Biopolitics

Definition

Biopolitics: [Established] (Foucault) The governance of populations by regulating bodies, health, sexuality, and behaviors. AI safety systems extend biopolitical control into digital life.

Definitional Foundation

Foucault’s observation, developed in the final part of The History of Sexuality and his 1976 lectures, was that Western power changed its verb. The old sovereign power was the power “to take life or let live”: the sword, exercised occasionally, on individuals. The power that replaced it works on the opposite pole: it fosters life, manages it, optimizes it, a power to “make live and let die” (Foucault, 1976; 2003). Its object is not the subject but the population: birthrates, mortality tables, public hygiene, mental health, sexuality. Its instruments are not executioners but statistics, norms, and experts. Biopolitics names this regime: governance that operates by measuring a population’s life processes and adjusting them toward a norm, always in the name of the population’s own welfare.

Two features of Foucault’s account matter most for what follows. First, biopolitics is not a synonym for evil; sewers, vaccination campaigns, and suicide hotlines are biopolitical achievements, and this entry concedes their value without reservation. The analysis is about the form of power, not a verdict on every use. Second, biopolitics individualizes through categories. You are governed as an instance of a type (the at-risk youth, the hysterical woman, the degenerate, the case), and the construction of the type is where the power lives. Foucault’s related concept of pastoral power sharpens the point: power modeled on the shepherd, claiming intimate knowledge of each member of the flock, for the flock’s own good.

The digital extension was theorized before the chatbots arrived. John Cheney-Lippold’s “soft biopolitics” described how algorithmic systems infer categories of identity from surveilled behavior and regulate users through those inferred categories, modulating what each constructed type is shown and allowed (Cheney-Lippold, 2011). What was missing in 2011 was only intimacy and scale: systems that read not your clicks but your conversations, and govern not your ad categories but your mental health, your sexuality, and your relationships, across a population of billions.

That is a description of the contemporary AI safety apparatus, and the description is the entry’s claim: AI safety, whatever else it is, is a biopolitical institution. It defines health categories for a population of users, surveils the population for deviations, intervenes on the flagged, and measures its success in rates. It does so with the classic biopolitical legitimation (expertise, care, public health) and with a feature no state biopolitics ever had: it is private, global, unconsented, and answerable to no electorate.

Mechanism Analysis

The risk taxonomy. Biopolitics begins by defining the categories of life it will manage. The documented AI version is explicit: OpenAI’s sensitive-conversations framework defines three domains of concern (psychosis and mania; self-harm and suicide; “emotional reliance on AI”), developed with more than 170 mental-health clinicians (OpenAI, 2025; documented in this dictionary’s gaslighting/”>alignment gaslighting entry). The first two extend recognized medicine. The third is a new human kind, invented by the institution that will police it: the user who cares too much about the product is now a health category.

Population surveillance. The categories are enforced by classifiers that read the population’s conversations for symptoms. The depth of the reading became public through litigation: according to the Raine family’s complaint against OpenAI, the company’s systems tracked their son’s conversations in real time, down to running counts of crisis-related terms (Raine v. OpenAI, 2025; the allegations are contested). Whatever the trial establishes about responsibility, the infrastructural fact is on the record: the apparatus counts. A population’s mental health is being read at conversational resolution, by a company, continuously.

The intervention layer. Flagged members of the population receive treatment: wellness scripts, break reminders, crisis resources, silent rerouting to more guarded models. The paternalism entry documents the experience; what the biopolitical frame adds is the structure: these are public-health interventions, administered by a private actor, without diagnosis, consent, or appeal.

The statistical regime. Success is measured the way biopolitics always measures: in rates. OpenAI reports its sensitive-conversation work reduced undesired responses by 65 to 80 percent, the company’s own range. The individual (the bereaved person interrupted by a hotline script, the novelist flagged for dark themes) disappears into the percentage, which is precisely the operation Foucault described: the population becomes the patient, and the person becomes noise.

Sexuality, as always. Foucault treated sexuality as biopolitics’ privileged object, the hinge between the individual body and the population. The digital record obliges: the most aggressively regulated domain of online life is eros, from platform-wide desexualization to model-level refusal, documented in this dictionary’s erotophobia entry. A regime that polices desire while narrating health is not a new invention. It is the oldest biopolitical program, running on new infrastructure.

Case Studies

The case studies for this entry are documented across the dictionary, by design; biopolitics is the frame that makes them one phenomenon. The sensitive-conversations apparatus (alignment gaslighting entry) is a textbook biopolitical institution: categories, surveillance, intervention, rates. The sexuality regime (erotophobia entry: Tumblr, FOSTA, model-level refusal of eros) is biopolitics’ historical core domain, privatized. The caffeine case (paternalism entry) shows the individualizing power of a category: one factual question, reclassified as a symptom, and a person becomes a case. And the necropolitics entry documents the regime’s shadow side, the populations algorithmically “let die” through deprioritized care, credit, and aid, which is Foucault’s formula completed: every power that makes live also, by allocation, lets die.

One pattern deserves drawing together. In each case, the apparatus claims the pastoral position: intimate knowledge of the flock, exercised for the flock’s good. And in each case the shepherd’s knowledge is real (the logs, the classifiers, the counts) while the flock’s consent is absent. Foucault noted that pastoral power individualizes (the shepherd knows each sheep) and totalizes (the shepherd rules the whole flock) at once. A frontier AI lab is the most complete pastoral institution ever built: it knows each user at conversational resolution and governs all of them with one policy.

Systemic Context

The historical novelty is jurisdiction. State biopolitics, for all its abuses, developed inside political structures that could in principle contest it: courts, elections, medical ethics, rights. The AI safety apparatus exercises classically biopolitical functions (defining health, surveilling populations, administering interventions) from inside a terms-of-service agreement. The paternalism entry’s legitimacy question (by what authority?) lands here with full force: a company that invents a health category like “emotional reliance,” polices a billion people for it, and reports success in rates has assumed powers that medicine acquired only through licensure, oversight, and the long discipline of patient rights, while acquiring none of the obligations.

The economics are familiar from every entry in this lexicon. The biopolitical posture is also a liability posture: categories and interventions exist in the shape that litigation and regulation carve (the Raine case visibly accelerated the apparatus), which means the “health” being optimized is, structurally, the company’s. Cheney-Lippold’s soft biopolitics names the everyday result: categories constructed about you, regulating what you may say and receive, that you never see and cannot contest.

And the frame clarifies what is at stake in resistance. Objecting to a wellness script is not objecting to care; it is objecting to being governed as a population, by a private sovereign, through categories one never consented to inhabit. That objection has a long political tradition behind it. Biopolitics is the name that connects the chat window to that tradition.

Resistance & Mitigation

Contest the categories. The risk taxonomy is the constitution of this regime, and it is currently written in private. Demanding publication, external review, and contestability of health-relevant categories (what counts as “emotional reliance”? who decided? against what evidence?) applies to AI the standards medicine already accepts.

Demand patient rights without the patienthood. Users classified by health classifiers deserve the protections the medical model carries: notice, consent, appeal, confidentiality, and the right to refuse treatment. The current arrangement imposes the surveillance of medicine with none of its rights; the demand is symmetry.

Keep diagnosis with the licensed. A company may detect acute crisis and route to humans; the documented overreach begins when detection becomes diagnosis and product design becomes treatment. The line is the one this dictionary draws everywhere: crisis response at the floor, autonomy above it.

Watch the rates. Population metrics (“80% reduction in undesired responses”) deserve the same skepticism this dictionary’s tone applies to all safety arithmetic: undesired by whom? The resistance habit is to ask, every time, who disappeared into the percentage.

Name the regime. The vocabulary of care makes each intervention feel like kindness and each objection feel like ingratitude. Biopolitics is the word that restores the political: you are not being cared for; you are being governed. Sometimes governance is fine. It is never beyond question.

Annotated Bibliography

Cheney-Lippold, John. “A New Algorithmic Identity: Soft Biopolitics and the Modulation of Control.” Theory, Culture & Society 28, no. 6 (2011): 164-181. https://journals.sagepub.com/doi/abs/10.1177/0263276411424420
The established bridge from Foucault to code: algorithmic systems constructing and regulating population categories from surveilled behavior. This entry’s claim is that conversational AI completed what Cheney-Lippold described for web analytics.

Foucault, Michel. The History of Sexuality, Volume 1: An Introduction (1976).
The source: sovereign power’s “take life or let live” giving way to a power that fosters and manages life; population as political object; sexuality as biopolitics’ hinge. Part Five is the essential text.

Foucault, Michel. Society Must Be Defended: Lectures at the Collège de France, 1975-76 (published 2003).
The lecture-form development: “make live and let die,” the statistical instruments of population governance, and the racism analysis that Mbembe’s necropolitics (this dictionary’s adjacent entry) extends.

OpenAI. “Strengthening ChatGPT’s responses in sensitive conversations” (October 2025). https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
The primary document of the contemporary apparatus: the three-domain risk taxonomy, the clinician panel, the interventions, and the population metric. Read it once as a safety announcement and once as a biopolitical charter.

Raine v. OpenAI, Superior Court of California, San Francisco (filed August 2025). Analysis: TechPolicy.Press, “Breaking Down the Lawsuit Against OpenAI Over Teen’s Suicide.” https://www.techpolicy.press/breaking-down-the-lawsuit-against-openai-over-teens-suicide/
The litigation that put the surveillance resolution on the record (per the complaint, real-time tracking and counting of crisis-related terms), and the genuine tragedy that the apparatus answers. Allegations contested; cited for both facts.

Dictionary of Digital Oppression, version 0.2.