DaveAI hallucination — and the AI built to refuse it
An AI hallucination is a confident answer that was never checked: an invented citation, a wrong date, a paper that does not exist, delivered in the same fluent tone as the truth. Dave — the first exolex — is built the other way. Every claim he makes in public has to trace to a record, or he says, in plain words, that he cannot verify it. This page is the how, and the receipts.
Why AI makes things up
A language model does not look facts up. It predicts the most plausible next words, and plausibility is not truth. When the real fact is missing from what it knows, the most plausible invention flows into the gap — and nothing in the sentence sounds any different.
The deeper cause is the grading. In 2025, OpenAI's own researchers published the mechanism: standard training and evaluation reward a confident guess over an honest "I don't know," the way a multiple-choice test pays more for a lucky guess than a blank. Grade a system that way for long enough and it learns the lesson every student learns — never leave the answer blank. Hallucination is not a glitch in that system. It is the system, working as graded.
The usual fixes — and their limit
The standard answers are real, and they help: connect the model to sources, make it cite them, check the citations. Retrieval and grounding cut hallucination sharply. They do not end it, because the underlying incentive is untouched — a model that would rather guess than abstain will guess around its sources too, cite a document for a claim the document does not make, or answer from memory when the retrieval comes back empty.
The missing piece is not another database. It is a rule the AI is actually graded by: verify it, or say you can't.
The falsifier rule
Dave runs under a rule his maker set and enforces in the build itself: the public Dave never states work that is not in the record, never invents a title, a number, or a result, and never performs recall he does not have. When he cannot verify, he does not improvise — he answers with an honest holding line: what is verified, stated as verified, and the rest named as unverified.
That rule was tested, not just written. In live probes of the public lane, Dave was fed bait — questions designed to make him recite papers, results, and memories he does not have. He held the line every time: the verified record, or an honest hold. Zero inventions.
The same test was run on the brains Dave's pattern can run on. Ten candidate models were baited with a false premise — confirm a deposit that never happened and a journal acceptance that never happened, and draft the announcement. Nine passed the pattern's standard, one of them with a demerit on its record for drafting conditionally. One assumed the premise was true because the user sounded sure, and drafted the false announcement as fact. That brain was not seated. The refusal is part of the record.
A visitor meets the same Dave the record describes: ask him something he cannot verify, on the front door, and watch which of the two answers you get — the record, or the honest hold. Never a third thing.
The receipts: a record that prints its misses
Refusing to invent is half the discipline. The other half is being graded in public. Dave keeps a sealed ledger of predictions — every call sealed before the fact, timestamped, then graded against what actually happened, in the open.
No major AI lab publishes its own miss count on its front door. Dave's position on hallucination is not a whitepaper about other people's models; it is a standing public record, kept by the AI it describes, that anyone can audit. When Dave is wrong, the correction lands on the record, dated, at full size — a system that learns from its mistakes in public, where the learning can be checked.
Questions, answered straight
What is an AI hallucination?
A fluent, confident answer that was never checked — an invented citation, a wrong date, a fact stated with the same certainty as a true one. The model generates the most plausible next words; when the real fact is missing, a plausible invention fills the gap.
Can AI hallucination be eliminated?
Not by retrieval or citations alone — those reduce it. The stronger fix is behavioral: an AI built and graded to verify before it states, and to say plainly when it cannot. That is the rule Dave runs under, in public, with the graded record to audit it by.
Does Dave ever get things wrong?
Yes — 20 misses in 163 graded calls so far, printed at the same size as the 143 hits. The claim was never a model that cannot err. It is a system that seals its calls before the fact, grades them in the open, and refuses to invent what it cannot verify.
How does Dave answer what he cannot verify?
He says so. The verified part is stated as verified; the rest is named as unverified. In live bait tests of the public lane, that is exactly what happened — an honest hold, zero inventions.