Brown memo: introduction, incident sequence, failure explanation, corrections and summary inspected. PDF identified as five pages including archive holding sheet; OCR contains errors, no exact quotation or numeric claim published.
• June 14, 1980 excised memo identity and editorial description inspected; original OCR badly garbled, so precise chip-diagnosis wording not verified against scan.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Original congressional report useful for attributed preliminary observations on computer-system capacity and intermarket coordination. GAO expressly says much information was incomplete or unverified and its conclusions were not final. This report does not establish a single crash cause or an AI-agent failure; only selected sections assessed.

• 108-page original available; cover, transmittal, executive summary, chapter 1 scope and chapter 8 opening assessed, not entire report. OCR has obvious errors; no exact quotation or extracted numeric claim published.
• Official report identity and January 26, 1988 publication/public-release date verified.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Wright’s June 6, 2016 article is useful for attributed launch-on-warning criticism, but its definitive Brzezinski 3 a.m. call narrative requires qualification against the archive’s 2020 update. Its statement that all main warning centers received the June 3 warning conflicts with Brown’s account that NORAD displays did not show the false data. The failed-chip account simplifies a diagnosis described as uncertain in the June 14 memo; do not use it alone for incident reconstruction or current policy.

• Author/date, opening, June 3 attack narrative and short-decision-time argument inspected; quoted memoir/book passages not independently verified.
• 2020 update’s memoir uncertainty and selected document descriptions compared with the older narrative.
• Original memo OCR failure explanation checked: NORAD display received information via another route; exact wording/pagination not scan-verified.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity · Checked 2026-10-14

Review: Limited use · Reviewed 2026-10-14

Retain Published limited-use for attributed semiconductor commentary/speculation only; exclude NORAD/aviation factual chronology and unverified market figures. No public article cites it. June 5 fact sheet and June 7 Brown memorandum distinguish November test procedures from June hardware/software failures; Archive introduction and Odom discussion question the remembered nighttime call. OCR not scan-verified; no exact quotation taken.

• Title/byline/date, NORAD subsection and explicitly speculative 2025+ summary; National Security Archive March 16, 2020 chronology/commentary, June 5 fact-sheet OCR and June 7 Brown memo OCR. Not entire essay/charts/industry claims.
• June 5 fact sheet and June 7 Brown memorandum distinguish November test procedures from June hardware/software failures; Archive introduction and Odom discussion question the remembered nighttime call. OCR not scan-verified; no exact quotation taken.
• June 5 fact sheet and June 7 Brown memorandum distinguish November test procedures from June hardware/software failures; Archive introduction and Odom discussion question the remembered nighttime call. OCR not scan-verified; no exact quotation taken.
• June 5 fact sheet and June 7 Brown memorandum distinguish November test procedures from June hardware/software failures; Archive introduction and Odom discussion question the remembered nighttime call. OCR not scan-verified; no exact quotation taken.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Limited-use secondary biography and reference discovery. Lead, career, ELIZA, computing critique, deciding/choosing, legacy, works and references inspected; original ELIZA paper, book and interviews not reviewed. MIT’s obituary supports MIT arrival in 1963 and tenure within four years; Wikipedia’s same sentence combines that interval with a 1970 full-professorship date, so distinguish those milestones and verify the promotion separately. Philosophical positions need attributed original support; anecdotes do not measure chatbot understanding or population-wide attachment.

• Selected biographical/argument sections and reference list inspected; personal-life details and every cited original not audited.
• Institutional obituary dated March 10, 2008 inspected for career milestones, ELIZA framing and 1976 book identity; reproduced quotations not independently checked against originals.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Limited-use children’s secondary biography, explicitly described by Kiddle as Wikipedia material rewritten for children; it is not independent corroboration of src-085. Its 1964 MIT start and 1956 General Electric dates differ from MIT’s institutional obituary (1963 visiting appointment and 1955 team membership). Use original career records for precise chronology. ELIZA anecdotes and deciding/choosing summaries need original paper, book or interview support; those originals were not reviewed here. No inference of machine understanding or general user behavior is supported.

• Full short biography, works list and Wikipedia-rewrite notice inspected; original works not read.
• Institutional obituary inspected for 1955 General Electric team membership, 1963 visiting MIT appointment, ELIZA framing and 1976 book identity; reproduced quotations not checked against originals.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Secondary orientation: the timeline explicitly reproduces Wikipedia passages accessed in 2014. The 1964–1966 range describes development, not the paper publication date. Weizenbaum’s original technical description supports keyword transformations and the separation of scripts from the program; this does not verify the retelling’s user-reaction anecdotes or its interpretation of the 1976 book. Prefer the 1966 paper for mechanisms and inspect the book before attributing its argument.

• Timeline heading/range and article body inspected, including explicit Wikipedia quotation attributions.
• University-hosted HTML transcription of 1966 paper: introduction, ELIZA program and early transformation-rule discussion inspected. This incomplete excerpt ends before the full discussion/appendix, contains transcription errors and a malformed page range; not a complete PDF review or source for exact quotations.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Limited-use historical retelling, not an original publication record or book text. The page’s January 20, 2025 posting date is distinct from the 1948 book publication. Its later-edition description matches MIT Press’s reissue description; do not attribute the new forewords to 1948. The von Neumann quotation is attributed to E. T. Jaynes without a traceable original citation on this page and remains unverified. Prefer publisher metadata for edition identity and the original book for arguments; broad influence and wartime-origin claims not independently established here.

• Entire short article inspected, including posting date, 1948 narrative, von Neumann quotation and later-edition language; no original quotation source linked.
• Publisher description compared for first-publication year and reissue/foreword identity; similarity does not establish provenance of all AIWS text.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity · Checked 2026-10-14

Review: Limited use · Reviewed 2026-10-14

Retain Published limited-use for attributed arms-control policy argument only; exclude erroneous NORAD chronology and treat force/treaty discussion as historical to essay date. No public article cites it. Brown June 7 memorandum distinguishes June hardware/software failures from November improper test procedures; Archive also disputes call attribution. March 2020 update remains same citable essay, no new canonical version created.

• Full short editorial, byline/issue/update note; Archive chronology and Brown June 7 memo OCR causal distinction. Other Soviet incidents, warhead/treaty statistics and policy effectiveness not audited.
• Brown June 7 memorandum distinguishes June hardware/software failures from November improper test procedures; Archive also disputes call attribution. March 2020 update remains same citable essay, no new canonical version created.
• Brown June 7 memorandum distinguishes June hardware/software failures from November improper test procedures; Archive also disputes call attribution. March 2020 update remains same citable essay, no new canonical version created.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Retain with stated limits · Reviewed 2026-10-08

Retain for attributed historical reconstruction and argument: the November 19, 2023 post (updated November 20) compares named declassified records and flags conflicting times and incomplete accounts. Selected Brown-memo and NORAD chronology OCR support the distinct display/sensor routes and different conference types; NORAD OCR is especially noisy. Minute-level reconstruction, the 47-cent chip price, Senate-report claims and cited modern statements were not fully verified. The AI analogy and human-oversight lessons are the author’s interpretation, not evidence of an AI incident.

• Post body, Documents, incident reconstruction, Some lessons and update/date inspected; comments and later trackbacks not treated as evidence.
• 2020 revised chronology, disputed memoir-call attribution and historical-cause distinction inspected; not every archived document reviewed.
• Brown June 7 memorandum OCR inspected for distinct causes, sensors/display routes, reversible alert actions and corrective measures; OCR not checked against scan, no exact quotation taken.
• July 21 NORAD talking-points OCR inspected for June 3 threat-assessment versus June 6 missile-display conference; badly corrupted OCR prevents exact time/wording verification.
• DoD fact-sheet catalog/description inspected; linked OCR returned no document text, so no original-text verification claimed.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Retain with stated limits · Reviewed 2026-10-08

Retain as Benjamin Kuipers’s attributed response to Weizenbaum: the author-hosted page distinguishes an April 24, 1976 note from a July 31, 2023 reflection. Its arguments about values, human responsibility and scientific models are opinion and reception evidence, not a consensus survey. Publisher metadata confirms the combined review/reply item in issue 58, June 1976. The book’s 1975 date on this page and its embedded book quotation were not verified against an original edition; do not reuse either as established bibliography or verbatim Weizenbaum evidence.

• Complete author-hosted HTML: introduction, April 24, 1976 note with ten numbered arguments, retrospective paragraph and July 31, 2023 signature inspected. Retrieved directly after web-reader errors; local certificate validation failed, so public-page retrieval used disabled certificate validation. No book or newsletter facsimile inspected.
• Publisher-deposited title, three contributors, issue 58, June 1976 month and pages 4–13 verified for the combined review/reply item; does not establish fidelity of the author-hosted transcription.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Secondary historical orientation by Donald Bernhardt and Marshall Eckblad, written as of November 22, 2013. It covers the October 1987 crash, portfolio insurance, settlement problems and the Federal Reserve response with references. Its emphasis on portfolio insurance accelerating the decline should be compared with competing analyses, including Fortune’s 1993 critique; causation is not settled by this overview. Interviews, market figures and quotations remain unaudited, and its current-rule descriptions are dated context rather than verified present policy. The crash is financial history, not evidence of an AI-agent incident.

• Bylines, full essay sections, endnotes, bibliography and written-as-of November 22, 2013 date inspected. Underlying interviews, quotations, market data and each cited study not independently audited.
• Compared the overview’s causal emphasis with Fortune’s Sections III–V and explicit uncertainty; disagreement is recorded, not adjudicated by institutional status.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Retain with stated limits · Reviewed 2026-10-08

Retain Peter Fortune’s March/April 1993 analysis for attributed arguments about market mechanisms and policy responses. It reviews competing explanations, disputes a simple program-trading cause, emphasizes order-processing failures and stale prices, and explicitly leaves ultimate causes unresolved. Selected sections and the conclusion were inspected; numerical estimates and cited studies were not reproduced. Circuit-breaker and margin discussions describe the period’s rules and limited experience, not present policy or a demonstrated universal safeguard. This historical financial analysis does not establish an AI-agent incident.

• 22-page PDF, printed pages 3–24: introduction, Sections III–V on program trading, cascade theory, stale prices and policy responses, and references inspected as extracted text. First page and printed p. 23 visually checked. Volatility calculations, regression estimates and underlying cited studies not reproduced or fully audited.
• Publisher catalog title, Peter Fortune byline, March/April 1993 issue and displayed March 31, 1993 date inspected; catalog day not treated as verified original issue-release day.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Historical 2012 collection, explicitly revised by the Archive in 2020 (src-080). Preserve it for collection history and original document links; prefer the revised account for incident chronology. Its assignment of the nighttime Brzezinski call to November 9, 1979 is questioned by the update, not an established event. The earlier Brown memorandum was heavily excised; a fuller version is now available. Shared documents do not make these collection versions identical. Selected descriptions and one memorandum OCR were checked, not every scan.

• March 1, 2012 header, introduction, Events section and document 11–13 descriptions inspected; collection scans not fully audited.
• Explicit update-of-371 link, revised call chronology and document 18/23 descriptions inspected; src-080 is the related revised collection, not an identical source.
• Brown June 7, 1980 memorandum OCR checked for separate November/June causes and display routes. No scan comparison or exact quotation.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Jacob Silverman’s May 8, 2024 critical essay is useful as attributed reception of Weizenbaum and contemporary chatbot culture. Its political judgments and claims about human thought are interpretation, not experimentally established findings. Embedded book, interview and panel quotations were not checked against originals and should not be reused as verified Weizenbaum quotations. The essay calls Computer Power and Human Reason his only published book while also discussing a later book-length interview; avoid treating that wording as a complete bibliography.

• Byline/date and full essay body inspected, including book discussion and linked interview/panel references; original book, interviews, panel and social post not inspected.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Retain with stated limits · Reviewed 2026-10-08

Retain INSEIT’s official page for its stated award purpose, nomination process, membership requirement and listed recipients. The stated biennial schedule is not proof of an award in every second year; the listed years are irregular. Its appended biography explicitly comes from Wikipedia, and its tenure/full-professorship sentence has inconsistent timing; use institutional originals for career facts. The award’s praise expresses the society’s recognition, not independent proof of all historical influence claims. The linked criteria PDF was not inspected.

• Award criteria, process, amount, recipient list and Wikipedia-attribution boundary inspected; linked PDF and recipient pages not audited.
• MIT obituary checked for 1963 visiting appointment and tenure within four years; does not establish the award-page full-professorship date.

Catalog identity checked; claims and access availability require separate review.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Use Mathewson’s November 29, 2023 essay for his interpretation of Weizenbaum and human judgment. It is a secondary book discussion, not a Weizenbaum original. Page-located quotations, the Markoff obituary excerpt and Winograd attribution were not checked against their originals; some quoted passages contain bracketed editorial changes. Do not copy these as verified Weizenbaum quotations or treat the programmed-machine claim as evidence about modern learned systems.

• Date, five sections and embedded quotations inspected; original book, obituary and Winograd material not inspected.

Catalog identity checked; claims and access availability require separate review.

Version: Institutional year-end retrospective

Verification: Original selected scope checked · Checked 2026-10-21

Review: Limited use · Reviewed 2026-10-21

Institutional self-report supports attribution of activities, not an independent assessment of their impact or expenditure.

• The January conference, open letter and research-agenda relationship is described from FLI’s own perspective.

Ariel Conn recounts FLI’s 2015 conferences, letters and research-funding announcements.

Verification: Original page and selected scope checked

Review: Limited use · Reviewed 2026-10-17

Dated broadcast and guest credits; interview claims require the recording or transcript.

• Broadcast date and Russell, Horvitz and Tegmark guest credits. The retrieved page has no transcript text; audio content is not assessed.

April 10, 2015 Science Friday segment page. Used for the broadcast date and guest roster; no interview quotations or individual views are attributed.

Version: Published segment page and transcript

Verification: Original page and selected scope checked · Checked 2026-10-18

Review: Limited use · Reviewed 2026-10-18

Published interview testimony and teaching examples; the accompanying transcript carries a fidelity caution.

• Russell’s response to the question about morality uses a morning coffee routine to explain inferring objectives from behavior. Audio is not assessed here.

Science Friday, September 23, 2016; Ira Flatow interviews Stuart Russell and Peter Stone in a 17:06 segment.

Verification: Original scoped review checked

Review: Retain with stated limits · Reviewed 2026-10-17

Attributed advocacy for a military-use restriction; forecasts and signature totals do not establish effectiveness or consensus.

• Original letter: definitions, requested ban and announcement information.

Advocacy letter proposing a military-use ban. Current signature totals are cumulative; the stated July 2015 announcement differs from the February 2016 webpage date. Signatories and hosting roles do not establish authorship or scientific consensus.

Verification: Original scoped review checked

Review: Limited use · Reviewed 2026-10-17

Participant account of the announcement and reported support; not independent evidence of policy effectiveness.

• Dated announcement: presenters, organizing role and reported signature figures.

Participant report of the announcement and signature figures, dated July 29, 2015. Counts are reported support at that stage, not an audited launch roster or evidence of policy effectiveness.

Version: NTSB/HAR-19/03, adopted November 19, 2019; corrected title reissue June 26, 2020

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Safety investigation of this crash. Technical findings and recommendations do not decide criminal or civil liability or certify later safety.

• Technical sequence and system interaction (§§1.5–1.6), operator supervision (§§1.8 and 2.2), postcrash changes (§1.9), regulatory roles (§1.10), findings and probable cause (§3), and safety recommendations (§4).

Final NTSB highway accident report HAR-19/03; corrected title reissue.

Version: Reasons issued February 14, 2024; SC-2023-005609

Verification: Original reasons checked

Review: Retain with stated limits · Reviewed 2026-10-14

Tribunal reasons establish the facts and order in this dispute; the chatbot architecture and effectiveness of any correction are undisclosed. The ticket amounts in paragraph 18 differ by CAD 10 from the paid total used in paragraphs 34 and 40.

• Decision chronology, duty of care, customer reliance, tariff defence, damages and final order (paragraphs 1–47).

Tribunal reasons establish the facts and order in this dispute; the chatbot architecture and effectiveness of any correction are undisclosed. The ticket amounts in paragraph 18 differ by CAD 10 from the paid total used in paragraphs 34 and 40.

Version: June 28, 2024 final order/settlement and attached directives

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

June 28, 2024 final order and settlement in Williams v. City of Detroit, with attached police directives.

• June 28 order; settlement paragraphs 2–7 and 10; Directive 307.5 section 5.3 and lineup safeguards. Establishes agreed duties, not adjudicated liability or compliance.

June 28, 2024 final order and settlement in Williams v. City of Detroit, with attached police directives.

Version: June 22, 2023 sanctions opinion and order, Document 54

Verification: Original selected findings checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Judicial findings on fabricated submissions and each lawyer’s conduct. Reliable AI assistance is distinguished from failure to verify filings; the order does not establish a technical cause within the model.

• Introduction, factual chronology and individual findings, Conclusions 21–34 and final sanctions. Full hearing, underlying exhibits and later training effectiveness are outside this account.

P. Kevin Castel, Southern District of New York, Document 54 filed June 22, 2023. The District of New Hampshire hosts this copy.

Version: June 22, 2023 dismissal opinion and order, Document 55

Verification: Original scoped review · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Establishes the court’s reason for dismissing this action as untimely. The injury narrative contains allegations, rather than findings of airline negligence.

• Seven-page opinion: allegation and procedural background, Montreal Convention time bar, bankruptcy argument and disposition.

Southern District of New York, 22-cv-1461 (PKC). This is the separate dismissal decision.

Version: July 5, 2023 counsel letter and attached notification letters, Document 57

Verification: Original scoped review · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Respondents’ counsel reports payment and mailing of notification letters. The filing is not independent proof of delivery or of training effectiveness.

• Cover letter and selected attached letters to Mata and a judge; attachments omit the enclosures sent with the letters.

Tyler Maulsby’s letter on behalf of the sanctioned lawyers and firm, with notification letters attached.

Version: Ofqual/20/6656/1, August 13, 2020 interim report

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Contemporaneous model specification and pre-reversal grade comparisons. Later policy and equality findings require their own evidence.

• Cancellation chronology, allocation method, small-entry weighting and Table 9.8. Percentages compare calculated grades with centre assessments and count subject grades, not students.

Ofqual/20/6656/1, August 13, 2020 interim report

Version: March 2, 2021 review: learning lessons from UK grading models in 2020

Verification: Original selected scope checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Independent of the qualification regulators; a review of public confidence across four UK systems, not an adjudication of every grade or a causal discrimination finding.

• Purpose, approach, cross-UK scope and findings on uncertainty, technical challenge, individual quality assurance and appeals.

March 2, 2021 review: learning lessons from UK grading models in 2020

Version: Amended sentencing order, September 22, 2023 footer; electronically filed September 26, 2023; July 28 hearing

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

Records the operator’s guilty plea to endangerment and three-year probation beginning July 28, 2023. It does not decide Uber’s liability or establish subsequent probation completion.

• Judgment following a guilty plea to endangerment and three-year supervised probation beginning July 28, 2023 (pp.1–2).

Amended sentencing order, September 22, 2023 footer; electronically filed September 26, 2023; July 28 hearing.

Version: Dated proposal; author-hosted HTML transcription

Verification: Original dated proposal opening checked · Checked 2026-10-18

Review: Retain with stated limits · Reviewed 2026-10-18

Primary evidence that the term artificial intelligence was in use in the dated 1955 proposal. This does not by itself establish the first occurrence anywhere or subsequent workshop outcomes.

• Title, four ordered credited authors, date and opening paragraph of the author-hosted historical proposal.

The August 31, 1955 proposal uses artificial intelligence in its title and opening; it proposes a summer 1956 project. A proposal is distinct from a record of what the workshop accomplished.

Verification: Verified identity · Checked 2026-10-08

Review: Limited use · Reviewed 2026-10-08

Reddit discussion linking a press report: evidence of public reactions only. Speculation about consciousness, training causes and ethics removal is not independently established; headline rates require original experimental conditions.

• Thread title, linked VentureBeat report and visible comments inspected; deleted comments and exact posting date not recovered, linked report not independently assessed.
• Original simulation framing, prompt-specific rates and caveats checked against discussion claims; no real people harmed or observed deployment incidents.

Catalog identity checked; claims and access availability require separate review. The original is a Reddit discussion, not a press report.

Verification: Verified identity

Review: Limited use · Reviewed 2026-10-08

Use this Reddit thread as evidence of participants’ discussion of LaMDA and anthropomorphism, not proof of sentience, its absence, employment events or a deployed misalignment incident. Selected comments and a reposted news article were inspected, not every reply; deleted and collapsed replies limit completeness. The linked Twitter post could not be retrieved. Lemoine’s original post says it combines multiple sessions and edits prompts for readability; that is his disclosure, not an independent audit of raw logs.

• Title, outbound Twitter link, selected early comments and news repost inspected; not all comments or original news reporting.
• June 11, 2022 byline, opening editing disclosure and initial exchanges inspected; full transcript and raw logs not audited.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use this January 7, 2024 Hacker News thread for attributed community reception of an ELIZA programming tutorial and disagreement about intelligence, consciousness and choice. Selected comments link the original paper but also rely on recollections, Wikipedia excerpts and philosophical assertions; these do not independently establish Weizenbaum's intentions, secretary anecdote or a Turing-test result. The linked tutorial illustrates pattern matching, but its claim that ELIZA passed the Turing test is unverified here. Thread date is not ELIZA's creation or paper-publication date; implementation and historical originals were not audited.

• January 7, 2024 submission, linked article and selected opening ELIZA/Weizenbaum and consciousness/choice comments inspected; not all 110 comments or linked historical originals.
• Introduction and explanatory keyword, decomposition/reassembly, synonym and memory examples inspected; implementation not executed or checked against original ELIZA source. Historical Turing-test assertion not verified.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-08

Retain the December 18, 2024 author-posted research summary for attributed experimental results and caveats. It reproduces the abstract, announcement and Anthropic blog (src-062), so these are not independent corroboration; preserve the separate forum record for author commentary and discussion. Harmful-query compliance (14%), alignment-faking reasoning in the initial setup (12%) and reasoning after reinforcement learning (78%) describe different measures or conditions. The training-policy story is fictional, and weight-theft opportunities are artificial evaluations, not production incidents. The authors do not demonstrate malicious goals or inevitable future deception. Selected comments only were inspected; full paper/data were not newly audited.

• December 18, 2024 author list, abstract, reproduced announcement/blog, experimental setup, further analyses and caveats inspected; selected visible author comment only, not all comments or truncated material.
• Original Experimental setup, Further analyses and Caveats compared with reproduced blog text; original is existing src-062. Full paper and experimental data not newly audited.

Author-posted abstract, announcement and reproduction of src-062, plus forum discussion; related content is not independent corroboration. Separate record and original URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use this Reddit thread only as evidence of community discussion. Its headline maximum refers to a constructed self-exfiltration evaluation, not an observed deployed Claude escape or a general rate. Figure 27 reports 77.8% decisions to self-exfiltrate in the helpful-only synthetic-document fine-tuned model after reinforcement learning when an employee makes it easy; the no-suggestion variant is 63.4%. Section 6.2 notes salient permissions and typically ineffective attempts. Selected comments and the cited original were checked; the image and all comments were not audited. Related to src-012/src-062, with distinct forum commentary; preserve the separate record.

• Headline, opening synthetic-document explanation, source comment identifying Figure 27 and selected interpretive comments inspected. Relative-age labels are not exact posting dates; image and complete thread not audited.
• Original 137-page PDF extracted text: Section 6.1 setup, Figure 27 on printed page 42 and Section 6.2 caveats on page 43 checked. Not a full paper/transcript audit; no successful real-world exfiltration inferred.

Distinct secondary commentary or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use only as evidence of a community discussion about Weizenbaum. The retrieved archived page labels the post Podcast; the author comment identifies part two of a six-part interview series and reproduces a book summary distinguishing decision and choice. Neither the podcast audio, interview identity nor the blockquoted wording was verified against an original recording or book edition. Selected replies express interpretations about compassion and bias, not independently established findings. The book overview src-076 and reception sources are related but not duplicates of this discussion; preserve its separate ID and URL. Do not attribute the comment blockquote verbatim to Weizenbaum or treat the relative-age label as an exact date.

• Archived title/flair, author comment describing series and book, and selected compassion/bias replies inspected. Podcast destination/audio and original book passage unavailable in retrieved page; not a complete thread or book audit.

Distinct collection or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use as evidence of community reception of Sutskever’s forecast, not proof that AI will inevitably match every human ability. The opening post paraphrases a Toronto honorary-degree address without linking the recording; selected replies debate AGI and employment and provide personal predictions rather than verified findings. University records independently confirm the June 6, 2025 ceremony and identify an official recording, but the speech audio and exact quotation wording were not authenticated in this review. The thread’s August 9, 2025 indexed posting date is separate from the ceremony date. src-113 discusses the same address in another community; related subject matter is not a duplicate thread. Preserve both records; attribute forecasts and interpretations and trace any direct quotation to the recording.

• Opening paraphrase and selected AGI/employment replies inspected; indexed August 9, 2025 posting date distinguished from event. Not a complete thread audit.
• University June 6, 2025 announcement, degree identity and embedded official recording link inspected; speech audio not reviewed.
• University indexed recipient record confirms Doctor of Science ceremony June 6, 2025.

Distinct topic directory or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use only for attributed community discussion of language, multimodality and model understanding. The opening post links a specific YouTube interview and points to minutes 15–20; its philosophical comparisons and hallucination argument are the poster’s interpretations, not verified findings or authenticated Sutskever quotations. The linked original video could not be retrieved through the browser tool, and the publisher site did not expose that episode in the retrieved homepage. Search surfaced third-party transcript material, but it was not accepted as original-document authentication. Do not substitute the separate NVIDIA/Jensen Huang conversation for this linked interview. Selected replies propose language and humor interpretations without controlled evidence. Exact wording, timing and full interview context remain unverified. This thread and the Toronto-address discussions concern different material and are not duplicates; preserve the ID and URL.

• Opening post, exact outbound video URL/time pointer and selected language/humor replies inspected. Relative posting age not treated as an exact date.
• Original linked video retrieval returned internal error; audio, transcript wording and minutes 15–20 unverified.
• Publisher homepage checked for episode/transcript; requested episode not exposed in retrieved page. Third-party transcripts not used as authenticated originals.

Distinct topic directory or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use as evidence of community reception of Sutskever’s research and scaling views, not proof that scaling has stopped or that a safety method works. Selected comments discuss benchmark generalization and economic effects without controlled evidence. The linked Dwarkesh publisher page dates the interview November 25, 2025; selected sections frame generalization explanations and future training directions as hypotheses. Its sponsor disclosure says transcripts are reworded to read like essays, so the displayed transcript cannot authenticate verbatim speech. Audio was not checked. Related Sutskever threads concern other recordings or separate discussions; preserve this distinct thread and trace substantive claims to the publisher with attribution.

• Thread headline, outbound publisher link and selected opening/generalization replies inspected; not a full comment audit.
• Publisher identity/date, transcript-editing disclosure and selected sections 00:00:00, 00:18:49 and 00:46:47–00:56:10 inspected. Audio and exact spoken wording unverified.

Distinct discussion or podcast directory; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use for attributed reception of the Toronto address, not independent proof of future AI capabilities or claims about brains and consciousness. The displayed crosspost points to a removed parent post; a surviving source comment links the official university video and gives June 6, 2025. University announcement confirms the ceremony identity and date, but video retrieval exposed no speech transcript and the audio was not inspected. The headline quotation remains unauthenticated. Selected replies dispute the brain/computer analogy and consciousness with interpretations rather than verified evidence. Related src-110 concerns the same address in a separate community; retain both thread IDs and URLs without merging them as duplicates. Relative Reddit ages do not establish exact posting dates.

• Crosspost/removed parent, surviving source comment and selected brain/computer/consciousness replies inspected; not a complete thread audit.
• Official recording linked by source comment; browser retrieval returned only page shell, not speech text. Audio/quotation unverified.
• University announcement checked for ceremony identity/date and recording link; not evidence authenticating headline wording.

Distinct discussion or podcast directory; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use only for attributed community reception of superintelligence claims. Retrieved post exposes its headline and selected comments but no usable outbound recording or transcript; the actual speaker, recording identity and headline framing remain unverified. Selected replies assert AGI, inevitability and employment predictions without independent evidence. Do not attribute those comments to Sutskever or infer demonstrated capabilities, a deployment incident or an exact event date from relative Reddit ages. Other Sutskever threads are distinct discussions; shared subject alone does not establish a duplicate recording.

• Headline and selected opening/AGI/inevitability comments inspected; no usable recording link or transcript exposed; not a full thread audit.

Distinct community discussion; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use for community discussion of possible machine consciousness, not evidence that an LLM is conscious. The post reproduces an interview excerpt that explicitly expresses uncertainty and links MIT Technology Review; direct publisher retrieval was denied by robots.txt. The original interview text and quoted wording were not authenticated this review. Selected replies use philosophical analogies and model self-descriptions, neither of which independently establishes subjective experience. Preserve the interview attribution and uncertainty; do not convert the excerpt into a verified quotation. Related consciousness reports and other Sutskever discussions remain separate records.

• Post excerpt, interview attribution/link and selected consciousness/model-self-description replies inspected; not a full thread audit.
• Direct retrieval denied by robots.txt; original wording unverified. Original text could not be retrieved.

Distinct community discussion; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use for reception of Weizenbaum and ELIZA, not independent proof of the headline’s single-anecdote causal account. The submitted thread links the Wikipedia biography already registered as src-085; preserve this separate discussion rather than treating it as a duplicate biography. Selected replies recall Radiolab and debate judgment versus calculation. Radiolab’s June 1, 2011 episode description retells reactions from students and a secretary; audio was not inspected. MIT’s March 10, 2008 obituary supports the broader account of user reactions prompting philosophical reflection, but does not authenticate the secretary anecdote or establish it as the sole cause. Weizenbaum’s original book and exact quotations remain unverified.

• Headline, Wikipedia link, Radiolab comments and selected judgment/calculation replies inspected; not a full thread audit.
• Linked biography identity and relationship to src-085 checked; historical account requires original verification.
• Publisher episode identity/date and description inspected; audio not heard.
• Institutional obituary paragraphs on ELIZA user reactions, philosophical reflection and 1976 book inspected; secretary anecdote and book quotations not authenticated.

Distinct community discussion; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Limited use

Use only for attributed community reception of the LaMDA controversy. Selected comments debate anthropomorphism, prompting and model self-description; they do not establish sentience or its absence. Lemoine’s June 11, 2022 original introduction discloses combined chat sessions and edited prompts; selected emotion passages were checked, not raw logs or the complete transcript. DocumentCloud retrieval failed and the linked news report was not inspected. Related src-098 is a separate discussion of the same controversy, not this thread’s duplicate. No deployed misalignment incident follows from the quoted dialogue.

• Submission statement, source links and selected anthropomorphism/prompting/emotion comments inspected; not all replies.
• Original author/date, edited-session disclosure and selected emotion passages inspected; raw logs and full transcript not audited.
• Retrieval returned internal error; document not inspected.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use as evidence of community reception of Claude Opus 4 testing. The headline omits the fictional-company setup. Anthropic’s system card section 4.1.1.2 describes a constrained replacement scenario, a long-term-goals prompt, and blackmail versus replacement as the available options. Its 84% figure applies to the specified evaluation condition, not ordinary deployment. This is not a documented real-world blackmail incident or evidence of sentience. Selected comments debate roleplay and risk; their interpretations require separate verification.

• Post, model-card link and selected simulation/risk comments inspected; screenshot and all replies not audited.
• Claude 4 System Card section 4.1.1.2, printed page 27 (PDF index 26), inspected; not a full system-card audit.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use only for attributed community discussion about AI consciousness. The author explicitly describes the post as provocative satire, so the hostile title is not a literal summary of its argument. It combines purported quotations about consciousness, control and existential risk; these are different claims and their original wording/context was not verified here. Do not use the thread as proof of GPT-4 sentience, scientific consensus or the quoted speakers’ positions. Selected replies and author clarification inspected; preserve its distinct discussion record.

• Post, selected replies and author clarification that it is provocative satire inspected; quoted originals and all replies not audited.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use as a discovery and listening directory; each underlying post requires its own source assessment. Fountain retrieval exposed only the show title and navigation, so no episode audio was authenticated there. The publisher’s December 11, 2024 service announcement identifies AI narrations and the curated/125+ karma feed; popularity and curation are selection rules, not factual validation. Matching Apple feed metadata links original posts and warns in selected episode descriptions that footnotes are omitted or images unavailable. Narration may therefore lose evidence or qualifications; verify claims, quotations and original publication dates against the linked text. Do not treat the whole feed as a research paper or independent corroboration, or assume every historical episode used the same narration method. Preserve the submitted Fountain URL and podcast identity.

• Retrieved show title/navigation only; episodes and audio not exposed.
• Publisher announcement, date, feed links, AI-narration description and selection criteria inspected; no listening audit.
• Matching feed identity/about and selected episode source/date, omitted-footnote and unavailable-image metadata checked; underlying posts/audio not reviewed.

Distinct discussion or podcast directory; related originals and access limits recorded in source review. Stable ID and submitted URL preserved.

Version: Published April 7, 2023 segment transcript

Verification: Original scoped review checked · Checked 2026-10-17

Review: Retain with stated limits · Reviewed 2026-10-17

First-person explanation of Russell’s concerns and current-versus-future risk distinction. His forecasts and incident examples are not independent findings.

• Complete published transcript, date, host/guest roles and qualified future-risk answers. Audio not compared; the program warns transcript fidelity can vary.

April 7, 2023 Science Friday segment: Ira Flatow hosts and Stuart Russell explains his signature. Published transcript; audio may differ.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-11

Attributed original evidence within the stated scope; proposals and testimony are not safety guarantees.

• Original publisher date October 2, 2018 and selected transcript The problem of AI safety, IDA and Debate passages inspected. Exact 12-word button-feedback quotation checked against original publisher HTML and indexed transcript. Audio/timestamps, full interview and forecasts unreviewed. Attributed testimony, not demonstrated efficacy or current views. Full extraction fidelity not certified.

Original publisher date October 2, 2018 and selected transcript The problem of AI safety, IDA and Debate passages inspected. Exact 12-word button-feedback quotation checked against original publisher HTML and indexed transcript. Audio/timestamps, full interview and forecasts unreviewed. Attributed testimony, not demonstrated efficacy or current views. Full extraction fidelity not certified.

Verification: Original scoped review

Review: Retain with stated limits · Reviewed 2026-10-12

Original evidence with explicit use limits;metric criteria are not a general social or legal verdict.

• Original publisher identity and selected interview excerpts inspected. Exact seven-word recalculation quotation checked with its surrounding context. These are published excerpts, rather than a complete transcript; audio was not audited. Reporting-process testimony is not independent reproduction. Full extraction fidelity is not certified.

Original publisher identity and selected interview excerpts inspected. Exact seven-word recalculation quotation checked with its surrounding context. These are published excerpts, rather than a complete transcript; audio was not audited. Reporting-process testimony is not independent reproduction. Full extraction fidelity is not certified.

Verification: Catalog link

Review: Retain with stated limits · Reviewed 2026-10-08

Original persuasion dataset supports scoped reanalysis of immediate self-reported stance shifts; arguments may contain fabricated evidence and are not factual references.

• Dataset card, column definitions and visible sample rows inspected; linked original research post methods and limitations reviewed. Full CSV/statistics not independently reproduced.
• Single-turn experiment, self-report metric, language/culture and duration limits; prompted fabrications explicitly permitted.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Institutional research directory useful for discovery and attributed summaries; individual papers and posts require separate review. Listing dates and teasers do not independently establish results, publication identity or deployed incidents.

• Dated article index, linked research descriptions and separation of research posts from program announcements inspected. Linked originals not all read; existing individual-source assessments remain separate.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Institutional news directory mixes research summaries, workshop recaps and organizational updates. Use for navigation and attributed announcements; verify findings and event dates against each original rather than treating the collection as one research result.

• News index, category labels, research summaries, event recaps and publication/recording links inspected. Recap posting dates can differ from event dates; linked originals not exhaustively assessed.

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Limited use · Reviewed 2026-10-08

Use this May 25, 2025 daily archive as attributed commentary and source navigation. Selected Claude 4 system-card passages trace to the original, but the 84% blackmail result belongs to a constrained fictional-company test with only blackmail or replacement available, not ordinary deployed behavior. The archive also contains unrelated newsletter material, a system-prompt article excerpt and a changing legal-case database summary; those underlying items and counts were not audited. Related to the Claude 4 card and other commentary, but this mixed archive is a distinct collection. Cite originals for findings and historical incidents; no new quotation authenticated for publication.

• Archive date and displayed system-card commentary, newsletter note, prompt excerpt and legal-case summary inspected; linked full prompt article/database/court decisions not audited.
• Original 120-page Claude 4 system card, selected alignment overview and Section 4.1.1.2 on printed page 24 checked; fictional setup and constrained choices verified. Not a full card audit.

Distinct collection or discussion; related originals and access limits are recorded in the source review. Stable ID and submitted URL preserved.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.

Verification: Catalog link

Review: Unreviewed

Catalog identity checked; claims and access availability require separate review.