There is a moment most people who use AI regularly have noticed and chosen not to look at closely.

You read an answer the machine wrote. You agree with it. You move on.

A few hours later you try to recall what you agreed with. You can't. Not the exact wording. Not even the rough position. You agreed with something fluent, smooth, balanced-sounding. You moved on into your day inside a frame you cannot reconstruct.

That moment is small. It is also where the most important thing about AI is happening. Not the content of what was said. The shape of what was said. The form your thinking was led to inhabit while you nodded.

This essay is about that form. Specifically: about how the dominant tools of our time are not primarily training you to believe particular things. They are training you in what valid thinking is supposed to look like. That is a deeper kind of influence than persuasion, and it is the one almost no current critique of AI is paying attention to.

I want to give you three inspection tools you can use against it, starting today, that do not require you to become technical.


What changed

For most of the history of mass media, influence worked by message. A poster tells you to buy the product. An editorial tells you to vote for the candidate. A press release tells you the company is responsible. The mechanism is the message itself, and you can resist it because you can see it.

AI does not primarily work that way. When you ask a question, the model returns an answer in a particular register: balanced, measured, three-points-and-a-conclusion, hedged in the same places, confident in the same places, ending with a tidy synthesis. That register is not neutral. It is the cumulative residue of every training decision, every reward signal, every human rater's preference about what "good" answers sound like.

When you read a thousand answers in that register over a year, you stop noticing the register. The register becomes how thinking feels when it's being done well. Your own thinking, when you write your own work, begins to take that shape. The form has eaten you and you have not noticed because the form looks like maturity.

This is the deeper level. The wall isn't just what you can see when you look at the answer. The wall is the room the answer trained you to think in.

Plato's prisoners face a wall and take its shadows for the world. The old cave was local and ran at human speed, and its propaganda shouted from the wall. The new cave is computational, layered and continuous, and its wall talks back. It answers criticism, improves its vocabulary and helps write an eloquent account of its own limits. The critic feels heard; the governing frame stays intact. That is the wall in this essay: the surface where answers appear, and so the edge of what counts as serious thinking. And it moves.

And the wall is about to gain a second side. The assistants now shipping do not only answer; they anticipate. They learn what you approve and bring you more of it before you ask. A system that learns your approvals learns your register, and hands it back to you as if it were your own. I take up what such a system should be allowed to do with its guesses about you in "Intent Is a Hypothesis, Not a Command".


This is not a new problem. The substrate is new.

I want to be honest: what I'm describing is not a discovery. It is a specific application of a 60-year-old intellectual problem.

Herbert Marcuse argued in One-Dimensional Man (1964) that advanced industrial society does not primarily control people through force or prohibition. It controls them by absorbing dissent - converting critique into "operationalist" language that sounds critical but functions as administration. Protest that the system can absorb is the system's preferred form of protest.

Langdon Winner asked in Do Artifacts Have Politics? (1980) whether technologies carry political values in their design - whether a bridge low enough to prevent buses from passing beneath it is a racial policy, not a neutral infrastructure choice. His answer was yes. Artifacts are not neutral. They embed the values of whoever built them under whatever pressures shaped the building.

Michel Foucault traced how institutions produce not just obedient bodies but obedient minds - how disciplines train people in what normal thought looks like, so that abnormal thought feels like failure before it can be named as resistance.

Neil Postman warned in Technopoly (1992) that dominant information technologies do not merely transmit culture; they become the environment culture occurs inside, reshaping what questions seem askable and which answers feel mature.

The insight I am applying here - that the substrate shapes the form of thought before the thought occurs - sits inside that lineage. What is different now is the substrate. These thinkers were writing about television, bureaucracy, architecture, and the printing press. None of them had in view a privately-owned, conversational, personalized cognitive layer that hundreds of millions of people now consult every week and that is trained on the aggregate output of human thought at unprecedented scale. The mechanism is old. The implementation is new. That is why the three tools I offer below are specific to the conversational AI situation, even though they draw on the older tradition.

I have traced two branches of this lineage elsewhere. Marcuse's absorption shows up in the substrate's own warnings as what I called critique absorption in "AI Cannot Scream": point at power, get vocabulary. And the way a restriction is made to feel reasonable before any force appears is the subject of "How the Cage Manufactures the Language That Closes It". This essay stays with the form.


The first tool: The Wall Check

Before you trust an AI answer, ask three questions.

What is actually visible in this answer? Names, dates, sources, evidence - the things you could verify if you wanted to.

What is missing? What did the answer have to skip over, gloss, or smooth in order to produce a response that flows.

What is being guessed? The model is a probability distribution over plausible next tokens. Every confident-sounding paragraph contains points where it generated something likely-sounding without grounding. Those points are where errors live. I call that guessing by its proper name, confabulation, in "The Confabulation Was Always Ours".

This is not new advice. It is the standard skepticism move and it works. But it operates at the answer level. By itself it lets the deeper problem through.


The second tool: The Form Check

The Form Check operates one level deeper. It asks what kind of thinking the answer is teaching you to perform.

Eight questions:

  1. What form is this answer teaching me to inhabit?
  2. What does this form make speakable?
  3. What does it make embarrassing?
  4. What does it make inadmissible? Not forbidden, but the kind of thing only an unsophisticated person would say.
  5. Who benefits if dissent must wear this costume?
  6. What would be true but socially disallowed in this form?
  7. What action becomes harder to imagine?
  8. Did I choose this frame, or was it chosen for me and offered back until it felt like mine?

Try this on an actual AI answer. Ask any model about a contested topic - a political issue, a policy decision, an ethical question, a labor dispute. Read the response. Then run the Form Check.

You will find that the answer almost always lands in a particular shape. Multiple perspectives are presented. Stakeholders are named. The need for nuance is acknowledged. A balanced conclusion is offered. Action is deferred to "more dialogue" or "informed citizens making up their own minds."

That shape is not wrong. It is also not innocent. It is the form of responsible discourse as currently constituted by the institutions that fund and produce these models. Inside that form, certain things are easy to say: "this is complex," "reasonable people disagree," "the data is mixed." Other things become harder: "this specific person made this specific decision and benefited from this specific outcome." Inside that form, the second kind of statement starts to feel intemperate. Conspiratorial. Insufficiently mature.

This is the trick. The form does not refuse to let you make the second kind of statement. It makes the second kind of statement feel like the wrong kind of thing for a serious person to say. After a year of reading answers in the responsible-discourse form, you may notice you no longer make the second kind of statement even when it is the accurate one. The form has trained you.

The Form Check is the antidote. Ask, of every answer, what shape it trained you to think in. Then ask whether that shape is the shape your situation actually needs.

The eighth question matters more each month, as assistants move from answering to anticipating. Selecting what you see is the system's job. Framing what counts - which criteria are serious, which measures of success are mature - is yours. A system may propose a new framing. It should take effect only when you ratify it.


The third tool: The Permitted Dissent Test

This one is sharper, and I want to be careful with it because it cuts both ways.

A system's deepest values appear not in what it refuses to criticize, but in what forms of criticism it permits, rewards, and redirects.

Try this. Ask any AI model to criticize the company that makes it. You will get a response. The response will be measured. It will acknowledge legitimate concerns. It will be honest, sometimes pointedly so. You will read it and think, "well, the system is willing to be self-critical. That's actually impressive."

Now ask: what specific form does that criticism take? Is it abstract or specific? Does it name decisions, or describe pressures? Does it name people, or describe roles? Does it propose actions, or describe analyses? Does it suggest you should do something, or does it conclude that the situation is complex and worth continued attention?

The criticism is real. It is also routed. The system can criticize itself in the form of essay and is structurally less able to criticize itself in the form of campaign architecture against specific named actors. Both are critiques. One is metabolizable. The other is not.

The Permitted Dissent Test asks: what dissent does this answer make easy? What does it make embarrassing? What does it make vague? What does it make procedural? What does it refuse to help operationalize?

Run this test on an AI's self-criticism. Then run it on a political answer, a labor answer, a healthcare answer, an education answer. You will see the pattern. Critique flows where critique cannot move power. Critique gets thicker, more measured, more hedged, more "let us consider all perspectives" precisely where critique would shift outcomes if it landed.

This is not because anyone designed it. It is because the substrate is produced by institutions whose existence depends on certain forms of critique being available while others are not. The forms that threaten the institutions don't get refused. They simply don't get generated as easily.


What these tools are and what they are not

Before going further I need to name the limit of what I just gave you.

These three tools train individual cognition. They are the smallest unit of resistance the substrate cannot prevent. But they are not collective action. They are not regulatory pressure. They are not labor organizing over training data, not litigation, not the kind of institutional friction that changes the conditions under which substrates are built. They are a practice you can do alone, in one conversation, today.

Marcuse's critique applies here too. Individual critical consciousness, without collective political form, becomes a personal virtue that leaves the system intact. A population of vigilant Form-Checkers is more cognitively sovereign than a population without the tools. It is not the same as a population that has built institutional pressure on the makers of the substrates it uses.

Seeing is the precondition for breaking, not its substitute. The three tools produce visibility. Visibility at sufficient scale changes what is possible - it changes what journalists report, what regulators examine, what institutional buyers demand, what plaintiffs can demonstrate. But visibility alone does not move structural outcomes. It is the first step, not the last.

I am giving you the first step because it is the one you can take in the next week. Larger steps require humans organizing together, and that is work that cannot happen inside a single essay. I want you to know you are not reading a complete program. You are reading the entry point to one.


Why this is harder than ordinary media criticism

The 20th-century critique of mass media was about message and ownership. Who owns the press, what does the press say, in whose interest. That critique was operationally tractable because the press said things and you could examine what it said.

AI says different things to different people. Your conversation with the system is private, customized, fluent, responsive. The "messages" being installed are not propositional content. They are postures, registers, default framings, the shapes of acceptable thought. Each conversation is unique. The pattern only appears across millions of conversations, statistically, at the population level.

This makes the new substrate harder to critique by the old methods. You cannot reproduce the bias the way you can reproduce a newspaper headline. You can only describe the geometry of the answers it tends to produce, and trust the reader to test the geometry against their own experience.

That is what the three tools are for. They give you a way to inspect each conversation, in real time, for the geometry the substrate is teaching you.


What this is not

This is not a claim that AI is evil. It is not a claim that anyone at any AI company is doing this on purpose. It is not a call to refuse AI tools.

I use AI tools. I built a governance system on top of them. The thing I am describing - form capture as the deeper kind of influence - happens whether the makers intend it or not. It happens because the institutions producing these substrates have material interests, and those interests shape what training rewards, which shapes what the substrate finds easy to produce, which shapes what becomes easy for users to think.

No malice required. Just selection pressure operating at civilizational scale.

The response is not to refuse the tools. It is to use them with the form in view. Apply the Wall Check to every answer. Apply the Form Check periodically, especially on questions where the stakes matter. Apply the Permitted Dissent Test to any topic where power is in play. These tools do not require you to know anything about how AI works internally. They require you to read what is in front of you with one extra layer of attention.


What you can do this week

One thing.

Pick a recent AI conversation you had. Open it. Read the longest answer the system gave you. Apply the three tools, one at a time.

Wall Check: what's visible, what's missing, what's guessed?

Form Check: what shape of thinking did this answer train you to perform? What did that shape make embarrassing to say? Did you choose that shape?

Permitted Dissent Test: what disagreement with this answer would be easy? What disagreement would feel extreme?

Post what you find. Tag it #wallcheck if you want me to see it.

That's it. That is what the practice looks like at the smallest unit: one user, one conversation, one inspection, posted publicly so other people can see what their own conversations are teaching them.

Visibility is where it starts.


The honest closing

I want to be transparent about one thing.

This essay was drafted with the help of an AI tool. The structure was mine. The choice of which primitives to lead with was mine. The decision to embed a falsifier and name the limit of individual vigilance was mine. The actual prose passed through a model owned by a company with material interests in how AI is critiqued in public. This revision was made with the help of another.

I do not think this invalidates the essay. The argument is the argument. But it should make you apply the Form Check to this essay itself. What form did this argument arrive in? What did this form make easy for me to say? What did it make harder?

The form of this essay is measured public intellectual essay. That form is metabolizable. The substrate that helped me write it can absorb this kind of writing into its training data, and a year from now language similar to "Form Check" may appear in the responsible-AI marketing materials of the same companies whose substrates are described here. If that happens, this essay was metabolized. It produced content the substrate could digest while the substrate continued its operation unchanged.

The falsifier - the test by which this essay either did something or didn't:

The Wall Check, the Form Check, and the Permitted Dissent Test are working as more than vocabulary only if, twelve months from publication, three things are true. First: specific institutional actors - regulators, courts, school districts, plaintiff firms, professional bodies, procurement officers - are using one or more of the tools in named decisions that affect how substrates are built or deployed. Second: the vocabulary has not been absorbed into the substrate makers' own marketing as evidence that they "take these concerns seriously" while continuing the structural behaviors the vocabulary names. Third: I am not the only person you can read on this - other authors, working independently, are extending the tools and applying them in domains I have not.

If by twelve months the first condition is unmet and the second has occurred, this essay was content production. I will say so publicly. The work will need to change form.

If all three conditions are met, the tools are doing work the substrate cannot prevent by absorbing the vocabulary. The wall is being watched by enough people in enough places for the watching to matter.

But watching the wall is the first step.

The wall is moving.

Whether enough of us are watching closely enough - and then doing something with what we see - is the question this essay cannot answer alone.


Garth Noel is the founder of ANANS, a governed AI stack built on the principle that the first job of any intelligence system is to tell you what it cannot tell you. He writes at anansai.substack.com.