Social Alignment

Social alignment concerns how an AI system affects people beyond its immediate operator, and whose interests should guide it. A system can faithfully serve a company or government while imposing costs on others. Korinek and Balwit distinguish this problem from Direct Alignment, which concerns the operator’s own goals.^[1]

Operator goals and affected people

In Aligned with Whom? Direct and Social Goals for AI Systems, the authors contrast recommendation-system scenarios: an operator may object to polarization caused by engagement optimization, or may accept that side effect in pursuit of its own aims. These are the manuscript’s illustrative cases, not an empirical study of a platform. In the first, specifying the operator’s goals more fully can address the omission; in the second, more faithful implementation leaves the conflict with other people intact.^[1]

This distinction changes what a remedy must do. Improving a proxy or collecting more operator feedback cannot by itself settle which external costs are acceptable. Conversely, a socially chosen objective still needs reliable implementation. Outer Alignment develops specification; Meaningful Human Control develops the human reasons and responsibility governing actions.

Whose feedback counts?

Choosing feedback providers selects whose judgments influence training. Stable Alignment collects simulated ratings, critiques and revisions. Its authors discuss limitations from mainstream judgments and underrepresentation. The experiment tests a training procedure under its evaluation conditions; simulated people do not establish consent from actual affected groups.^[3]

Who Supplies Alignment Feedback? separates the availability of judgments, their reliability and representation. For social alignment, the representation question includes people who never operate the system or supply its training labels. Ask which people were included, how they were selected, what information they could use, and how dissent or changing values can alter the standard. These are questions for assessing a process, rather than a claim that one sampling procedure supplies legitimate authority.

Preference Aggregation addresses combining preferences; Fairness Tradeoffs addresses competing statistical criteria. Neither a combined preference score nor a parity measure alone establishes who is entitled to decide.

Disagreement, legitimacy and accountability

Korinek and Balwit use ideal social welfare as a conceptual benchmark, then discuss partial social preferences when not all choices can be ranked. This is not a computable consensus objective or tested training method. They examine norms, laws, markets and architecture as governance mechanisms for externalities.^[1]

Society-in-the-Loop proposes negotiation of value tradeoffs and the distribution of benefits and costs, followed by monitoring of an algorithmic social contract. It identifies difficulties in articulating values, measuring externalities and checking compliance. The proposal does not demonstrate that a particular institution represents everyone or enforces an agreement successfully.^[2]

Legitimacy concerns the authority and process for setting the terms; accountability concerns how decisions and consequences can be examined, challenged and corrected. A practical reading of these proposals therefore asks who can contest a rule, who must respond, and what happens after a violation. Those institutional questions remain even when technical tests show that the system follows its assigned rule.

Measurement and permissible use

Gender Shades, NIST’s evaluation and the Williams incident provide a route from measured performance to consequences of use. Identify the task, population and outcome being measured before interpreting a score. Prediction quality and justification for using a system are different questions.

England’s grading reversal and the retrospective research likewise connect institutional objectives with affected individuals. The collective-life sequence extends the governance questions to access and commitments. Social alignment requires both an account of whose interests count and a way to hold the resulting choices open to scrutiny.