Hallucination review checklist
Large language models produce fluent, confident text — including confident fabrications. The riskiest hallucinations hide inside specific-sounding facts: a precise statistic, an exact date, a named person or company, or a URL that does not exist. This tool scans any LLM output and extracts those claims into a checklist so you can verify each one against a real source before you publish, send, or act on the text.
How it works
The tool splits your text into sentences and runs local pattern matching for the claim types that most commonly turn out to be hallucinated: numbers and percentages, four-digit years and date phrases, capitalised named entities, URLs, and citation-style references. Each matching sentence becomes a checklist item tagged with what triggered it, so you know whether to verify a figure, a name, or a link. You then tick items off as you confirm them, and the running counter shows how much of the output still needs checking.
Why these specific claim types?
Not all text is equally likely to contain hallucinations. Language models are trained to be helpful and fluent, which means they are most likely to invent facts when asked to be specific — to name a person, cite a figure, or provide a date. The tool focuses on the pattern types that have the highest hallucination rate in practice:
Precise numbers and percentages — Models invent statistics with false authority. A claim like “productivity improved by 23.7%” sounds authoritative but may have no basis in any study. Even when a real study exists, the number may be subtly wrong.
Years and specific dates — Models frequently misattribute when something happened. The launch date of a product, the year a law passed, or the publication year of a paper are all frequently wrong in model output.
Named entities (people, organisations, products) — Models confuse names, invent people who do not exist, attribute quotes to wrong speakers, and combine details from different individuals into a single plausible-sounding biography.
URLs and citations — Dead links and invented citations are among the most damaging hallucinations because they look verifiable but waste a reader’s time or lead them to unrelated content. Always click before trusting.
Citation-style references — “According to a 2019 study by Smith et al.” — when the model cannot remember a real citation, it often constructs a plausible-sounding one with a real author name and a real journal, but a wrong title and year.
Practical verification workflow
- Paste the model output and run the scan.
- Start with the URL items — click each one and confirm it exists and says what the model claims.
- Move to statistics — search for the exact figure or the study it supposedly came from.
- Check named entities — confirm each person exists and holds the role attributed to them.
- Verify dates — these are fast to check and frequently wrong.
- Tick each item as confirmed. Anything you cannot verify should either be cut or rewritten with appropriate hedging.
Tips and notes
- Start with numbers and links. Fabricated statistics and dead URLs are the fastest hallucinations to catch and the most damaging to miss.
- Trace to a primary source. Verifying one model’s claim with another model is not verification — confirm against documentation, an official site, or a peer-reviewed source.
- Re-run after edits. If you ask the model to revise, paste the new version and rescan; revisions frequently introduce fresh fabricated details.
- Nothing leaves your browser. The scan is entirely local, so it is safe to paste sensitive drafts.