Hallucination Check for Summaries

Compare a model summary with its source and flag unsupported statements.

Added October 15, 2026. Written with AI assistance and checked for clarity and placeholders; not reviewed by a lawyer, accountant, clinician or other professional. How these prompts are made

The Open buttons send the prompt text as currently shown below to that site only when you click. Fill in the placeholders first.

Check a summary against its source. Source text: [SOURCE_TEXT]. Summary: [SUMMARY_TEXT].

Split the summary into individual claims. For each claim, quote the source sentence that supports it, or write unsupported. Mark claims as supported, partly supported (the source says something weaker or different), or unsupported. Pay close attention to numbers, names, dates, causes and words like always, never and all.

Return a table with Claim, Source quote, Verdict and Fix, a count of each verdict, and a corrected summary that keeps only supported claims. If the source is too short to judge a claim, say so instead of guessing.

Also add a second pass that looks for omissions, listing important points in the source that the summary leaves out, and flag any place where the summary changes the emphasis, such as turning a possibility into a certainty.

Finish with a short list of habits that reduce unsupported claims in future summaries, such as asking for quotes with every number and limiting the summary to what the source actually states.

Use this layout for the claims table so it is easy to audit. Claim one: the quoted sentence from the summary. Source quote: the exact words from the source, or the word none. Verdict: supported, partly supported or unsupported. Fix: the corrected wording, or delete. As an example, if a summary says sales doubled last year and the source says sales rose by 60 percent, the verdict is partly supported, because the direction is right and the size is wrong, and the fix is to state the 60 percent figure. Count the verdicts at the end, for example three supported, one partly supported and one unsupported out of five claims, and report the share of fully supported claims as a percentage with the arithmetic.

Close by listing the three summary habits that most reduced unsupported claims in the test, so they can be written into the prompt for next time.

Fill in before use

Optional: type your details and the prompt above updates. Nothing is saved or sent.

Replace each bracketed placeholder with your own details, then paste into Any AI or another assistant.

Invented source: a quarterly note saying sales rose 60 percent, the Leeds shop closed in May, and hiring is paused.
Summary claims and verdicts:
1. Sales doubled. Source: "rose 60 percent". Partly supported; fix: "rose 60 percent".
2. The Leeds shop closed in May. Supported.
3. Hiring is paused. Supported.
4. The closure caused the sales rise. Source: none. Unsupported; delete.
Count: 2 supported, 1 partly, 1 unsupported. Fully supported share = 2 / 4 = 50 percent.
Omission: the source's note about a delayed audit is missing from the summary.

• quality