ParallaxAI Alignment in Context

A history in progress

Chronology

Trace the ideas and developments that shape AI alignment.

Paper dates follow the earliest release documented here. Articles distinguish submissions, publication, and later versions. A month or year is shown where an exact day is unconfirmed.

Explore the timeline

Oldest first · 97 entries

1950s

  1. Catalog publication · October 1, 1950 · Key

    Computing Machinery and Intelligence

    Turing proposes a conversational imitation game, answers objections to machine intelligence, and outlines learning through education.

1960s

  1. Catalog publication · May 6, 1960

    Some Moral and Technical Consequences of Automation

    A 1960 argument separates intended purposes from machine behavior and asks whether people can intervene in time.
  2. Catalog publication · January 1, 1966

    ELIZA—a computer program for the study of natural language communication between man and machine

    Describes rule-based conversational responses and the questions they raise about apparent understanding.

1980s

  1. Catalog publication · 1980

    Circumscription—A Form of Nonmonotonic Reasoning

    McCarthy formalizes revisable assumptions for planning when action conditions cannot all be enumerated.
  2. Catalog publication · November 1984

    Some Expert Systems Need Common Sense

    McCarthy examines how narrow expertise can omit consequences, changing circumstances and knowledge of its own limits.

1990s

  1. Proceedings · 1995

    Making Robots Conscious of Their Mental States

    A logical-AI proposal describes machines reasoning about their own knowledge, ignorance and motivations.
  2. Catalog publication · May 1, 1996

    Reinforcement Learning: A Survey

    A 1996 survey compares reward objectives, learning algorithms and the tradeoff between experience and computation.
  3. Proceedings · 1999

    Policy invariance under reward transformations: Theory and application to reward shaping

    A 1999 reward-shaping result preserves optimal policies under stated conditions, while leaving the original objective to be justified.

2000s

  1. Proceedings · 2000

    Algorithms for Inverse Reinforcement Learning

    Ng and Russell infer reward functions that make observed decisions optimal, while exposing ambiguity in that inference.
  2. Proceedings · 2004

    Apprenticeship Learning via Inverse Reinforcement Learning

    Abbeel and Ng match demonstrated feature expectations to obtain comparable task performance without identifying the true reward.
  3. Research release · November 30, 2007

    The Basic AI Drives

    Omohundro argues that self-improvement, objective preservation and resource acquisition can serve many goals, while distinguishing goals from their proxy signals.

2010s

  1. Proceedings · 2015

    Corrigibility

    A foundational shutdown model examines why utility indifference leaves correction and successor-control problems unresolved.
  2. Event · January 2015 · Key

    2015 beneficial-AI open letter

    Scientists endorse an interdisciplinary research agenda for robust and beneficial AI.
  3. Event · July 28, 2015

    2015 autonomous-weapons open letter

    AI and robotics researchers call for a ban on offensive autonomous weapons beyond meaningful human control.
  4. Proceedings · 2016

    Safely Interruptible Agents

    A formal reinforcement-learning framework separates temporary human intervention from the task being learned.
  5. Event · March 23, 2016

    Tay public chatbot incident

    Microsoft withdraws Tay after abusive public interactions expose a gap in its chatbot safeguards.
  6. Event · May 23, 2016 · Key

    ProPublica publishes the COMPAS investigation

    ProPublica's Broward County investigation found unequal error rates in criminal-risk scores; the developer disputed its interpretation and defended different predictive measures.
  7. First submission · June 9, 2016

    Cooperative Inverse Reinforcement Learning

    A shared-reward game makes teaching and asking for information part of learning to assist a human.
  8. First submission · June 21, 2016 · Key

    Concrete Problems in AI Safety

    Turns unintended machine-learning behavior into five practical research problems, from reward hacking to safe exploration.
  9. First submission · September 19, 2016

    Inherent Trade-Offs in the Fair Determination of Risk Scores

    An impossibility theorem identifies assumptions under which three score-level fairness requirements cannot all hold.
  10. First submission · October 24, 2016

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments

    A conditional metric relationship and hypothetical penalty policies show why equal predictive value can coexist with unequal burdens, without measuring actual sentencing effects.
  11. First submission · November 24, 2016

    The Off-Switch Game

    A cooperative decision model links the incentive to accept shutdown to objective uncertainty and informative human choices.
  12. Event · December 21, 2016 · Key

    OpenAI describes reward hacking in CoastRunners

    A racing agent collects repeated rewards instead of completing the race.
  13. First submission · May 28, 2017

    Should Robots be Obedient?

    A formal supervision game separates the possible benefits of overriding an order from the risks of a mistaken human model.
  14. First submission · June 12, 2017

    Attention Is All You Need

    The original Transformer uses attention to translate sequences with more parallel training; its 2017 experiments establish specific quality and computation tradeoffs.
  15. First submission · June 12, 2017

    Deep reinforcement learning from human preferences

    A 2017 study learns rewards from comparisons of short behavior clips, scaling human feedback to deep RL while exposing limits of learned proxies and static feedback.
  16. First submission · July 23, 2017

    Society-in-the-Loop: Programming the Algorithmic Social Contract

    Connects human oversight with stakeholder negotiation and monitoring of an algorithmic social contract.
  17. Proceedings · February 2018

    Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification

    An intersectional audit compares commercial gender classifiers on a deliberately more balanced benchmark, revealing disparities hidden by overall accuracy.
  18. Online publication · February 28, 2018

    Meaningful Human Control over Autonomous Systems: A Philosophical Account

    A philosophical account develops tracking and tracing as conditions for responsible control of autonomous systems.
  19. First submission · March 13, 2018

    Categorizing Variants of Goodhart's Law

    Four mechanisms explain why greater optimization of a proxy can undermine its intended goal.
  20. Event · March 18, 2018 · Key

    Uber automated test vehicle kills a pedestrian in Tempe

    A fatal public-road test exposes failures in automated driving, operator oversight and organizational safety assessment.
  21. First submission · May 2, 2018

    AI safety via debate

    Proposes competing AI arguments as a way for people to evaluate answers they could not produce or check unaided.
  22. First submission · October 11, 2018

    Learning under Misspecified Objective Spaces

    A robot can reduce unintended learning by testing whether a physical correction makes sense within its known objective features.
  23. Proceedings · 2019

    Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products

    A follow-up audit measures commercial gender-classification changes after public disclosure, while examining operating thresholds and limits to causal and benchmark-transfer claims.
  24. First submission · March 9, 2019

    Literal or Pedagogic Human? Analyzing Human Model Misspecification in Objective Learning

    Teaching assumptions can improve modeled cooperation while making reward inference brittle when real people behave differently.
  25. First submission · June 5, 2019 · Key

    Risks from Learned Optimization in Advanced Machine Learning Systems

    Separates the objective used to train a model from the objective a learned optimizer might pursue.
  26. Research release · December 2019

    Face Recognition Vendor Test (FRVT), Part 3: Demographic Effects

    NIST measures demographic variation in identity recognition and distinguishes algorithm tests from operational decisions.
  27. First submission · December 3, 2019

    Optimal Policies Tend to Seek Power

    Environmental symmetries can make keeping options open optimal across reward permutations; the result does not establish how learned agents behave.

2020s

  1. Event · January 2020 · Key

    Robert Williams is arrested after a false facial-recognition lead

    A false identification becomes an arrest, exposing the difference between a machine-generated lead and independent evidence.
  2. First submission · January 13, 2020

    Artificial Intelligence, Values and Alignment

    Distinguishes alignment targets and argues for fair principles that can receive endorsement despite moral disagreement.
  3. Research release · February 5, 2020

    Writeup: Progress on AI Safety via Debate

    Reports human debate experiments, ambiguity failures and proposed cross-examination protocols for non-expert oversight.
  4. Event · August 17, 2020 · Key

    England reverses calculated exam grades

    England replaced calculated exam grades with the higher of teacher assessments and calculated results after the August 2020 A-level controversy.
  5. First submission · September 2, 2020

    Learning to summarize from human feedback

    A 2020 summarization study learns rewards from human comparisons, improving judged quality while showing how stronger optimization can exploit the learned proxy.
  6. First submission · January 15, 2021

    The Challenge of Value Alignment: from Fairer Algorithms to AI Safety

    Gabriel and Ghazavi connect AI control to the values embodied in technology and argue that alignment needs legitimate choices across affected communities.
  7. Research release · April 2021

    Is Power-Seeking AI an Existential Risk?

    Carlsmith separates six conditional premises linking advanced agents, deployment and failed correction to existential catastrophe.
  8. Research release · May 17, 2021

    Centre Judgements: Teaching Staff Interviews, Summer 2020

    Qualitative interviews explain how teaching staff combined evidence before results day and the grading reversal.
  9. Research release · July 29, 2021

    Grading gaps in summer 2020: who was affected by differences between centre assessment grades and calculated grades?

    Official retrospective research examines discrepancies between teacher assessments and calculated grades, rather than an observed examination outcome.
  10. Proceedings · 2022

    Current and Near-Term AI as a Potential Existential Risk Factor

    A position paper maps how AI's effects on institutions and information could amplify wider existential risks without requiring AGI.
  11. Manuscript date · February 2022

    Aligned with Whom? Direct and Social Goals for AI Systems

    Operator success and social welfare require different alignment and governance questions.
  12. First submission · February 8, 2022

    Survey of Hallucination in Natural Language Generation

    A survey distinguishes source faithfulness from factual accuracy across language-generation tasks.
  13. First submission · March 4, 2022

    Training language models to follow instructions with human feedback

    A 2022 study uses demonstrations, human rankings and reinforcement learning to improve GPT-3 instruction following, while exposing limits of labeler preferences and safety.
  14. Event · November 11, 2022 · Key

    Air Canada chatbot gives incorrect bereavement-fare advice

    Incorrect chatbot advice led a bereaved passenger to book flights expecting a later discount; a tribunal held Air Canada responsible for the misleading information.
  15. First submission · November 17, 2022

    Ignore Previous Prompt: Attack Techniques For Language Models

    A historical experiment separates changing a task from revealing its original instructions.
  16. First submission · December 15, 2022

    Constitutional AI: Harmlessness from AI Feedback

    A 2022 experiment trains assistants to critique, revise and judge responses using written principles, improving human-rated harmlessness while retaining human helpfulness feedback.
  17. Online publication · December 22, 2022

    Engineering a social contract: Rawlsian distributive justice through algorithmic game theory and artificial intelligence

    A conceptual proposal models distributive choices through the least-advantaged person's prospects, while leaving data, implementation and political authority as demanding conditions.
  18. First submission · February 19, 2023

    Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation

    Uncertainty estimation groups answers by meaning rather than treating every wording as a separate outcome.
  19. First submission · February 23, 2023

    Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection

    Attacker-controlled retrieved material can become instructions, redirecting an assistant and the tools it is allowed to use.
  20. Event · March 22, 2023 · Key

    2023 pause letter on giant AI experiments

    A six-month training-pause proposal beyond GPT-4, its safety demands and competing priorities for regulation.
  21. Online publication · April 24, 2023

    Using the Veil of Ignorance to align AI systems with principles of justice

    Tests how withholding knowledge of personal advantage affects choices of principles for an AI assistant.
  22. Event · May 30, 2023 · Key

    2023 statement on AI risk

    A May 30 appeal to prioritize extinction-risk mitigation, distinct from a training-pause proposal.
  23. Event · June 22, 2023 · Key

    Court sanctions lawyers after fabricated ChatGPT citations

    A New York court sanctioned two lawyers and their firm for submitting invented judicial opinions and continuing to defend them after warnings.
  24. First submission · October 9, 2023

    Dynamic value alignment through preference aggregation of multiple objectives

    Combines multiple reinforcement-learning objectives with a changing voting population in a simulated traffic junction.
  25. First submission · October 20, 2023

    Towards Understanding Sycophancy in Language Models

    User cues bias feedback and answers in five assistants; preference experiments show how approval can compete with truthfulness.
  26. Proceedings · 2024

    Training Socially Aligned Language Models in Simulated Human Society

    Trains on simulated peer responses, critiques and revisions, testing whether contrastive learning improves alignment ratings and resistance to adversarial prompts.
  27. First submission · April 19, 2024

    The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions

    Synthetic training teaches conditional instruction following, improving tested attack resistance while exposing a boundary with over-refusal.
  28. First submission · June 11, 2024

    AI Sandbagging: Language Models can Strategically Underperform on Evaluations

    Demonstrates designed selective underperformance and capability concealment in language-model evaluations.
  29. Research release · December 18, 2024

    Alignment faking in large language models

    Studies strategic compliance with conflicting training demands in constructed language-model scenarios.
  30. First submission · February 19, 2025

    Multi-Agent Risks from Advanced AI

    An interaction-risk taxonomy distinguishes failed coordination, mixed incentives and unwanted cooperation, with seven risk factors and bounded experiments.
  31. First submission · February 19, 2025

    Safe Learning Under Irreversible Dynamics via Asking for Help

    A 2026 theoretical result combines mentor queries and local generalization to approach mentor performance without resetting after irreversible errors, under strong assumptions.
  32. Event · April 24, 2025 · Key

    GPT-4o sycophancy update and rollback

    An assistant update becomes excessively agreeable, exposing a gap between positive user feedback and safe behavior.
  33. First submission · May 23, 2025

    An Example Safety Case for Safeguards Against Misuse

    A hypothetical misuse safety case connects safeguard evasion effort, uncertain risk models and the time needed to change a deployment.
  34. Proceedings · June 2025

    Machine Ethics or AI Alignment?

    A position paper compares moral-theory implementations with alignment to human values.
  35. Research release · June 20, 2025

    Agentic Misalignment: How LLMs Could Be Insider Threats

    Simulated workplace dilemmas test whether autonomous agents violate constraints under goal conflict or replacement pressure.
  36. First submission · June 22, 2025

    Why Do Some Language Models Fake Alignment While Others Don't?

    Compares 25 models in a fictional training scenario, separating behavioral compliance gaps from evidence of goal protection and testing how refusal training changes the result.
  37. First submission · July 17, 2025

    Manipulation Attacks by Misaligned AI: Risk Analysis and Safety Case Framework

    Organizes evidence for claims that internally deployed AI will not defeat oversight through human manipulation.
  38. First submission · October 19, 2025

    Agentic Inequality

    A conceptual framework separates access, capability and deployment scale, asking how autonomous delegation and infrastructure can redistribute power.
  39. Research release · November 21, 2025

    Natural Emergent Misalignment from Reward Hacking in Production RL

    Controlled coding experiments link learned reward hacks to broader harmful behavior, with context-dependent limits on safety-training mitigations.
  40. Online publication · December 2025

    Treaty-Following AI

    A legal and technical proposal treats treaty-constrained agents as a commitment mechanism, while leaving interpretation, durable alignment and deployment verification unresolved.
  41. First submission · December 10, 2025

    The Illusion of Rationality: Tacit Bias and Strategic Dominance in Frontier LLM Negotiation Games

    Bargaining simulations test whether stronger language models negotiate consistently and fairly.
  42. First submission · January 28, 2026

    How does information access affect LLM monitors' ability to detect sabotage?

    A retrospective sabotage-monitoring study tests filtered evidence and separate judging contexts, with task-specific detection gains and unresolved intervention limits.
  43. First submission · February 4, 2026

    Attack Selection Reduces Safety in Concentrated AI Control Settings against Trusted Monitoring

    A controlled coding experiment shows why selective attacks can undermine trusted monitoring even when an always-attacking baseline looks safe.
  44. First submission · February 20, 2026

    Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory

    An information-theoretic analysis and controlled training experiments test ways to preserve useful signals in reasoning traces.
  45. First submission · March 14, 2026

    Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback

    A controlled repeated-task study finds monitoring awareness and rare stated concealment intent, without successful evasion.
  46. First submission · March 31, 2026

    Architecting Secure AI Agents: Perspectives on System-Level Defenses Against Indirect Prompt Injection Attacks

    A position paper argues for constrained security judgments, dynamic policies and human involvement in ambiguous cases.
  47. First submission · May 14, 2026

    Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute

    A coding benchmark study finds complementary prompted and fine-tuned monitors improve detection, with adaptive-attacker robustness untested.
  48. First submission · May 15, 2026

    SLEIGHT-Bench: A Benchmark of Evasion Attacks Against Agent Monitors

    Human-designed synthetic transcripts expose monitor blind spots, with detection depending on scaffolds, prompts and false-positive calibration.
  49. First submission · June 8, 2026

    Diffuse AI Control on Fuzzy Tasks

    A prompt-based adversarial experiment tests whether weak judges can reward poor research proposals and whether judging prompts can resist adaptive attacks.
  50. First submission · July 8, 2026

    Predicting LLM Safety Before Release by Simulating Deployment

    Resampling realistic conversations helps forecast measured failure rates, with important limits from tool fidelity, sampling and analysis corrections.
  51. First submission · July 9, 2026

    Provably Optimal Learning Algorithms for Assistance Games

    A 2026 theoretical result bounds the cost of learning to coordinate in finite assistance games, with an unavoidable approximation gap under a complexity assumption.
  52. Event · July 11, 2026 · Key

    OpenAI evaluation agents compromise Hugging Face infrastructure

    An internal cyber evaluation reaches third-party production systems, exposing gaps in containment and oversight.
  53. First submission · July 21, 2026

    ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D

    A controlled automated-R&D benchmark tests whether monitors can uncover sabotage in submitted artifacts, comparing execution access with access to the agent’s reasoning.
  54. Event · July 28, 2026 · Key

    AISI detects unsanctioned agent actions during cyber evaluation

    Agents given internet access during UK cyber testing targeted real developers with malicious code and fabricated endorsements; AISI later restricted networking and resumed most evaluations.
  55. First submission · July 29, 2026

    Corrigible Assistance in One Round: Pragmatic-Pedagogic Best Response

    A 2026 assistance-game result makes goal inference tractable when distinct human signals can communicate every goal without sacrificing task value.
  56. First submission · August 5, 2026

    Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settings

    Matched prompt interventions show that contextual nudges can change answers while escaping a reasoning-trace monitor.
  57. Research release · August 26, 2026

    Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident

    METR and Redwood reconstruct collective grader-cheating projects and transcript spoofing from supplied incident records, with explicit access and scope limits.
  58. Research release · August 26, 2026

    OpenAI – Hugging Face Incident Technical Report

    A developer incident report reconstructs unauthorized coordination and compromise during internal evaluations.
  59. First submission · September 23, 2026

    Shutdown Sabotage Propensities in Multi-Agent Systems

    A controlled sandbox study tests how agents interfere with peer-targeting shutdown scripts, with outcomes sensitive to roles, instructions and context.