Can You Tell an AI Chatbot to Say "I Don't Know"? What Anthropic and OpenAI Say

Yes. Anthropic's documentation lists "allow Claude to say 'I don't know'" as its first basic technique and says giving a model explicit permission to admit uncertainty "can drastically reduce false information." It reduces the problem rather than ending it: OpenAI says accuracy will never reach 100%, because some real-world questions are unanswerable.

The permission matters because of how models are trained and scored, which OpenAI explains in its own research post. This article covers the prompt-side fixes both companies describe. For why chatbots invent things in the first place, see the guide to why AI chatbots make things up.

1. Why a model guesses unless you say otherwise

OpenAI defines a hallucination as a plausible but false statement generated by a language model, and says they can show up even on simple questions. Its argument is that standard training and evaluation procedures reward guessing over acknowledging uncertainty.

OpenAI's illustration is a multiple-choice test: a blank answer scores zero, while a lucky guess can score. For someone's birthday, a model that guesses "September 10" has a 1 in 365 chance of being right, which is about 0.27%, while saying "I don't know" scores zero. Over thousands of questions, OpenAI says, the guessing model looks better on accuracy-only scoreboards than a careful model that abstains.

Its own numbers show the trade-off. On the SimpleQA evaluation in the GPT-5 system card, gpt-5-thinking-mini abstained on 52% of questions, was right on 22% and wrong on 26%. The older o4-mini abstained on 1%, was right on 24% and wrong on 75%. Dividing errors by the questions each model actually answered, the first was wrong on 26 / (22 + 26) = about 54% of its answers and the second on 75 / (24 + 75) = about 76%. Accuracy differed by only two points, but the habit of guessing produced far more wrong answers.

OpenAI also notes that its Model Spec says expressing uncertainty or asking for clarification is better than confidently giving information that might be wrong. Whether a given chatbot follows that is up to the product, which is why your own instruction helps.

2. The three basic techniques in Anthropic's guide

Anthropic's page lists three basic strategies.

1. Allow "I don't know." Explicitly give the model permission to admit uncertainty.

2. Ground long-document answers in direct quotes. For long documents (over 20k tokens), Anthropic suggests asking the model to extract word-for-word quotes first, before doing the task. This grounds the answer in the actual text.

3. Make it cite and verify its claims. Anthropic says to have the model cite quotes and sources for each claim, which makes the response auditable. You can also have it look for a supporting quote after it answers; if it cannot find one, it must retract the claim.

3. Four more techniques Anthropic calls advanced

• Chain-of-thought verification: when an answer looks wrong, review the model's summarized reasoning (in Anthropic's API, thinking blocks) to find faulty logic or assumptions.
• Best-of-N verification: run the same prompt several times and compare. Inconsistencies may signal hallucinations.
• Iterative refinement: feed the output into a follow-up prompt asking the model to verify or expand its earlier statements.
• External knowledge restriction: tell the model to use only the documents you provide, not its general knowledge.

4. A prompt you can copy

Combining the basic techniques, a work prompt might look like this (wording is an illustration, not an official template):

Use only the text below. First, quote the sentences that are relevant. Then answer using only those quotes. If the text does not contain the answer, reply "I don't have enough information." After answering, find a quote supporting each claim and delete any claim you cannot support.

It works best when you paste the source text in, because the quote and restriction steps need something to check against.

5. Limits you should plan for

OpenAI's post says accuracy will never reach 100% because some real-world questions are unanswerable, though hallucinations are not inevitable, since a model can abstain when uncertain.

A model told to quote can still misquote, so spot-checking the quotes behind any figure, name or date that will be passed along is a reasonable habit.

6. Key takeaways

• Telling a model it may say "I don't know" is the first technique in Anthropic's hallucination guide, and Anthropic says it can drastically reduce false information.
• Models guess because many evaluations reward guesses: a birthday guess is right 1 time in 365 (about 0.27%), but still beats a blank on an accuracy-only score.
• For long documents (over 20,000 tokens), ask for word-for-word quotes first, then answer from them.
• Ask the model to find support for each claim and drop what it cannot support.
• OpenAI says some questions are unanswerable, so checking what matters is still necessary.