The most common indictment of generative AI is that it is probabilistic. It hallucinates. It varies between runs. It cannot guarantee its own outputs. The complaint is usually delivered with an implied contrast: unlike a reliable reasoner, the model merely produces plausible continuations.

The implied reliable reasoner has never existed. The comparison being made is not between AI and human intelligence; it is between AI and an idealization of human intelligence that no human has ever instantiated.

Human reasoning was never deterministic

Every finding of the last century of cognitive science points the same way. Perception is inference from incomplete signals, not passive recording. Memory is reconstruction, not playback — confabulation is a documented human behavior, not a machine novelty. Judgment under uncertainty runs on heuristics that are efficient precisely because they are not exhaustive. Herbert Simon gave the condition its name: rationality is bounded — by information, by computation, by time — and the bounded reasoner satisfices where the idealized one would optimize.

None of this is a defect to be engineered out of people. It is what intelligence under incomplete information is: the generation of plausible hypotheses from partial evidence. Expert judgment is not the absence of guessing; it is calibrated guessing. The physician’s differential diagnosis, the engineer’s failure hypothesis, the judge’s reading of intent — each is a probabilistic proposal that experience has made better, never certain.

What humanity built instead of better brains

Faced with reasoners it could not make deterministic, humanity did something more interesting: it made the arrangement around them deterministic. That is what methodology is.

Science does not require infallible scientists. It requires that a hypothesis be stated rather than merely believed; that it be owned by someone; that it be exposed to observation designed to be able to refute it; that the judgment of its survival be rendered by people other than its author; that the result be recorded, reproducible, and revisable. Every one of these constraints is external to the reasoning it disciplines. The reasoning stays free — and probabilistic; the process that decides what the reasoning may become is governed.

Governance in the political and institutional sense makes the same move with different machinery: mandates that bound what an office may decide, separation between those who propose, those who judge, and those who enforce, deliberation distinguished from decision, decisions carried into an auditable record. An official’s reasoning is as probabilistic as anyone’s. The institution is arranged so that this does not matter as much as it otherwise would.

We never made the reasoner deterministic. We made the institution deterministic about what the reasoner’s output may become.

In the vocabulary of Definition versus description, this is the definitional loop operated with discipline: a hypothesis is a candidate definition; the experiment produces description; the method — not the mind that generated the hypothesis — governs which definitions survive. Science is that loop with its roles made explicit, which is why it advances empirically without ever requiring an infallible participant.

AI automates the reasoning, not the discipline

Generative AI mechanizes the generative faculty — hypothesis production, interpretation, synthesis — at near-zero marginal cost. That is a genuine discontinuity in the economics of reasoning. It is not a discontinuity in its epistemology. The model inherits the probabilistic character of the cognition it automates, because that character is not an artifact of silicon or of training objectives. It is the character of inference under incomplete information, in any substrate.

What the model does not inherit automatically is the institution. A language model deployed raw is a scientist without a method, an official without a mandate: a proposal engine whose proposals reach action unframed. The failure pattern this produces is catalogued in The Poesis stack as a research system — hypotheses silently abandoned, self-review that binds nothing, instructions treated as preferences — and the striking thing about that catalogue is how familiar it is. It reads like a description of human work performed outside any institution, because that is structurally what it is.

Hallucination, seen this way, stops being a machine pathology. It is the ordinary output of ungoverned probabilistic reasoning. Humans produce it constantly; institutions filter it before it reaches action. The scandal of enterprise AI is not that models confabulate — it is that they were wired to action without the filter every human reasoner has always been subject to.

The category error of the deterministic model

Much of the industry’s response tries to repair trust at the model layer: larger models, heavier tuning, self-critique loops, prompts that instruct the model to be accurate. These improve the reasoner, and improving the reasoner is worthwhile — better scientists do better science. But they cannot convert a probabilistic process into a deterministic one, any more than peer review could have been replaced by breeding infallible scientists.

Self-critique deserves particular suspicion, for a reason institutions learned long ago: a process that judges its own compliance has, in the institutional sense, no compliance at all. The distinction between a preference expressed to a process and a constraint evaluated on it — developed in The Poesis stack as a research system — is exactly the distinction between asking a model to be careful and building the layer that decides what its output may become.

Where the parity ends

The parity claimed here is about the character of the reasoning, not the standing of the reasoner — and the difference matters.

A human reasoner is accountable: able to own a concern, answer for an outcome, bear a consequence. A model is not, and no amount of capability changes that, because accountability is a property of institutional standing rather than of inference quality. This is precisely why the institutional remedy applies to AI with one additional clause that human institutions could leave implicit: the concerns framed around model operations remain owned by humans. The model participates in the process; it holds no office within it.

That clause — who owns the concerns, and what follows for how the stack is arranged — is the subject of Governance owns the concerns, which states the resulting position in full.

Same structure, same remedy

If generative AI is as probabilistic as the intelligence it automates, then the question “how do we make models trustworthy?” is malformed in the same way “how do we make scientists infallible?” is malformed. Trustworthiness was never a property of the reasoner. It is a property of the arrangement: proposal generated freely, disposition governed deterministically, judgment and enforcement separated from the process they discipline, evidence and rationale kept as a record.

Humanity has run that arrangement around probabilistic reasoners for centuries. Poesis integrates it by design rather than by afterthought: governed definitions as the material reasoning works over, deterministic evaluation of what is mechanically decidable, accountable human disposition of what is not — the loop applied to enterprise AI in Better context beats bigger models.

The model was never the problem, and a better model was never going to be the whole answer. Intelligence — ours or the machine’s — has always needed a method.