When a result sounds too good to be true, the useful question is not “is it right?” but “what would have to be true for it to be right, and how could I tell?” Today’s prompt makes a model break a claim into premises, tests and ways it could fail. It will not decide whether the claim is correct, but it gives you a checklist for asking sharper questions of the people who can.
Why this prompt, today
On October 26, 2026, OpenAI released model-generated results in mathematics, computer science and physics and named outside experts to check them. According to The Verge, it also published an open letter asking how the community should handle early access to AI-generated results, credit for people who check and explain them, and support for students. Whatever the outcome of those proofs, the situation shows a general skill: when a claim is announced before it is verified, you need a way to organize the checking. Our report on the three claimed discoveries has the details; this prompt is for any claim, from a product benchmark to a news headline or a colleague’s proposal.
The prompt
Paste this, then add the claim and any source text you have.
You are a careful, skeptical reviewer. I will give you a claim. Do not decide whether it is true. Help me check it.
CLAIM: {{paste the claim, with its source}}
MY BACKGROUND: {{e.g. non-specialist, manager, student}}
WHAT I HAVE: {{links, text, data, or "nothing besides the claim"}}
Produce:
1. RESTATEMENT: the claim in one plain sentence, without adjectives.
2. WHAT MUST BE TRUE: 3 to 7 premises that would all have to hold. Mark each as [STATED] if it appears in my material or [ASSUMED] if you are filling it in.
3. HOW TO CHECK: for each premise, the cheapest test I could run or question I could ask, and who is best placed to answer.
4. WHAT WOULD FALSIFY IT: two or three specific findings that would show the claim is wrong or weaker than stated.
5. WHO HAS A STAKE: the parties who benefit if the claim is accepted, with no accusations, just a list.
6. CONFIDENCE NOTE: say plainly which parts you cannot judge without outside expertise. Do not invent sources, numbers or quotations.
Keep each item short. If my material is too thin to do step 2 properly, say so and ask up to three questions instead.Why it works
The prompt separates three jobs that people often blend: understanding the claim, listing what supports it, and deciding whether to believe it. By forbidding a verdict, it avoids the model’s habit of sounding confident. The [STATED] versus [ASSUMED] labels make it visible where the model is filling gaps, which is where invented details usually creep in. Asking for falsifiers is a form of checking that mirrors how scientists test ideas: a claim worth taking seriously is one that could be shown wrong.
Prompt guides from model makers commonly advise being explicit about the role, the task and the output format. OpenAI’s prompt engineering documentation, for example, discusses structuring instructions clearly and giving models examples of the output you want. This prompt applies those basics to a task where overconfidence is the main risk.
A short illustrative example
This example is invented to show the shape of an answer, not a real transcript.
Claim: “Our new support bot resolves 90% of tickets without a human.”
A good response would restate it as “The bot closes 90% of tickets with no human involvement.” Premises might include: [ASSUMED] “resolved” means the customer did not reopen the ticket; [ASSUMED] the 90% covers all ticket types, not only easy ones; [STATED] it was measured over a specific period. Checks: compare reopen rates within 14 days, and ask for the split by ticket type. Falsifiers: if reopen rates are high, or the figure is only for password resets, the claim is weaker than it sounds. Stakeholders: the vendor selling the bot, the team that adopted it. Confidence note: the model cannot judge your data quality.
Limits and failure modes
- It is a checklist, not evidence. A model can list good questions and still be wrong about facts. Treat step 2 as a draft.
- Specialist claims need specialists. For advanced mathematics, medicine or law, the real check is a qualified person or a formal tool, not a chatbot.
- Anchoring. If you paste a long, persuasive source, the model may echo it. Give it the bare claim first and the source second.
- False balance. The prompt lists ways a claim could fail, but not all are equally likely. Weigh them yourself.
- Privacy. Do not paste confidential documents into a tool your organization has not approved.
Variations
- For your own work: swap “claim” for “plan” and add “assume it failed; why?” to link it with our pre-mortem prompt.
- For research: ask for the exact figure, table or section of the source where each premise should be verifiable.
- For teams: have two people run it independently and compare the lists.
- For agents: pair it with an explain-it-back check so the model states its understanding before acting.
What to watch next
- Community norms for releasing and crediting AI-generated research results.
- Tools that formalize claims, such as proof assistants, so checking is less manual.
- Whether labs adopt shared disclosure practices for results they cannot yet verify.
Key takeaways
- Ask what must be true for a claim to hold, not just whether it is true.
- Label premises as stated or assumed to expose where a model is guessing.
- Always ask for falsifiers: findings that would show the claim is wrong.
- The prompt organizes checking; it does not replace experts or evidence.
- Keep confidential material out of unapproved tools.
Frequently asked questions
Can this prompt tell me whether a claim is true?
No. It is designed to produce questions and checks, not a verdict. The final judgment needs evidence and, for technical claims, qualified reviewers.
Which models does it work with?
It is written for any modern chat model. Results vary, so review the list rather than trusting it blindly.
Why does the prompt forbid inventing sources?
Language models can produce plausible but false references. Telling the model to say what it cannot judge reduces, but does not remove, that risk.
