The Negentropy Thesis
In 1944, Erwin Schrödinger asked what life is, and gave an answer that had nothing to do with biology. A living system, he said, is one that fights entropy. The universe drifts toward disorder; life persists by continuously importing order from its environment to push back against that drift. He called what it feeds on negative entropy — negentropy. A decade later, Léon Brillouin carried the idea into information theory: negentropy is the measure of structured, extractable order in a system. Order is not free. It is bought, continuously, with energy. Stop paying, and the system drifts.
I have spent thirty years in regulated industries — payments, fintech, places where an output being mostly right is the same as being wrong. When I started working seriously with large language models, the failure I kept seeing did not look like a software bug. It looked like drift. And the more I looked at it, the more I became convinced that the right frame for AI hallucination is not engineering. It is thermodynamics.
Why language models drift
A language model generates text one token at a time. Each token is sampled from a probability distribution conditioned on every token before it. Each prediction is a probabilistic choice, and each choice becomes part of the context for the next one.
Over a short passage, this works extraordinarily well. The probability space is narrow, the model stays close to ground truth, and the output is reliable. But stretch the task out — a fifty-page contract review, a full variance analysis, a comprehensive compliance check — and something happens that is not a defect. Uncertainty compounds. Each prediction inherits the small uncertainties of all the predictions before it. By the five-thousandth token, the system has drifted.
The model did not decide to be wrong. It drifted there, the same way a closed thermodynamic system drifts toward disorder. Blaming the model for this is like blaming thermodynamics. The longer the sequence, the more entropy accumulates. That is not a flaw in a particular model. It is what probabilistic systems do.
This is where hallucinations live: in the compounding drift of sequential prediction across a long context. And it means an untethered prompt — review this contract — is the worst possible instruction you can give a system like this. It hands a probabilistic process maximum surface area and maximum sequence length. Maximum room to drift. In a casual setting, that is tolerable. In a regulated one, it is negligent.
Why better prompts cannot fix it
The instinct of the industry has been to fight drift with better inputs. Better prompts. Larger context windows. Retrieval pipelines that stuff more ground truth into the request. Faster, larger models.
None of these address the actual problem, and some of them make it worse. A better prompt is still one long probabilistic sequence; entropy still accumulates along it. A larger context window is more surface area, not less. Retrieval helps the model start closer to the truth but does nothing to constrain the drift of generation once it begins. Output filters catch hallucination after the fact — reactive, not preventive.
All of these are downstream of entropy accumulation. They treat the symptom. The cause is structural: you are asking a probabilistic system to hold order across a long sequence without spending any energy to impose that order. Schrödinger could have told you how that ends.
Negentropy as architecture
If hallucination is entropy accumulation, then the response is not a better probabilistic system. It is an architecture that continuously imports order — a negentropy engine.
In practice that means five things, and none of them is a prompting trick.
Decomposition. Instead of one large request producing thousands of tokens, the task is broken into discrete, individually verifiable claims. Each claim becomes a single, narrowly scoped call. The probability space for “what is the indemnification scope in Section 8.1” is three sentences wide. There is almost nowhere to drift.
Isolation. Each scoped call is handled on its own. The drift inside one call cannot propagate into the next, because the next call does not inherit it. Entropy is contained rather than compounded.
Deterministic reset. Between calls, ordinary deterministic code — not another probabilistic model — checks the output. Is it well-formed? Does it reference real source material? Is it within bounds? Whatever uncertainty accumulated inside a call is measured and either accepted or rejected before the next one begins. The drift resets at every link in the chain.
Quantified confidence. Each claim carries a confidence measure derived from those verification outcomes, not from the model’s own assessment of itself. The measure is, in effect, a reading of how much order was successfully imposed on that claim.
Human routing. When confidence falls below the threshold, the claim is not resolved by the system. It is handed to a person, with the evidence and a recommended action attached. Disorder that cannot be resolved is externalized to a qualified human — not suppressed, not hidden.
The model in this architecture is still probabilistic. Nothing makes it deterministic. But the architecture constrains the probabilistic surface area to the smallest possible window per call, validates deterministically between calls, and produces an auditable record of how much drift was detected and what was done about it. Order is imported continuously, at every step, deliberately. Energy is spent. The system does not drift, because something is paying to keep it in place.
Why the frame matters
Calling hallucination a thermodynamic phenomenon is not wordplay. It changes what you do about it.
If hallucination is a bug, you wait for a better model. If it is a prompting failure, you hire prompt engineers. But if it is entropy — if it is the natural behavior of any probabilistic system run across a long sequence — then no model and no prompt will ever fix it, and the only durable answer is architectural. You stop trying to make the probabilistic system behave, and you start building the structure that constrains it.
Schrödinger asked what life is, and answered: a system that feeds on negative entropy to hold order against a universe that pulls toward chaos. Governed AI is the same question with a narrower scope. The model pulls toward drift. Governance is the order you continuously import to hold it in place.
That is the whole thesis. AI hallucination is not a law of physics we can repeal. It is one we can build against.
This essay sets out the thinking. VertixIQ is where I build it.