OpenAI traced confident wrong answers to how tests are graded

Start here

Ask an assistant for a person's date of birth or the title of an obscure paper, and it can answer with total confidence, and be wrong, sometimes three different wrong answers in three tries. Vendors call this a hallucination: a fluent, plausible statement that is not true. OpenAI's help center defines the pattern in a single line: ChatGPT will occasionally make up facts or hallucinate outputs, in its frequently asked questions page. A research paper OpenAI published on 5 September 2025 goes further than naming the behavior; it argues for a specific, checkable cause.

What the documents say

The research post states models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty, comparing this to a multiple-choice test where a wrong guess scores the same as an unanswered one, zero, while a lucky guess can score full marks. The paper cites a benchmark called SimpleQA, comparing two OpenAI models: one abstained 52 percent of the time with a 26 percent error rate, while an older model abstained just 1 percent of the time with a 75 percent error rate, even though its raw accuracy looked slightly better. Its conclusion states plainly that "Accuracy will never reach 100%" on open-ended questions, since some have no determinable answer regardless of model size. Anthropic's separate guidance on reducing hallucinations treats hallucination as ordinary even in advanced models and offers techniques such as permitting the model to say it does not know, while cautioning these reduce, not eliminate, the problem.

Check this

A reader can reproduce the incentive described in the paper on a small scale: ask a narrow factual question the assistant is unlikely to know precisely, such as an exact obscure date, and see whether it guesses or states its uncertainty. The post's own recommendation, that scoring should reward an honest "I don't know" over a wrong guess, is a check a reader can apply to their own use too: treat a hedge as more trustworthy than false confidence, not less.

What holds and what fails

The claim that hallucination has an identifiable statistical cause, not an unexplained glitch, holds according to the cited paper's own framing, and it is worth naming this as the vendor's account of its own system, not an independent audit. It is an editorial extension, not something the paper claims, to say a reader should distrust any unsourced factual claim regardless of how confidently it is phrased. Where a question has a genuine, verifiable answer, checking a primary source remains the only fix the documentation endorses; no amount of better grading removes that need.

• Ask an assistant to state its confidence, or say it does not know, rather than pressing for a guess.
• Treat any unsourced date, quote, name or statistic as a claim to verify, not a fact.
• Read the vendor's own hallucination guidance before assuming a helpful-sounding answer is accurate.

Hallucination is not a mysterious failure mode; the vendors' own documents describe a specific incentive that produces it and a specific caution that reduces it. Neither promises the problem is solved.