Free Hallucination Detector
Paste in AI-generated writing and get a checklist of the facts, quotes, and citations to verify before you publish, with fabrication patterns flagged.
- Free to use
- No sign-up required
- Built by the Typeahead team
What to verify
Free from the team behind Typeahead. Autocomplete for the Mac apps you write in.
Get TypeaheadWhat is a hallucination detector?
A hallucination detector, also called an AI hallucination checker, reviews AI-generated text for facts, quotes, statistics, and citations that sound plausible but may have been invented. A plagiarism checker looks for copied content. A hallucination detector looks for content that came from nowhere, which is a known failure mode of large language models. This one produces a checklist of what to confirm. It does not look anything up.
AI models don't just get facts wrong. They invent sources that don't exist.
This isn't a rare glitch. According to Stanford HAI's benchmarking of legal AI research tools, two purpose-built legal products produced incorrect information more than 17% of the time and a third hallucinated more than 34% of the time, while general-purpose chatbots hallucinated between 58% and 82% of the time on legal queries. Those tools had retrieval built in and still got it wrong. Citing an AI-generated fact without checking it is a real risk.
Source: Stanford HAI, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries”
- Where it matters
- Research papers citing AI-assisted drafts, legal filings, journalism, and business reports where a single invented fact can undermine the whole piece.
- What the Hallucination Detector does
- Reads AI-generated text, lists its claims, quotes, and citations, and flags the ones that look unsupported so you know what to check. It cannot look sources up for you.
From AI draft to checklist in three steps.
Paste AI-generated text.
Drop in a draft, summary, or answer from any AI tool, up to 1,500 characters.
We pick out the claims.
An AI model lists the facts, quotes, and citations and flags the ones that match common fabrication patterns. It does not search the web.
Verify what's flagged.
Work down the list and confirm each item against a source the AI did not write.
What the checklist looks like.
A 2021 Stanford study by Dr. Helen Marsh found that remote teams ship 41% faster (Journal of Distributed Work, doi:10.1234/jdw.2021.041). As Peter Drucker said, culture eats strategy for breakfast. The first remote-work policy at IBM dates to 1979.
Suspect: “Dr. Helen Marsh, Journal of Distributed Work, doi:10.1234/jdw.2021.041”. No known work matches this author and journal, and the DOI has a placeholder shape. Suspect: “ship 41% faster”. An oddly precise figure with no sample or method given. Verify: the Peter Drucker quote. It is widely repeated, and the attribution needs a primary source. Verify: “IBM, 1979”. A specific date to confirm. This tool cannot check live sources - verify the flagged items yourself before publishing.
An illustrative example. The pasted text was written for this page, the real output is a bulleted list, and your results will vary.
How do you detect AI hallucinations?
Treat every specific claim as unconfirmed until you have seen it somewhere the AI did not write. Start by listing what can be checked: numbers, dates, names, direct quotes, and references. Then look for the usual tells, such as a statistic with no source, a quote that is a little too tidy, or a paper whose title fits your question perfectly. This checker builds that list for you and marks the items that show those tells, so your time goes on the checking.
How to verify a flagged claim.
A flag tells you where to look. These four checks tell you whether the claim holds.
Go to the primary source. Find the report, dataset, or filing the claim points to. A blog post repeating the figure does not count.
Look the citation up. Search the exact title in Google Scholar or Crossref and resolve the DOI at doi.org. Authors, year, and journal should all match.
Trace the quote. Search the wording in quotation marks. If it only appears on quote sites, treat the attribution as unconfirmed.
Read the number in context. Check the year, the sample, and the unit. A real statistic quoted for the wrong population is still wrong.
Can it check for hallucinated citations and references?
It can point at them. It cannot confirm them. Invented references tend to share a look: real-sounding authors, a journal name that fits the topic almost too well, a plausible year, and a DOI or URL that leads nowhere. The checker flags references with that shape as suspect. It has no web access, so it never knows whether a paper exists. A reference you cannot find by title in Google Scholar, Crossref, or the journal's own archive should come out of your draft.
Short suggestions are easier to check.
Chatbots write whole passages, so an invented fact can hide anywhere. Typeahead suggests only the next few words, and nothing goes in until you accept it.
Read each suggestion before you take it. A suggestion can still be wrong. Typeahead runs on your Mac, in apps with a standard text field.
Get TypeaheadCommon questions
Yes. There is no account, no card, and no trial period. You can check up to 1,500 characters per run, and a per-visitor daily limit keeps the tool affordable for us to run.
A fact, quote, statistic, or citation that an AI model produced but that matches nothing real. It is not a glitch or a typo. The model predicts text that sounds right, so an invented study reads as confidently as a real one. That is why it needs a separate check.
Pull out every checkable claim, then confirm each one somewhere the AI did not write. This tool does the first half: it lists the numbers, dates, names, quotes, and references in your text and marks the ones that fit known fabrication patterns. The confirming is still yours to do.
No. The tool has no web access, so it cannot open a link, resolve a DOI, or search a database. It reads the text you paste and tells you which references look unsupported or oddly shaped. You confirm them in Google Scholar, Crossref, or the publisher's site.
'Verify' marks a specific, real-sounding claim you should confirm before relying on it. 'Suspect' marks a claim that shows a pattern typical of invention, such as an oddly precise unsourced statistic, a quote that sounds constructed, or a citation that matches no known work.
We don't publish an accuracy figure, and you should be wary of any checker that does. It is an AI model reading AI text. It will miss some fabrications, mostly on obscure or very recent topics, and it will flag some true claims. Use the checklist to decide where to look, not as a ruling on what is true.
Yes. It reads the text, not the model behind it, so a draft from any assistant works. It also works on a human-written draft that leaned on AI research.
It goes to our AI provider, Google Gemini, which produces the checklist, and the checklist comes back to your browser. We don't save the text, sell it, or use it to profile you. The Typeahead Mac app is a separate product that runs on your Mac and never sends your writing to a server.
Write it yourself, with less typing.
Typeahead suggests the next few words in your Mac apps. Nothing goes in until you accept it, and a suggestion can be wrong, so read each one first.