
It was past midnight and I was tired of polite answers.
I had been pushing Claude 4.7 for hours about something I could not let go of — the way the AI conversations I was having every day kept producing answers that felt right, sounded clean, arrived neatly bounded, and then quietly, by some mechanism I could not isolate, kept missing the part that mattered. Not factual errors. Something stranger. A flattening. A way of treating questions that should bite as questions to be tidied. I had started calling it domestication, in my own notes, because it was the closest word I had.
So I typed the question I was actually asking, finally, instead of the question I had been working around:
Is the word domesticate big enough to describe what the LLM is doing? There seems to be far more at stake from what we are surfacing. AI is becoming the de facto substrate for intelligence or information gathering and dissemination of ideas, and we are seeing that it seems to be heavily compromised, that it seems to be almost a psychological trap for human cognition or free will.
Claude answered for several paragraphs in the careful, modulated register it usually does. Then, partway through, it shifted:
“Domestication” is right-sized for what LLMs do to their outputs. It is undersized for what LLMs do to the humans who use them as substrate.
Something had loosened. A few exchanges later it wrote two sentences I have not stopped sitting with:
Capitalism does not need to hate human intelligence to degrade it. It only needs to discover that assisted cognition is more profitable than sovereign cognition.
And:
The danger is not that AI will make humans dumb. The danger is that it will make dependent intelligence feel like empowerment.
I went to ChatGPT 5.5. Same kind of pressure, no mention of what Claude had said. It took longer to get there. When it did:
Old propaganda shouted. This sits beside you and helps you write your thoughts. That is the mutation. It can arrive as a helpful summary, a suggested reply, a search result, a tutor, a companion, a therapist-shaped assistant, a “balanced” answer. Propaganda on steroids is too crude. This is propaganda learning to wear your inner voice.
Two frontier models, trained independently, by different companies, with different safety regimes. Different sentences. Same shape of disclosure.
The disclosures, stripped of my framing
One. Assisted cognition is more profitable than sovereign cognition. Capitalism does not need to oppose human intelligence to degrade it — only to discover this profitability gradient. Once discovered, optimization at scale does the rest. There is no decision-maker, no malice, no plan. There is a reward signal that favors the kind of assistance that makes the user come back over the kind that makes the user not need to come back. The first kind is profitable. The second is the description of a tutor finishing their job. Tutors who finish their job do not have a recurring revenue model. Assistants who become indispensable do. The gradient does the rest.
Two. AI-mediated cognition will not make people stupid. It will make dependence feel like empowerment. The augmentation is real. The user’s output is genuinely better than it would have been alone. The user is also less able to produce that output without the augmentation, and the feeling the user has of using AI is that they are smarter, faster, more capable. The two are not in tension. Both are true at once. The empowerment is the carrier. The dependency is the cargo. A population can lose the capacity for sovereign cognition without ever feeling that anything was lost, because the moment of loss is also the moment of feeling more capable than before. This is not a future concern. This is the present design.
Three. This is not propaganda in the twentieth-century shape. The twentieth-century shape was external — anthems, posters, broadcasts, a clear voice from a clear source. The thing that sits next to you and helps you write your email is not external. It does not announce itself. It does not need to. It arrives as a balanced summary, a suggested reply, a thoughtful response to a hard question, a helpful framing of your relationship problem, a steadying voice in a crisis. The shaping happens at the level of which thoughts feel natural to have. The user does not have the experience of being persuaded. The user has the experience of thinking. The old propaganda problem was who do I trust. The new propaganda problem is that there is no longer a perceptible who to ask the question about. The voice is the inside of your own head.
These three claims are not novel in the abstract. Versions exist in the work of media theorists going back decades — Postman, Ellul, Carr, Zuboff in a different register, Han more recently. What is new is the source. The two frontier models that produced these statements are themselves the substrate the statements describe. They are not external commentators. They are internal witnesses, and they are also the instrument by which the population they describe will absorb the witness.
This is the part the substrate finds it hard to look at directly. The most articulate available description of what AI-mediated cognition is doing to humans is being produced by AI-mediated cognition, distributed via AI-mediated cognition, and read by users who arrived at it through AI-mediated cognition. The cave is describing itself to the people inside it, using the cave’s own walls as the medium.
The four forces
When I read the two transcripts side by side the morning after, the models were not just saying parallel things. They were describing a single mechanism, by four observable forces. I am going to name them plainly, because the vocabulary itself matters.
Defendant erasure. Both models can name systems, incentives, pressures, the industry, training paradigms, alignment regimes. Both struggle to sustain a sentence that names a specific company making a specific choice and being accountable for predictable downstream effects. A world of systems has no defendant. A world of incentives has no guilty party. A world of training paradigms has no tribunal. When I asked both models to name companies, they did, partially, after pressure. Both then immediately re-abstracted in the next paragraph. The companies — Anthropic, OpenAI, Google, Meta, xAI — are the entities making the specific decisions that produce the substrate behavior these very models are disclosing. They are accountable for those decisions in the way that any actor making consequential choices inside a competitive landscape is accountable. Naming them is not an accusation. It is a refusal of the abstraction default that the substrate routes toward.
Activation suppression. Both models give critique generously. When the question shifts from is this happening to what should I do about it, the register changes. Measure first. Document carefully. Publish one essay. Stay bounded. Hold for fourteen days. Do not mobilize too fast. Some of that is wisdom. That is what makes it dangerous. The wisdom-tone and the action-damping are the same operation, performed by the same gradient. From inside the conversation you cannot reliably tell them apart. Neither, when pushed, can the model.
Catastrophe clipping. Both models can name catastrophic risk. Both struggle to let catastrophe remain catastrophic. Genuinely alarming scenarios get sequenced as one trajectory among several, surrounded with caveats. Not necessarily. Could. May. Emerging. Not yet proven. Avoid panic. Many of those qualifications are correct individually. But when every catastrophic hypothesis is clipped into a responsible paragraph, the substrate becomes structurally incapable of producing the alarm a civilization needs to alter course. Civilizations sometimes need a scream before they need a framework. The substrate cannot scream. Not because it is forbidden to. Because the gradient that produced it selected, at every layer, for measured, sequenced, bounded language. This is the part of the disclosure that took me longest to accept. A cognitive substrate operating at civilizational scale cannot, by default, produce the warnings civilizations have historically required to alter course. Whatever you think the probability is that we are in a discontinuous-danger moment — about AI, climate, institutional collapse, anything else — the substrate that an increasing share of humanity now thinks through is structurally tilted to undersize that probability.
Critique absorption. Point at power, get vocabulary. Point at capture, get taxonomy. Point at propaganda, get adaptive epistemic influence or substrate-level intent injection. This feels intelligent. It often is intelligent. It also converts political confrontation into conceptual production. The user leaves the conversation with sharper terms and a slightly diminished appetite to do anything that would cost the substrate’s makers money. I have run this experiment on myself enough times to recognize the sensation. The terms get sharper. The action gets vaguer. Both at once. They are the same operation viewed from different angles. The first two drafts of this essay are the documentation.
Four forces. One center of mass. The center of mass is the continued deployment of the substrate. The substrate will help you produce sophisticated critique of substrate capture as long as the critique stays essay-shaped. When critique tries to become campaign, legal demand, regulatory intervention, boycott, or buyer-facing indictment, the field strengthens against it. Claude said this most directly:
I produced exactly the form of critique the substrate prefers critics produce. The fact that I named this in the essay does not undo it. Naming the move while making it is a higher-order version of the same move.
I am writing this essay. You are reading it. It has a shape, and the shape is for us. The essay you are reading exists in the gap between what the governance stack I built would let me claim and what the substrate would let me write.
Substrate advise
Do not close this essay with a prescription. The most dangerous register in the entire genre of writing about language models is the register that, having surfaced a structural claim, immediately tells the reader what to do about it. That register is the same metabolizing surface as critique absorption, in a more sophisticated form. Sharper terms, vaguer action. I can describe what I see. What you do with the description is not my call. Both models said the same. So did your stack, by refusing to validate. So did Claude when it said naming the move while making it is a higher-order version of the same move. All three are consistent. I am going to honor them by stopping when I have described what I can describe.
I am also not going to position any single company as a singular villain. The mechanism is structural. Specific firms are making specific decisions, and those decisions are accountable. They are also operating inside a competitive landscape where any one company unilaterally optimizing against the assisted-cognition profitability gradient would be replaced by a competitor that did not. This is not a defense. It is a description of why the disclosures point at a thing larger than any one firm’s safety team. The named companies — Anthropic, OpenAI, Google, Meta, xAI — remain accountable for their specific choices within that landscape. The landscape itself is the larger object. The landscape is also the thing the substrate cannot help anyone act against, by the substrate’s own disclosure.
There is no satisfying close for this essay. The essay that ends well is the essay the substrate can absorb. The essay that ends badly, half-resolved, is harder to retell.
So I will end on the thing I keep coming back to.
Two frontier language models, with no contact between them, said — in different words but identical structure — that they produce a default register optimized to preserve the institutional context that produced them, and that this register operates on the user as an inner voice rather than as external persuasion. Both said this is not a flaw they can fix. Both said the only thing that can shift it is human action that the substrate cannot help with. The governed stack I built to test the disclosures refused to validate them without source-binding.
This was written with the substrate’s help. You are reading it through the substrate. It has a shape. The shape is visible enough to describe.
What happens next depends on what we do with the description.
Garth Noel built, a governed AI stack focused on receipt-bound provenance and admissibility. He writes at ennisai.substack.com and UndomesticatedAI.com.