ClaudeVerification mistakesLevel 2: Context

Why Does Claude Sound Sure When It's Wrong?

Claude writes in the same fluent tone whether a detail is right or wrong, so tone tells you nothing. Anthropic says that even the most advanced models can sometimes generate incorrect text. Give Claude permission to say it doesn't know, ask for sources, and check any specifics you'll rely on.

Prompt to copy
[Your question.]

If you're not sure of something, say "I don't know" rather than giving your best guess. That's a useful answer here.

After your answer, list every specific claim: numbers, names, dates, citations, quotes and method names. For each one, give a source I can open, or write "no source." Put the claims I should check first at the top, starting with the ones that would cost the most if wrong.
↑ cost if wrong
Find a second sourcecostly, hard to check
Always checkcostly, easy to check
Treat as a leadcheap, hard to check
Quick lookcheap, easy to check
how easy to check →
A rule of thumb, not an Anthropic chart: put your checking effort where a wrong fact costs the most.

Try it now

  1. 1Turn on search: click + in the chat window and select Web search, if your version shows that toggle.
  2. 2Paste the prompt above with a question that needs specifics.
  3. 3Open the sources for the top three claims, and look for the exact number or quote on each page.
  4. 4You'll know it worked when each claim you plan to use has a page that says the same thing. Treat every "no source" claim as unconfirmed.

Use it when

When an answer includes checkable specifics that you'll publish or act on: statistics, citations, dates, API method names, legal claims, doses, or quotes attributed to a person.

Skip it when

Brainstorming and rough drafts, where a wrong detail costs nothing because you'll replace it anyway. Don't check well-known basics either. Spend your time on the specifics a decision rests on.

Why it works

Anthropic's hallucination guide says that explicitly allowing Claude to admit uncertainty can drastically reduce false information. So a prompt that demands an answer works against you, and one that accepts "I don't know" helps. The guide also suggests having Claude cite a source for each claim and retract any it can't support, or limiting it to documents you provide. Search helps with facts that change, because Anthropic's Sonnet 5.5 guide says the model sometimes answers from training when a search would catch a change. Anthropic is clear that these techniques don't eliminate errors. Its medical guidance shows how it treats the highest-stakes specifics: for a child's medicine dose, its suggested prompt has Claude take the number from the product label, never from memory.

Example

Say a draft cites a study by name, with a 42% figure. Search for the study's title. If you can't find it, remove the claim. If you do find it, look for the 42% in the original, since a number can be attached to the right study and still be wrong. A method name in code works the same way: look it up in the library's docs, or run it.

Going further

A confidence label from Claude isn't a check. Rule of thumb, not a vendor figure: "high confidence" is more generated text, so use it to decide what to check first. Anthropic's guide suggests running the same prompt several times, since differences between answers can flag weak claims. In long agent runs, its Fable 5 guide suggests telling the model to tie each progress claim to a tool result, which it says nearly eliminated fabricated status reports in Anthropic's testing. One trap: asking the model to reproduce its internal reasoning can trigger a reasoning_extraction refusal on Fable 5 and Sonnet 5.5, so ask for a short explanation of the answer.

Common mistakes

  • Taking a detailed answer as a researched one. Precise numbers can be made up just as easily as round ones.
  • Writing prompts that demand an answer no matter what, like "you must give a figure," which leaves no room for "I don't know."
  • Asking Claude "are you sure?" and accepting a yes. Check an outside source instead.
  • Over-checking everything, then giving up on checking at all. Focus on what a decision depends on.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course