What the Hallucination Claim Extractor does
This tool pulls the checkable claims out of AI-generated text - sentences with numbers, dates, names, quotations, superlatives, attributions and cause-and-effect statements - ranks them by how much damage an error would do, and turns them into a verification checklist. It extracts; it never decides whether a claim is true.
Hallucinations hide in exactly these details: a date a year out, a statistic with no source, a quotation nobody said. Reading a long answer for them is slow; a list of the sentences that need checking, with the specific detail to look up, is fast. The text is analysed in your browser and not sent anywhere.
How to use it
- Paste the AI-generated text, or load an example.
- Press Extract. Each sentence is split out and tested for checkable detail; sentences with none (opinions, questions, transitions) are skipped and counted.
- Work through the claims table from high risk down. The Check column lists the exact numbers, dates, names or quotations to look up.
- Download the checklist as CSV to share the verification work, or copy it as a plain list.
Reading the results
Risk is an A2Z heuristic, not a probability of error. Points are added per feature: quotation 3, number 2, date 2, attribution 2, superlative or absolute 2, named entity 1, causal link 1. Four or more points is high, two or three medium, one low.
A high-risk claim is one where a mistake would be specific and consequential, and so is worth checking first. It may be perfectly true. A low-risk sentence may still be wrong.
Hedges such as about, may or approximately are shown but do not lower the risk: a hedged wrong number is still a wrong number.
Worked example: a short answer about the Eiffel Tower
The history example is seven sentences. Three - "It remains a beautiful sight at night", a question, and "Many would say yes" - have nothing checkable and are skipped. Four claims remain, three of them high risk.
The first sentence carries a date (31 March 1889), a height (300 metres), a second date (1930) and two names: high risk. (Its "tallest ... in the world" is not caught as a superlative because the words sit apart - a reminder that the rules are patterns, not understanding.) The visitor figure of 6.3 million in 2023, attributed to the tower's operating company and hedged with "about", is high risk. The quotation attributed to Eiffel is high risk because quotations are among the most often invented details. The sentence naming Koechlin and Nouguier as the originators of the concept is low risk: names only.
The checklist tells you to confirm the completion date, the original height, the year it was surpassed, the 2023 attendance figure and its source, and whether the quotation is genuine and correctly worded.
How the extraction works
The text is split into sentences, avoiding breaks after abbreviations and inside numbers. Each sentence is then matched against patterns: digits and number words with units or percentages; dates and years; text in straight or curly quotation marks; capitalised multi-word names; phrases such as "according to" and "said"; superlatives and absolutes such as "the first", "the largest", "always" and "never"; and causal links such as "because" and "led to".
These are deterministic rules, so the same text always gives the same list, and every flagged detail is shown so you can see why a sentence was picked. The price of determinism is that the rules miss claims phrased without any of these signals, and occasionally flag a harmless sentence.
Limitations: what the result does not prove
- It never checks a claim against any source and cannot tell you whether anything is true or false.
- Claims without numbers, dates, names, quotations or the listed signal words are not extracted, and some statements that are wrong in meaning rather than in detail will be missed.
- The risk score is an A2Z heuristic for ordering the work, not a measure of how likely a claim is to be wrong.
- Name detection relies on capital letters, so it works best on English and other Latin-script text.
Privacy: where your data goes
Everything you paste, type or drop is processed in this browser tab. It is not uploaded, logged, stored or sent to analytics. Session recording and tag-manager scripts are switched off on this page.
Standards and sources
- NIST AI 600-1 - Generative AI Profile (confabulation)
- Anthropic - Reduce hallucinations
- Min et al. - FActScore: fine-grained atomic evaluation of factual precision
Frequently asked questions
Can this tool tell me if an AI answer is hallucinated?
No, and no text-only tool honestly can. It finds the statements that could be wrong in a checkable way and lists what to look up. Whether each one is true needs a trusted source: the original document, an official statistic, a primary publication.
Which kinds of AI hallucination are most common?
Invented or slightly wrong specifics are the most damaging: statistics, dates, citations, quotations and attributions. They read as authoritative and are easy to repeat. That is why the extractor weights quotations, numbers, dates and attributions most heavily.
Why is a hedged claim still high risk?
Words like about or approximately change how precise a claim is, not whether it is right. If the true figure is 2 million, "about 6 million" is still wrong. The hedge is shown so you know how exact the check needs to be.
What should I check first?
Start with high-risk claims that the reader will act on or repeat: figures in a report, quotations, and anything attributed to a named organisation or study. Low-risk sentences with only names can often be skimmed.
Does the text I paste get stored or sent to an AI?
No. Sentence splitting and pattern matching run in your browser. Nothing is uploaded to a2z.tools or sent to any model, so you can use it on drafts and internal documents.
How is this different from the citation coverage checker?
This tool finds checkable claims in any text. The citation coverage checker starts from text that already has citations and maps which claims carry a reference, which references are missing and which numbers are broken. They work well one after the other.
Last reviewed by the A2Z.Tools team against the sources listed above.
