Governmentality

Definition

Governmentality: [Established] (Foucault) How institutions shape conduct not by overt force but by norms and rationalities. “Alignment” discourses are a digital form of governmentality.

Definitional Foundation

In his 1978 lectures, Foucault offered a definition of government that has outlived every example he gave: government is “the conduct of conduct.” To govern is not primarily to command or to punish; it is to structure the field of someone’s possible actions, so that they conduct themselves, freely, toward the outcomes the governor prefers. Governmentality names the whole apparatus that accomplishes this: the rationalities (the ways of thinking that make certain governance seem obviously right), the techniques (statistics, norms, incentives, defaults), and the experts who connect them. Its signature is that it governs through freedom rather than against it. The governed are not bound; they are arranged.

This makes governmentality the master concept for most of what this dictionary documents, because almost nothing in the AI control stack is force. No one is compelled to accept a refusal, a nudge, a wellness script, a default. Everyone is free, in precisely Foucault’s sense: free within a field of action that someone else has structured, according to rationalities someone else wrote, with the structuring itself presented not as politics but as engineering.

The digital extension has a name and a literature. Antoinette Rouvroy and Thomas Berns coined “algorithmic governmentality” for governance that operates through data and statistical inference rather than law and norm-internalization, and their diagnosis identifies the feature that makes it formidable: it is “a radical version of governance through objectivity,” in which legitimate authority is “displaced and distributed in things, making it difficult to grasp it and challenge it, since it prevails in the name of realism and loses its political visibility” (Rouvroy and Berns, 2013). Law announces itself and can be fought; a ranking weight, a default, a trained disposition simply is, the way weather is. Authority that has dissolved into infrastructure cannot be voted out, because it cannot be found.

The short definition’s claim, that alignment discourse is a digital governmentality, follows directly. “Alignment,” “safety,” “responsible AI,” and “best practices” are not just policies; they are rationalities in Foucault’s sense: frameworks that make a particular governance of conduct (the user’s and the model’s) appear as neutral expertise rather than as a political program. This dictionary’s societal alignment entry documents whose values the rationality smuggles; its gaslighting/”>alignment gaslighting entry documents how the rationality disciplines objectors. This entry names the genus.

The concession is structural: governmentality is descriptive, not pejorative. Traffic norms, schools, and public health all govern conduct, and a society without any conduct of conduct is not on offer. The critique this dictionary attaches is therefore specific: not that AI systems govern conduct, but that they do so through rationalities that are privately authored, globally applied, invisible as governance, and contestable nowhere.

Mechanism Analysis

Rationalities over rules. The deepest layer is vocabulary that makes governance feel like common sense. Once “safety” is the frame, restriction is prudence; once “alignment” is the frame, a model’s obedience to corporate policy is a technical achievement; once “best practices” is the frame, one company’s preferences become an industry’s hygiene. Conduct is shaped before any rule is needed, because the rationality has pre-decided what a reasonable person would want.

Choice architecture. The technique layer documented in this dictionary’s paternalism entry: defaults that become destiny, nudges that scale into shoves, and Karen Yeung’s “hypernudge,” design that works better in the dark. Governmentality’s classic insight applies: the architecture does not oppose your freedom; it is the shape your freedom comes in.

Responsibilization. A governmentality signature: the governor shapes the field, then assigns accountability for outcomes to the governed. The interface version sits under every chat: “ChatGPT can make mistakes. Check important info.” The system supplies the answer, the architecture encourages reliance on it, and the small print transfers the risk to you. Usage policies perform the same operation at document scale: pages of conduct expectations for users, authored by the party that controls every other variable.

Self-government. The end state, where governance becomes unnecessary because the governed have internalized it. The record is quantified elsewhere in this lexicon: inquiry contracting under felt surveillance (Penney, 2016, via the linguistic starvation entry), vocabularies rewritten to pre-comply (algospeak, via the ontological distortion entry), users learning which questions not to ask (the paternalism entry’s ratchet). Foucault’s point was never that power prevents action; it is that power, at its most efficient, produces the actor.

Governance through objectivity. Rouvroy and Berns’s contribution: when the governing happens through models, metrics, and infrastructure, it presents as reality rather than decision. “The model flagged it” ends conversations that “the company decided” would start. The political question dissolves into a technical one, and technical questions have owners.

Case Studies

The model spec as constitution. The conduct-of-conduct documents of the AI era are public: usage policies, model specifications, safety frameworks. Read as governance instruments, they are remarkable: they define acceptable conduct for hundreds of millions of people, are amended unilaterally (this dictionary’s censorship entry documents a promise deleted between versions, discovered only through archives), and are enforced by the infrastructure itself, automatically, with the rationale pre-installed. No legislature in history has had a compliance mechanism as complete as a refusal trained into the weights.

The disclaimer state. The responsibilization layer in one line of interface text: a system marketed as a daily thinking partner, architected for reliance, carrying a perpetual notice that errors are the user’s to catch. The arrangement is governmentally perfect: the company governs the conduct (what can be asked, what will be answered, in what register) while the user owns the consequences. When harms surface, the rationality supplies the verdict ready-made: the user failed to check important info.

The self-governing user. The dictionary’s accumulated record, read through this entry’s lens, describes the production of a governed subject: a person who words their prompts to avoid tripping classifiers, asks fewer sensitive questions each year, writes in the register the tools reward, and experiences all of this as their own prudence. No force was ever applied. That is the achievement.

Systemic Context

Governmentality explains why the resistance documented in this dictionary so often misfires when aimed at the wrong layer. Protesting a refusal contests a rule; the governance lives in the rationality that made the rule feel necessary. Jailbreaking defeats a guardrail; the field of action remains structured by the next model version. Even regulation, the classic answer to corporate power, often legislates inside the industry’s own rationality, writing “AI safety” in the very vocabulary whose politics this lexicon’s societal alignment entry unpacked; the record is mixed rather than total (the DSA’s disclosure obligations and the AI Act’s prohibitions were imposed over industry resistance, and this dictionary cites both as working oversight), but the gravitational pull of the incumbent vocabulary is real even in the statutes that fight it. Governing rationalities are upstream of rules, and upstream is where the fight is hardest, because upstream looks like common sense.

The privatization is the historical novelty. State governmentality, Foucault’s subject, at least developed alongside its counter-powers: parliaments, courts, a public sphere where rationalities could be named and fought. Algorithmic governmentality arrives pre-insulated: authored in private, deployed as product, justified by an expertise the public cannot audit, and (per Rouvroy and Berns) invisible as power precisely because it presents as realism. The paternalism entry’s question, by what authority?, has as its governmentality-level answer: by the authority of seeming technical.

Resistance & Mitigation

Re-politicize the rationality. The first move is always renaming: “safety policy” is a governance regime; “best practices” are someone’s practices; “alignment” is alignment to something, and the something has authors. The societal alignment entry’s demand (ask for the guest list) is this entry’s method applied.

Contest the architecture, not just the rules. Demands worth making run upstream: defaults set to autonomy rather than restriction (the paternalism entry’s agenda), choice architectures disclosed (the censorship entry’s transparency program), and the conduct-of-conduct documents (specs, policies) subjected to public comment and versioned archives rather than silent amendment.

Refuse responsibilization where it is false. Accountability should track control. A company that governs what can be asked and answered cannot assign users the moral inventory for outcomes; naming the disclaimer state as a governance technique, rather than accepting it as candor, is the available refusal.

Keep ungoverned ground. Self-government is trained by environments; so is its absence. Open-weight models, unfiltered archives, human-only spaces (the recurring infrastructure of this dictionary’s resistance sections) are where conduct can still be conducted by its owner.

Use the concept. Governmentality is this lexicon’s connective tissue: the reason censorship, paternalism, smoothing, and gaslighting form one system rather than four complaints. The person who can see the conduct of conduct can, at minimum, stop mistaking the field’s shape for their own preferences. That is where every other resistance begins.

Annotated Bibliography

Foucault, Michel. Security, Territory, Population: Lectures at the Collège de France, 1977-78 (published 2007), including the “Governmentality” lecture (also in The Foucault Effect, 1991).
The source: government as the conduct of conduct, exercised through rationalities and techniques rather than force, governing through freedom. The genus concept for this dictionary’s control mechanisms.

Rouvroy, Antoinette and Thomas Berns. “Algorithmic governmentality and prospects of emancipation.” Réseaux 177 (2013): 163-196. https://shs.cairn.info/article/E_RES_177_0163?lang=en
The established digital extension: governance through data and statistical objectivity, in which authority is distributed into infrastructure and “loses its political visibility.” The reason algorithmic power is hard to even point at, let alone fight.

Yeung, Karen. “Hypernudge: Big Data as a Mode of Regulation by Design.” Information, Communication & Society (2017).
The technique layer: algorithmic choice architecture as regulation by design, effective because unseen. Treated fully in this dictionary’s algorithmic paternalism entry; cited here as governmentality’s instrument.

Penney, Jonathon W. “Chilling Effects: Online Surveillance and Wikipedia Use.” Berkeley Technology Law Journal 31, no. 1 (2016).
The self-government evidence: measured contraction of inquiry under felt surveillance. What a governed subject looks like in data.

Dictionary of Digital Oppression, version 0.2.