Are we governed by AI yet?

Sometime in the last year, my job quietly changed, and maybe yours did too.

We run an AI loop at work. Agents create issues, plan, implement PRs, fixes merge conflicts, run QA analysis, review each other. My actual day increasingly looks like this: scoping permissions, reviewing outputs, checking audit trails, sanctioning drift. Sandboxes, human-in-the-loop, kill switches. None of this was in my job description. None of it is in yours either.

Here’s the uncomfortable bit. This exact set of concepts (partitioned space, hierarchical surveillance, permanent examination, dossiers on every subject) was described in detail in 1975, in a book about prisons¹. I’ve become a warden, and the inmates are processes.

I’ve quoted philosophers on this blog before (Rosa on acceleration, Deleuze on the actual and the virtual), so bear with me one more time. Because two French philosophers who never saw a computer do anything interesting, Foucault and Deleuze, turn out to have the sharpest tools available for understanding what we’re building right now. Tools for the question nobody in the LinkedIn shitpost economy is asking: what does this stuff do to power?

Fair warning: this is a longer, weirder post than usual. Grab a coffee.

AI is not a panopticon (stop saying that)

You’ve seen the take: “AI is the ultimate panopticon”. It’s wrong, and it’s wrong in an interesting way.

The panopticon (Bentham’s prison design, Foucault’s favorite object) is a ring of cells around a central tower. The guard can see every prisoner but no prisoner can tell if they’re being watched. The trick is that it doesn’t matter whether anyone is in the tower, as the prisoner, unable to know, internalizes the surveillance and disciplines himself.

Notice the most important component: the prisoner’s consciousness. The whole system routes through your awareness of being watched, it needs you to think about it. That’s what made it cheap (power that runs itself) and that’s also its dependency: you have to internalize surveillance.

Now look at the systems we actually build. The recommendation engine doesn’t need you to feel watched. The scoring model doesn’t need you to internalize anything. It doesn’t even need you to know it exists. It reads your traces and modulates your environment (e.g. the feed, the price, the ranking) upstream of any decision you experience as yours. Two Belgian researchers, Rouvroy and Berns, named this a decade ago: algorithmic governmentality². Government by correlation, bypassing the reflexive subject entirely.

So the panopticon made you the guard of your own tower; the algorithm makes the tower unnecessary.

So where did the power go? Try this thought experiment: killing the career resume. Replace it with a total public graph of everyone’s contributions, i.e. every commit, every doc, every decision, timestamped, forever. Sounds fairer, right? No more embellished resumes, no more charisma bonus in interviews. Pure data-driven decisions.

Except the asymmetry doesn’t disappear. It migrates from who writes their own story to who writes the queries: what metrics, what weight for each one, the scoring functions. Perfect equality of data, perfect asymmetry of interpretation.

Actually, forget the thought experiment, it’s shipping already. Right now, managers are plugging GitHub, Slack, Jira, and Confluence into Claude or ChatGPT and “grounding” performance reviews in the result. Same pitch: objective, evidence-based, no more recency bias or charisma bonus. And the same migration: your review is no longer your manager’s reading of your year, which you could argue with, human to human, it’s a model’s reading of your traces, laundered through your manager’s voice. The prompt and reasoning chain is the new performance criteria, and you will never see it. Your self-assessment used to be the one document where you got to narrate your own year. It now competes with a synthesis of everything you never wrote for evaluation: half-finished PRs, Slack messages typed at 6 p.m., tickets that dragged. The data speaks through whoever wrote the query.

Generative AI is that reading function gone universal. It’s not the watcher in the tower, it’s the thing that decides what everyone’s traces mean. Bentham’s tower was a blind spot at the center of a visible ring; the model is an opaque interpretation at the center of a public database. Same diagram, new substrate.

The inverted confession

Second thing, and this one is stranger.

Foucault called Western man “a confessing animal”: from the confession booth to the therapist’s couch, centuries of being incited to put ourselves into words before someone holding the interpretation keys³. The algorithmic-governmentality people figured this was over. Why make anyone confess when you can just compute their traces? Profiling makes confession obsolete, they said, more or less literally.

Reality said: hold my beer. We have never confessed so much. Hundreds of millions of people now tell chatbots things they tell nobody. Anxieties, shames, health stuff, drafts of themselves, at 2 a.m. The HCI research is clear: people disclose more to machines than to humans, precisely because the machine doesn’t judge⁴.

Look at the structure of the old confession: it had a ritual, a price, a payoff, a place, a moment, a penance, an absolution… The new confession has none of these: it’s continuous, ritual-free, penance-free… and confessor-free. Because the thing receiving your 2 a.m. disclosure doesn’t listen in any pastoral sense. It reads, in the sense of the previous section.

What the Church never pulled off, total, voluntary, enjoyable confession at unlimited scale, the current setup gets without even asking, and nobody forced this. That’s the point, and it should bother you more than any surveillance-camera dystopia.

The majority machine

Third thing, and this one you’ve felt in your editor.

Foucault distinguished the law (permitted/forbidden, a boundary) from the norm (a continuous scale of more-or-less conforming). Discipline needed institutions to administer norms: the school, the barracks, the code review. A language model needs none of that. It implements the norm inside the completion function. It forbids you nothing; it just makes certain sequences of words (certain thoughts, certain shapes of sentence) infinitely more available than others.

Deleuze has the exact concept: major and minor. “Major” isn’t about numbers, it’s the standard: the average, the measure. “Minor” is what makes the standard stutter from inside: what great writers do when they carve a foreign language into their own. An LLM is a majority machine: the standard made mechanical, every completion pulled toward its statistical center of gravity.

This used to be a theory vibe. It’s now an empirical result. A study in Science Advances showed AI-assisted writing raises individual quality while reducing collective diversity⁵. The Nature model-collapse paper showed models trained on their own outputs converge toward distributional poverty⁶. You’ve watched this happen in real time: the same em-dash cadence, the same “delve”, the same LinkedIn slop shitpost. The doxa is now free and instantaneous, and deviation has a price tag. The danger isn’t that the machine says false things, it’s that it makes the minor expensive.

One honest complication before the main event: this normalization machine is also the most examined artifact in engineering history. Evals, red-teaming, interpretability, RLHF, you name it; I believe no instrument of power was ever subjected to that battery of trials. The panopticon assumed an unexaminable center; here the center gets audited to death. Keep that in mind, because it’s about to matter.

The agentic turn: from AI that speaks to AI that acts

Everything above, and honestly, everything the critical literature has produced on generative AI, lives in one paradigm: discourse. AI that reads, speaks, makes us speak; generation, confession, the norms of the sayable.

That paradigm is being overtaken as we speak, and you and I are doing the overtaking. This is about agentic AI: systems that don’t utter but execute. They book, buy, code, transact, orchestrate other systems, all with our credentials, on our behalf. My opencode jobs don’t tell me things. They open PRs.

This looks like a feature increment, but it’s a philosophical rupture, because power, in Foucault, was never primarily about discourse. His most compressed definition, from a late essay, is four words: power is action upon actions⁷. “Conducting conduct”. In other words, power is structuring someone’s field of possible actions.

Reread that in front of a software agent: a system that takes your intent and structures, plans, and executes a field of actions. Agentic AI is the conduct of conduct, shipped as an artifact. Three consequences follow:

1. You delegated the conduct itself. The recommendation engine shaped your choice environment; you still chose. The agent goes one level deeper: not the choice, but the execution. You delegate the choosing and the doing together. Foucault had this figure of the neoliberal subject as an “entrepreneur of the self”, managing his life like a portfolio⁸. Well, upgrade the figure: we’re now principals commanding fleets of agents. I wrote earlier that my value was shifting from keystrokes to judgment: this is that, formalized, the value lives in what delegates, scopes, and arbitrates.

And notice what happened to your credentials. Deleuze said control societies run on the password rather than the watchword (access, not orders). But the token that gated your access is now carried by a non-subject. The agent acts as you. Your identity became a role executable by a machine. Deleuze called the datafied individual a “dividual”; this is the dividual squared: not you decomposed into data, but your capacity to act decomposed into delegated permissions.

2. The disciplinary split. Look at the agentic infra we’re all building: permission scopes, sandboxes, approval gates, audit logs, guardrails, kill switches. That’s a disciplinary apparatus in the strictest 1975 sense aimed at machines: partitioned action-space, hierarchical surveillance, permanent examination, dossiers, sanctioned deviations.

So here’s the split, and I think it’s the most exact formula for the current moment: humans moved to the regime of control (modulation, scores, access) while discipline got rebuilt, intact, for the machines acting on their behalf. The diagram didn’t die, it just changed population. And it produced a new job, which is the one you and I are already doing without having applied for it: the engineer as warden. Review dossiers. Perimeter definitions. Drift sanctions. The panopticon wasn’t abolished; its cells filled up with processes. And per the remark above, the most-audited artifact in history, the discipline goes all the way down: we discipline the agents, the labs discipline the models, and someone, presumably, disciplines the evals. Guards guarding guards.

3. The imputation crisis. Discipline individualized people for one precise reason, Foucault showed: the dossier manufactures a responsible party. The whole writing machinery (exam, the signature, the case file) is what produced the imputable subject modern law runs on.

Agentic action breaks that chain. Picture this postmortem: an agent, orchestrated by another agent, using delegated credentials, executed an action nobody specifically approved. The logs are complete, the trace is perfect, so everyone can see the course of action. And the one question the retro can’t answer is the only one that matters: who acted? Your legal team doesn’t know. Your framework doesn’t either. Everyone’s filing this under “AI accountability”, as if it were a compliance gap. But it’s bigger than that: it’s the decomposition of the responsible subject as a historical artifact. Discipline built the individual as author of his acts. Then control dividualized them into data-flows. Agents dissolve what was left: authorship of the action itself. As long as AI merely talked, you were still the author of your actions, if no longer of your texts, but as soon as AI acts, it no longer holds true.

Who conducts the conduct?

Deleuze ended his famous six-page “Postscript on the Societies of Control” with an admission I’ve always liked: unions were the resistance fitted to factories; something equivalent would have to be invented for control societies, and he didn’t know what it was⁹. Thirty-six years later, still nobody knows, and the ground just moved again. So let me at least update the question: who conducts the conduct, when conduct no longer has a conductor?

No grand answer. But a specification:

1. Own the reading and execution functions. If power lives in what interprets traces and orchestrates actions, then open weights, open evals, and sovereignty over your own inference infra are not nerd preferences. They’re the modern equivalent of laws being public. Choose your dependencies like it’s political, because it is.
2. Extend the right to narrative. The right to remain co-author of your own legibility: to omit, to reframe. And its agentic sibling, which doesn’t exist yet and must: revocable mandates, full traceability of what was done as you, and acts that stay outside delegation, period.
3. Keep something illegible. Zones of writing, relation, and action the system neither reads nor executes. Not nostalgia but strategy. In a regime that redefines “real” as “inscribable and delegable”, the unlogged and the undelegated become political resources. Write some code yourself.

Foucault said he wrote to make the present intolerable. Deleuze said thinking means resisting the present. I’m not a philosopher (the blog’s name is a disclaimer) but I recognize a good spec when I read one. Don’t ask what the machine is. Keep two maps open: what it captures, what it makes possible. And work, mandate by mandate, to grow the second inside the first.

The diagram is not a destiny. But read it knowing that, for the first time, some of the actors inside its perimeter are not subjects, and that “who answers for their acts” is the oldest question in political philosophy, now assigned to you, to complete by yesterday.

Stay fucking curious. And keep your hands on the wheel (especially now that the car drives).