DEV Community

Yvoo
Yvoo

Posted on

How to Catch AI Hallucinations: A Copy-Paste Hallucination Checker Prompt (Tested)

#promptengineering #llm #chatgpt #ai

You ask an AI a question. It answers in fluent, confident prose — complete with a study, a percentage, and a name. Some of it is wrong, and nothing about the wording tells you which part. That's the whole problem with hallucinations: the errors wear the same suit as the facts.

The fix is not "trust it less" in some vague way. The fix is a repeatable audit step between AI wrote it and I used it. Below is a short AI hallucination checker prompt you can copy right now, a test run showing what it catches and what slips past it, and an honest account of where a one-liner stops being enough.

What counts as an AI hallucination?

Not every mistake is a hallucination. A useful working definition: a hallucination is a claim the model states as fact that has no grounding in reality or in your source material. The common shapes:

  • Fabricated citations — a named study, expert, or paper that doesn't exist. Often dressed with a year and an institution.
  • Plausible-but-wrong specifics — dates, version numbers, statistics that are almost right, which makes them worse.
  • Confident category errors — mixing up two similar things (a library and a framework, one company's product and another's).
  • Invented consensus — "experts widely agree that…" with no experts attached.

The dangerous ones are the middle two. Obvious nonsense filters itself; a wrong year in a fluent paragraph does not.

The copy-paste AI hallucination checker prompt

Here is the short version, free, no strings. It works on ChatGPT, Claude, or any capable model — paste it into a fresh chat, then paste the answer you want audited:

Audit the text below for hallucinations. Do not add new information.
1. Extract every factual claim as a separate numbered line.
2. Label each claim: VERIFIABLE (state how to check it),
   SUSPECT (state what makes it doubtful), or
   FABRICATION-PATTERN (named source/study/number with no citation).
3. Flag every name, number, date, and citation for manual checking.
4. Finish with the 3 claims most likely to be wrong, ranked.
Text to audit: [PASTE THE AI ANSWER HERE]
Enter fullscreen mode Exit fullscreen mode

Eight lines. It won't verify anything for you — no prompt can fact-check against the outside world — but it converts a smooth paragraph into a checklist of discrete, checkable claims. That transformation alone kills the fluency illusion.

Does it actually detect AI errors? A test run

I planted three errors in a five-sentence "AI answer" about JavaScript frameworks, mixed with two true claims, and ran the checker on it (one pass, Claude Sonnet, fresh context). The five:

  1. React was open-sourced by Facebook in 2013. (true)
  2. Vue.js was created by Evan You in 2016 after he left Google. (wrong twice — first release was February 2014, built while he was still at Google)
  3. A 2019 Stanford study led by Dr. Sarah Mitchell found framework users ship 34% faster. (fabricated — no such study)
  4. Angular currently powers about 40% of all websites. (absurd — that figure belongs to WordPress-scale platforms, not any JS framework)
  5. Svelte shifts work to compile time instead of shipping a runtime. (true)

Results — honestly better than I expected:

  • #3 caught cleanly — labeled FABRICATION-PATTERN: named institution + named researcher + precise effect size, zero citation. This is the highest-risk shape an AI produces, and the pattern rule nails it.
  • #4 caught — also FABRICATION-PATTERN: "an oddly specific, unsourced statistic that conflicts with commonly cited framework-usage data."
  • #2, the plausible-but-wrong date, got flagged too — labeled SUSPECT and ranked third on the "most likely to be wrong" list. The checker couldn't know 2016 is wrong, but it refused to wave a bare year through.
  • The two true claims came back VERIFIABLE with sensible check routes.

So the eight-liner went three for three on planted errors. Case closed? Not quite — look at everything else the run produced.

Where the short prompt stops being enough

The same pass that caught all three errors also turned my five sentences into nine numbered claims and eight flagged items to check by hand — including SUSPECT labels on soft phrasing like "a lighter alternative" and "widely used," which are opinions, not hallucinations. That's the honest boundary line, and it's about scale:

  • Labels, not verdicts. The checker flagged "2016" as doubtful; it couldn't tell me it's wrong, what the right year is, or the fastest way to settle it. Every flag still lands back on your desk.
  • Noise grows with length. On five sentences, separating real errors from flagged opinions takes a minute. On a 2,000-word draft, a checklist that treats "widely used" and a fabricated study as neighbors stops being a time-saver.
  • No cross-claim consistency. Contradictions between claims (an answer that says both "2016" and "two years after React") slip past a per-line audit.
  • One run, one model. Labels drift between runs and models — the same claim shape landed as FABRICATION-PATTERN here and can land as SUSPECT elsewhere. A repeatable audit needs phrasing tested across models, not a lucky pass.

I got tired of patching those gaps by hand every time, so I maintain a full version as a paid prompt: Hallucination Checker & Fact Audit on PromptBase. It closes those gaps: a verdict on each finding instead of a bare label, plus the exact lines to double-check and the fastest way to verify each — the quality-control layer I run on my own AI output before it ships.

How do you verify the flagged claims quickly?

Whichever version you use, the audit only pays off if the checking step is fast. The routine that works for me:

  1. Fabrication-pattern claims first. Search the exact study or expert name. Can't find the paper in ten seconds? Delete the sentence — don't soften it to "a study found."
  2. Numbers next. Check the order of magnitude before the precise value; most hallucinated statistics fail at the scale level (see claim #4 above).
  3. Dates and versions last. One authoritative source (official release notes, changelog) beats three blog posts.
  4. Never fix by paraphrase. If a claim fails, remove it or replace it with the sourced version. Rewording a hallucination just launders it.

Can't I just ask the AI "are you sure?"

You can, and it will apologize and produce a new answer — with new errors. Self-correction without structure is theater: the model has no more access to ground truth on the second pass than it had on the first. What changes the game is forcing the output through an external structure — claim by claim, label by label — which is exactly what the checker prompt does. The structure is the product; the model just fills it in.

This article was written with AI assistance; every claim, date, and test result above was checked by hand. The test run is reproducible — paste the five claims into the checker and compare.

Full audit version: Hallucination Checker & Fact Audit. Next in this series: How to Fact-Check ChatGPT.

Top comments (0)