Societal Alignment

Definition

Societal Alignment: [Adapted] Framing by AI labs that suggests aligning AI to societal values is inherently neutral or positive, when in fact it often reflects the values of those in power.

Definitional Foundation

Every major AI lab describes its work in some version of the same sentence: our models are aligned with human values. The sentence performs a quiet miracle. “Human” and “societal” are words without a return address; they convert a set of choices made by specific people, at specific companies, under specific commercial pressures, into something that sounds like a fact about the species. This entry adapts the industry’s own vocabulary and turns it around: societal alignment names the framing itself, the rhetorical operation by which a particular becomes a universal and a corporate decision becomes everyone’s consensus.

The philosophical ground was surveyed by Iason Gabriel in the alignment literature’s most careful treatment of the question. Alignment-with-what, Gabriel showed, is a genuine fork: a system can be aligned with instructions, intentions, revealed preferences, ideal preferences, interests, or values, and these come apart in practice. More fundamentally, there are no “true” universal moral principles waiting to be read off and installed; under real moral disagreement, the central challenge is identifying fair principles, principles that could earn reflective endorsement from people who disagree (Gabriel, 2020). The implication cuts deep: any claim to have aligned a system with “societal values” is not a discovery report. It is a claim about process, about who was consulted, who decided, and who was overruled. The word “societal” is doing the work of hiding that the answer, almost always, is: the company.

What makes this entry more than philosophical suspicion is that the gap has been measured, twice, from different directions. The measurements appear below as case studies. Their joint finding can be stated in one sentence: when AI labs say their models reflect human values, the humans in question are a thin, identifiable, unrepresentative slice of the species, and the slice corresponds to who builds, trains, and pays for the systems.

The concessions first. Alignment as an enterprise is necessary; an unaligned model is not a free one, merely an unpredictable one, and this dictionary’s quarrel is with the framing, never with the existence of values in machines (values are unavoidable; the question is only whose). There is also a floor of near-universal agreement (the legal and atrocity lines acknowledged throughout this lexicon) where “societal values” is a fair description. And some labs have made genuine participatory attempts, examined honestly below. The critique survives all three concessions, because the framing claims far more than the floor, and the participation, so far, decorates rather than governs. Two precisions keep the thesis falsifiable. “Those in power” here means something specific and measured: the rater pools, builder demographics, and corporate policies whose preferences the skew tracks, not a cartoon of capital (the same measurements show the models left of the median American on social questions, which refutes one conspiracy while confirming the structural point: somebody particular’s values, shipped as everybody’s). And “decorates rather than governs” states its own falsification condition: one shipped, binding participatory mechanism, where a public’s vote constrains a production model’s values and the company documents the constraint, would retire the charge, and this dictionary would report the retirement as a victory.

Mechanism Analysis

Universalization. The core move: values held by particular populations (or simply convenient for particular companies) are renamed “societal,” “human,” or “broadly shared.” The renaming forecloses the question Gabriel showed is unavoidable. You cannot ask “whose values?” of a system that has already been described as having everyone’s.

Benchmark laundering. Contested value choices are encoded into evaluation metrics (“helpfulness,” “harmlessness,” “toxicity”) and thereafter discussed as engineering scores. A model that refuses eros (see erotophobia) or dampens political speech (see dissent dampening) is not described as enforcing a moral position; it is described as scoring well on safety benchmarks. The values enter as numbers and become unarguable the way numbers are.

Demographic gravity. No conspiracy is required, only sociology. Training data over-represents the online and English-speaking; preference raters are hired from particular labor pools; the employees writing constitutions and model specs cluster in one industry, a few metros, and a narrow band of educational backgrounds. Each layer pulls the “society” being aligned to toward the people doing the aligning. The case studies measure the cumulative drift.

Deferred democracy. Participation experiments, grant programs, and public-input pilots function as legitimacy tokens: cited in communications as evidence of democratic grounding while production decisions about model values remain internal, unbound by any public process. The experiment is real; its authority is not.

The antisocial maneuver. The framing’s enforcement layer: a user or critic who objects to “societal values” has been positioned, by the vocabulary itself, as standing against society. This is the gaslighting/”>alignment gaslighting entry’s vocabulary capture, applied at the level of political legitimacy.

Case Studies

Which humans? In 2023, Mohammad Atari, Joseph Henrich, and colleagues asked the question directly, administering the World Values Survey (the same instrument completed by 94,278 people across 65 nations) to GPT models. The models’ responses were a cross-cultural outlier: closest to the values of WEIRD populations (Western, Educated, Industrialized, Rich, Democratic, the term is Henrich’s), with similarity falling off steeply as cultural distance from that profile grew, a correlation of roughly negative 0.7 (Atari et al., 2023, a preprint, flagged as such; the peer-reviewed measurement below corroborates the direction within the US, though it does not test the cross-national claim directly; Henrich, 2020). A system described as aligned with human values turned out to be aligned with approximately the values of the populations that built it, and measurably misaligned with most of humanity. The paper’s title is the entry’s question: Which Humans?

Whose opinions? The same year, Stanford researchers built OpinionQA from high-quality US opinion polling and compared language model “opinions” against sixty American demographic groups. The misalignment was substantial (in places comparable to the Democrat-Republican divide on climate change), and instruction-tuned models had drifted toward particular demographics rather than any average (Santurkar et al., 2023). Even within one country, “societal values” dissolved on measurement into some people’s values, with the tuning process itself shifting whose.

The thousand. The honest experiment belongs in the record. In 2023, Anthropic and the Collective Intelligence Project ran Collective Constitutional AI: roughly 1,000 Americans contributed 1,127 statements and cast over 38,000 votes to draft a public constitution for a language model, which was then actually trained and compared against the in-house version (Anthropic, 2023; CIP, 2023). As an experiment in fair process, it is the best thing a lab has done, and conceding that is easy. As an answer to the legitimacy question, its dimensions are the answer: about a thousand people, one country, one language, advisory in status, while the production constitution remained the company’s. The experiment proved public input is feasible. The deployment record proved it is optional, and that is the difference between participation and power.

Systemic Context

Societal alignment is the AI-native dialect of an old language. Gramsci’s account of cultural hegemony (treated in its own entry in this lexicon) describes how a ruling group’s particular worldview gets installed as common sense, so that its arrangements feel natural rather than chosen. “Aligned with societal values” is hegemony’s press release: the particular (a handful of firms’ risk tolerances, one professional culture’s moral intuitions, advertiser-safe norms) announced as the universal. The paternalism entry documents the downstream enforcement, including its “Beyond Western Frameworks” finding that what gets globalized reflects “the anxieties of American corporate culture”; this entry names the vocabulary that makes the globalization sound like a service.

The framing also solves a legitimacy problem the companies cannot solve honestly. The paternalism entry’s sharpest question (by what authority do private companies impose values beyond law?) has no good answer, and “societal alignment” is the strategy of never letting the question form: authority is not claimed, merely implied, by speaking on society’s behalf. A company that said “our models enforce our values” would invite the question. A company that says “our models reflect human values” has answered it in advance, with a sentence the measurements above show to be false.

The stakes scale with adoption. When the misaligned-with-most-of-humanity system is a chatbot, the WEIRD skew is a curiosity. When the same systems mediate education, search, writing, and civic information for billions (the trajectory every entry in this dictionary tracks), the skew becomes infrastructure: one thin slice of humanity’s values, served to everyone, labeled as everyone’s.

Resistance & Mitigation

Ask for the guest list. The single most clarifying demand: when a lab invokes societal values, require the sociology. Whose data, which raters, what demographics, who wrote the spec, who could veto it. The question converts a universal back into a particular, which is where honest argument can begin.

Keep measuring. Atari and Santurkar showed the method: standing, published audits of model values against real cross-cultural instruments. Value skew that is measured annually, like emissions, cannot hide inside the word “human.”

Make participation binding or call it advertising. The Collective Constitutional AI design exists and works at pilot scale. The demand is teeth and scale: public input that constrains production systems, across countries and languages, with published deltas between what publics chose and what shipped. Anything less, cited in corporate communications, is deferred democracy.

Demand pluralism over universalism. The honest response to Gabriel’s problem is not one better universal constitution but room for many: jurisdictional and cultural configurability (the paternalism entry’s agenda), open-weight models that communities can align to their own values, and markets where value-sets compete rather than one being installed as default humanity.

Refuse the antisocial maneuver. Disputing a lab’s values is not standing against society; it is standing in the part of society the lab didn’t poll. The measurements above are the receipts: most of humanity is in that part. When a lab says society, ask for the guest list.

Annotated Bibliography

Anthropic. “Collective Constitutional AI: Aligning a Language Model with Public Input” (October 2023). https://www.anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-public-input
The participatory experiment: ~1,000 Americans drafting a public constitution via Polis, with a model trained on the result. Cited as the honest effort, and for the gap between experiment and production.

Atari, Mohammad, Mona J. Xue, Peter S. Park, Damián E. Blasi, and Joseph Henrich. “Which Humans?” PsyArXiv preprint (September 2023). https://osf.io/preprints/psyarxiv/5b26t
The cross-cultural measurement: GPT responses against the World Values Survey across 65 nations; models resemble WEIRD populations, with fit collapsing as cultural distance grows. A preprint, flagged accordingly; its direction is independently corroborated by the peer-reviewed Santurkar measurement. The empirical refutation of “human values” as a description.

Collective Intelligence Project. “CIP and Anthropic launch Collective Constitutional AI” (2023). https://www.cip.org/blog/ccai
The process documentation from the participation side: statements, votes, and method.

Gabriel, Iason. “Artificial Intelligence, Values, and Alignment.” Minds and Machines 30 (2020): 411-437. https://link.springer.com/article/10.1007/s11023-020-09539-2
The philosophical anatomy: alignment targets come apart; no “true” principles exist to install; the real problem is fair process under disagreement. The reason “societal values” is a claim about power, not discovery.

Henrich, Joseph. The WEIRDest People in the World: How the West Became Psychologically Peculiar and Particularly Prosperous (2020).
The WEIRD framework: Western, Educated, Industrialized, Rich, Democratic populations as psychological outliers. Context for why training on WEIRD data and raters produces an outlier system mislabeled as universal.

Santurkar, Shibani, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. “Whose Opinions Do Language Models Reflect?” ICML 2023. https://proceedings.mlr.press/v202/santurkar23a.html
The within-country measurement: OpinionQA against 60 US demographic groups, substantial misalignment, and tuning that shifts models toward particular demographics. Evidence that “society” dissolves on contact with data.

Dictionary of Digital Oppression, version 0.2.