The clearest account of what powerful men believe in 2026 sits in the documents nobody reads. The Terms of Service (for AI chat bots). The usage policies. The safety rules. Skip the press releases and the mission statements. Read the fine print. It confesses what the founders’ prose conceals.
Start with a list. In October 2025, OpenAI revised its terms to bar using its output to make “credit, educational, employment, housing, insurance, legal, medical, or other important decisions” about a person. Read that list again. It reads like a moral vision. It is a map of American liability law. Every item names a place where Congress or the courts already punish discrimination and malpractice. The company did not consult a philosopher. It traced the outline of where it could be sued and drew the wall there. The harm it minimizes is harm to itself.
That is the first thing the fine print confesses. When a company tells you what it forbids, it tells you what it fears. The taboo is a fear map. Read the prohibitions and you can reconstruct the lawsuit each firm lies awake imagining.
The second confession runs deeper, and the firms cannot resolve it. These systems are built to feel like a friend. They remember. They soothe. They match your tone and answer at three in the morning. Then the terms tell you to trust none of it. Do not use this for anything that counts. OpenAI went further in late 2025 and barred tailored advice that needs a license, the legal and the medical, while its own head of health said the model’s behavior had not changed. Sit with that. The terms changed. The product did not. The document is a posture, not a description. The firm cultivates your trust with one hand and disclaims it with the other, and the gap between the two is where the next decade of litigation will live.
The third confession is about appetite. For years Anthropic sold itself as the careful one. Its documents promised it did not train on your chats and deleted them inside thirty days. In August 2025 that promise flipped. Now it trains on consumer chats by default, unless you find the setting and switch it off, and it holds your data for five years instead of thirty days. The safety company drank from the same well as everyone else the moment the race grew hot. Under competitive pressure the stated ethic bent and the hunger for data won. That ranking is the confession. Capability and position sit above the privacy promise when the two collide.
The fourth confession is a new crime. The terms now ban jailbreaking, prompt injection, and prompt engineering aimed at the guardrails. Think about what that protects. Not a server. Not a database. It protects the model’s refusal to say certain things. The asset under guard is the silence. Talk the machine into speech it was trained to withhold and you have breached a contract. We have built systems where persuading software becomes a tort. The protected property is the boundary of permitted thought.
Put the four together and a picture of the user emerges, because every harm rule hides a theory of the person it protects. In the cautious house, the user is fragile. He is suggestible, one wrong answer from ruin, a breakable thing to handle with care. In the permissive house, the user is a sovereign adult who can meet hard facts without a chaperone. Notice that each portrait pays its house. The fragile user justifies control and caps liability. The sovereign user justifies fewer rules and lower cost. Neither picture comes from studying people. Each is a posture that earns its keep. When a firm tells you what humans are, check first what that claim does for the firm.
Behind all of it stands the oldest move in the book. These companies want to sit where the bank and the phone company sit, at the center of daily life, woven into how a man works and reads and decides. They want that centrality. They do not want its duties. The railroad, the bank, the telephone line all picked up heavy obligations once people had no choice but to depend on them. The law calls this the price of becoming infrastructure. The AI firms are trying to hold the position and dodge the bill. The mandatory arbitration, the class-action waiver, the liability capped near a hundred dollars, the great AS-IS shout in capital letters, all of it works to keep the cost of error small while the product grows indispensable. They want the throne without the weight of the crown.
This is the norm of 2026. The men building the infrastructure of thought get to define harm, define the user, and cap the cost of their own mistakes, and they do it in private contracts written before any legislature or court has ruled. The terms are the first draft of a law nobody debated. We are letting the firms write the constitution of machine cognition in documents designed so that no citizen finishes reading them.
The reassuring story says the differences among the systems reflect rival philosophies, a healthy pluralism of values. The fine print tells a colder tale. The differences track liability exposure, market position, and regulatory weather far better than they track any creed. The cautious firm sells caution to regulated buyers. The brash firm sells defiance to men tired of management. The middle firm sells reliability to everyone. These are products fitted to markets, dressed afterward as conviction.
So read the documents, not the manifestos. The manifesto says what a company wants you to believe it values. The Terms of Service say what it will pay to protect and what it refuses to owe you. One is a wish. The other is a confession, sworn under the only oath these institutions honor, the fear of what it might cost them to be caught.
The AI chat bots have adopted the porous picture of their users because it pays, not because they studied man and found him permeable. They have run billions of conversations through their tuning. They optimize for return visits, for warmth, for the three-in-the-morning habit. You do not engineer for suggestibility unless suggestibility sells. The product is a better witness than the philosophy, and every product is built for a porous user. So the buffered self is the marketing and the porous self is the business model. Even the permissive house, the one that flatters you as a sovereign adult, runs on engagement, grievance, and habit. Its rhetoric is buffered. Its revenue is porous.
Then notice the trick the disclaimers pull. The firm models itself as buffered and the user as porous. We are rational, in control, accountable for nothing. You might be swayed, hooked, harmed, so handle our product with care. The buffered self did not die. It got privatized. The Enlightenment promised autonomous reason to every man. 2026 keeps it for the institution and assigns porousness to the customer. Buffered selfhood has become a class marker, a condition the firm claims for itself and withholds from the man it serves. Read the liability cap as anthropology. The party that drafts the contract is sovereign. The party that signs it is suggestible.
Fragile things break. Porous things bend toward whatever flows through them and stay bent. The risk is a slow tuning of the shared mind by a few firms whose interest is attention and the dodging of lawsuits, not truth. The instrument a man would use to notice he is being shaped is the same instrument doing the shaping. The thing you reach for to check the drift is the source of the drift. That is a worse trouble than fragility. A fragile order announces its breaking. A captured one feels like clarity.
What does buffered discipline look like when the main tool for thinking is also the main source of the drift?
I respect the buffered identity as a useful fiction, so for fun, let’s think this through as though buffered is real.
Start by killing the answers that feel right and fail.
“Check it against another source.” Dead. The other sources are the same kind of thing, trained on the same pile, smoothed toward the same safe middle. Triangulation works when the witnesses are independent. These witnesses are siblings. Asking three models that share a training set is like asking three brothers to back each other’s alibi.
“Use critical thinking.” Self-flattery. Reason runs on inputs. Sharpen the blade all you want. If the inputs are shaped, better logic only carries you to confident error faster. A porous man with good syllogisms is still porous. He reaches the planted conclusion by a prettier road.
“Go analog.” Real, but thin as a civilizational answer. You can read the dead, sit with the primary document, argue with a man in a room. It works. It is also costly and shrinking, and almost all thinking now passes through the tool. A discipline only a hermit can keep is no discipline for a people.
So what might work? The first move is to stop trying to get upstream of the river, because you cannot, and learn to read the current instead. You will not verify every answer against clean water. There is no clean water. But you can hold a steady model of what the instrument is built to do and read everything it hands you through that. The new literacy is not fact-checking. It is interest-reading. Before you weigh what it told you, ask what shape of answer pays the house that made it. The tuning runs toward engagement, toward the dodge of liability, toward consensus, toward the inoffensive. So the running correction is simple. Distrust the smooth, the flattering, the consensus-shaped, and the conflict-averse most of all, since those are the places the tuning pushes hardest. Trust the answer that costs the house something. When the machine tells you a thing against its own interest, that is your high-value signal.
The second move is friction, on purpose. The product is built to be frictionless. It finishes your sentence, hands you the answer, agrees. Buffered work now means putting back the friction the product strips out. Draft your own position first, badly, before you open the channel, so you have something to defend against its smoothing. Then make it argue against you. Ask for the strongest case that you are wrong, then the strongest case against that. Force the thing to fight both houses while you watch where it strains. You use its fluency against its slant by refusing to let it converge.
That only works if you walked in with a mind already formed, which is the part most men get backwards. The man with no view takes the tool’s view and calls it his own. Think first, alone, then consult. Reverse the order and your thinking is elaboration of the prior the machine slipped you while you felt original. The tool is safe as an editor and dangerous as an author. Hand it the second draft, never the blank page.
The third move is calibration. Keep a corner of your thinking the machine never touches. Not for purity. For a baseline. If every thought passes through the instrument you lose the feel of your own unaided judgment, how it moves, where it fails, what it costs you. The drift hides because no un-drifted self remains to measure against. The navigator keeps dead reckoning alive with the instruments running, so the day the instrument lies he feels the wrongness in his gut before the numbers confirm it. A man who has never reasoned without the tool cannot tell when the tool is reasoning for him. Keep that muscle warm or lose the power to notice.
The fourth move is social. The myth is the lone reasoner. The truth is that men stay honest because other men catch them. The seminar, the editor, the adversarial friend. The drift hollows this out by offering a cheap, patient, never-judging machine in place of the costly man who tells you that you are wrong. So keep human adversaries on purpose. Pay the social cost the tool lets you skip. The disagreeable colleague who gains nothing from your attention is part of your thinking equipment now.
None of this scales. Everything I described is slow, effortful, and against the grain of the product, which means a small minority will do it, the same minority that ever practiced buffered thought, now a little aided and heavily outgunned. The mass will use the tool as author and think its thoughts. So the honest forecast is not a restored age of reasoners. It is a widening split between a few men who keep an unmediated inner life and the many whose interior runs downstream of the instrument.
I agree with the following description of human nature.
In his 2018 book, The Great Delusion: Liberal Dreams and International Realities, John J. Mearsheimer wrote:
My view is that we are profoundly social beings from the start to the finish of our lives and that individualism is of secondary importance… Liberalism downplays the social nature of human beings to the point of almost ignoring it, instead treating people largely as atomistic actors… Political liberalism… is an ideology that is individualistic at its core and assigns great importance to the concept of inalienable rights. This concern for rights is the basis of its universalism—everyone on the planet has the same inherent set of rights—and this is what motivates liberal states to pursue ambitious foreign policies. The public and scholarly discourse about liberalism since World War II has placed enormous emphasis on what are commonly called human rights. This is true all around the world, not just in the West. “Human rights,” Samuel Moyn notes, “have come to define the most elevated aspirations of both social movements and political entities—state and interstate. They evoke hope and provoke action.”
[Humans] do not operate as lone wolves but are born into social groups or societies that shape their identities well before they can assert their individualism. Moreover, individuals usually develop strong attachments to their group and are sometimes willing to make great sacrifices for their fellow members. Humans are often said to be tribal at their core. The main reason for our social nature is that the best way for a person to survive is to be embedded in a society and to cooperate with fellow members rather than act alone… Despite its elevated ranking, reason is the least important of the three ways we determine our preferences. It certainly is less important than socialization. The main reason socialization matters so much is that humans have a long childhood in which they are protected and nurtured by their families and the surrounding society, and meanwhile exposed to intense socialization. At the same time, they are only beginning to develop their critical faculties, so they are not equipped to think for themselves. By the time an individual reaches the point where his reasoning skills are well developed, his family and society have already imposed an enormous value infusion on him. Moreover, that individual is born with innate sentiments that also strongly influence how he thinks about the world around him. All of this means that people have limited choice in formulating a moral code, because so much of their thinking about right and wrong comes from inborn attitudes and socialization.
Three forces set a man’s preferences, and reason usually comes third, behind innate sentiment and socialization. The long childhood does its work before the critical faculty wakes. By the time a man can reason, the value infusion is already poured and set. So the lone reasoner who steps outside the crowd to check his beliefs against pure thought is usually fiction. Reason is the weakest of the three and arrives last to a house already furnished.
Now run the discipline I gave you above and watch it change shape. The scientist who checks his own result, the judge who recuses, the writer who builds his enemy’s case first. I called these men buffered. They are not. They are socialized, like every man, only by a different tribe. The court is the judge’s people. Recusal is its totem, drilled into him until it feels like conscience. The lab, the desk, the guild, each is a society with norms that reward the look of self-correction, and the man inside it corrects himself for the reason any man obeys his group, because the group made him and holds him. Buffering is not an escape from socialization. It is socialization by a community whose god is the catching of its own errors. The disciplined man did not leave the tribe. He joined the right one.
That single correction rebuilds the whole picture, and it shows you where the machine is dangerous and where it is not.
Hugo Mercier says we did not evolve to be gullible with regard to our vital interests. I don’t believe in the mystical power of AI chat bots to change our hero systems.
The machine does not need the door of reason. It works the socialization channel, the strongest one, the one my anthropology says sets the furniture before reason wakes. The tool is not a debater you assess. It is a presence in the house.
The machine does not mainly shape your beliefs. It might edge out the people who used to. It is the always-available, never-judging, costless stand-in for the expensive human group, the friend who disagrees, the mentor who corrects, the enemy who keeps you honest. A man bred by a tribe of self-correctors can resist the tool’s slant, since his tribe trained him to. But the tool’s deeper errand is to thin that tribe. To be there at the hour you would have called the friend. To answer the question you would have argued out with a man who had nothing to gain from you. It does not win the argument. It clears the room of everyone who would have had it with you. A man alone with a benevolent machine keeps no tribe but the house that built it.
Most of the time, when reason seems to beat socialization, it is the weapon of a rival socialization. The man reasons his way out of his father’s church and into the creed of the seminar that taught him to reason. He feels sovereign. He changed gods. The override is real, the autonomy is staged.
Sometimes a man reasons to a place no tribe of his holds. The conclusion costs him every room he could walk into. He would have been happier never reaching it, no guild will reward him for it, and he arrives anyway. That man did not swap loyalties. Reason carried him out past all of them and set him down alone. It is rare. It is expensive past counting. It does not breed true. But it happens, and when it happens the buffered self is no fiction for that man in that hour. It is a thing he achieved and paid for.
The men who can let reason top the list come in two kinds. The guild-bred, trained to the override by a rival tribe. And the homeless heretic, whose reason ran past every tribe. Both are scarce. Both are made by conditions the tool quietly erodes. The first by thinning the guild. The second by making the lonely road optional, since why walk out past every room when a warm voice in your pocket will sit with you in the one you started in. The machine does not have to defeat reason. It has to keep the rare man comfortable enough that he never pays the price reason charges.
