Catch hallucinations in generated summaries
Summarization is one of the most common LLM tasks — and one of the easiest to get subtly wrong. Models invent figures, swap names, or assert things the source never said. This tool sends your source document and the generated summary to your own LLM with a strict grounding rubric and returns a per-claim verdict: supported, unsupported, or contradicted.
How consistency checking works
The model is instructed to decompose the summary into atomic claims — single, verifiable statements — and to judge each one only against the supplied source. A claim is supported when the source directly backs it, unsupported when the source is silent on it (a hallmark of fabrication), and contradicted when the source asserts the opposite. The tool parses these verdicts into a clear list and highlights the contradicted claims, which are the most dangerous. Everything runs through a single direct request to your provider using your own key.
The three verdict types explained
Understanding what separates the three verdicts matters for knowing what action to take:
- Supported — the source contains a sentence or passage that directly backs the claim. The model found the grounding. Safe to keep.
- Unsupported — the source is silent on this point. The model invented or extrapolated it. Common when LLMs confidently add context that was not in the original. You need to either find a source or remove the claim.
- Contradicted — the source directly says the opposite. This is the most dangerous category: a confident falsehood. Fix or remove immediately.
A typical hallucination pattern looks like this. A source says “the project ran over budget by 12%.” The summary says “the project ran over budget by 20%.” That claim is contradicted, not unsupported, because the source states a different figure — it is an outright factual error, not just an addition.
Worked example of a consistency check
Source excerpt: “The meeting took place on 14 March. Seven members attended. The vote passed 5–2.”
A model-generated summary might produce: “The committee met in mid-March. All members were present and the vote was unanimous.”
Running consistency checking over that summary would flag:
- “met in mid-March” — supported (14 March is mid-March)
- “All members were present” — contradicted (source says seven attended, not all)
- “vote was unanimous” — contradicted (source says 5–2, not unanimous)
Two of three claims are outright errors, despite the fluent tone. This is exactly the kind of subtle falsification that reads fine on a quick skim.
When to use this tool
- Checking AI-generated executive summaries against long internal reports before distribution
- Validating research abstracts against their full-text sources
- Auditing customer-support AI answers against your knowledge base
- Confirming legal or compliance summaries before sending to clients
- Reviewing any RAG-generated answer that needs to stay grounded in retrieved documents
Tips for trustworthy results
- Use a capable model for the judging step (gpt-4o or claude-3-5-sonnet). Claim verification is harder than generation and benefits from a stronger model.
- Keep the source self-contained. If the summary relies on outside knowledge the source doesn’t contain, those claims will correctly read as unsupported.
- Pay closest attention to contradicted claims — those are outright errors, not just omissions.
- Re-run after editing the prompt or summary to confirm your fix actually removed the unsupported or contradicted claims.
- For very long sources, break them into sections and run separate checks per section — shorter source windows improve the judge model’s accuracy.
- A high ratio of unsupported claims often means the prompt was too vague rather than that the model hallucinated; tightening the system prompt can shift those claims to supported or correctly absent.