An AI hallucination detector reads text an AI produced and flags the parts that look invented. A hallucination is a statement made with full confidence that is false or unverifiable: a made-up statistic, a citation to a paper that does not exist, a date that is wrong, an event that never happened. Because the prose is fluent, these are easy to miss by reading.
What separates a useful one from a confident one is whether anything is checked outside the model. Some findings come from querying a real database. Others are a model judging that a sentence looks invented. Both are useful; only one is evidence, and knowing which you are looking at is the whole skill of using these tools.
More specific than "is this true", which is why the categories are the useful part of the output.
Fake citation. A reference identifier that does not resolve: a DOI with no CrossRef record, an arXiv ID that is not in arXiv. The most checkable category and the most common academic failure.
Non-existent paper. A plausible title, journal and author that no database carries.
Made-up source. A publication or organisation that does not appear to exist.
Invented statistic. A figure presented as established fact with nothing behind it.
Suspicious specificity. The tell that catches the most careful readers out: false precision. A model reaching for authority produces things like "87.3% of workers improved by exactly 41 minutes", and the implausible exactness is itself the signal.
Dead link. A URL that is fetched and does not resolve. Hard evidence, because it is a live check.
Plausible but uncited. A reasonable-sounding assertion with no support offered. Often not wrong, and worth reading first anyway, because it is where an unsupported claim hides inside otherwise sound prose.
They do not carry equal weight, and the output does not pretend they do.
Verifiable. A live source confirmed the record exists. The only verdict that rests on a lookup, and the response names the database that answered.
Fabricated. Arrives from two different routes under one word. Either a lookup returned nothing, which is strong, or a model judged the claim invented, which is weaker. So read the source field before acting on it.
Suspicious. Always a model judgement. Good for ordering your work, not for concluding anything.
Likely true. Also a judgement, and the one to treat most carefully, because it reads as reassurance and is not a check.
A result that reports every flag with the same confidence is concealing which of its findings are evidence. In a real test of our own tool, five flags came back on a fabricated passage and only one of them had been verified against anything; the other four were model assessment, and the output said so. That test is published in full here.
When you evaluate any tool in this category, the single most useful probe is to feed it a citation you invented and read how it justifies the flag. "This appears fabricated" is a guess. "CrossRef has no record of this DOI" is a check.
Three tools get reached for interchangeably and answer different questions.
An AI detector estimates whether text was machine-written. That is independent of truth: a person can write something false and a model can state something correct. Using one to find fabrications gives you no information about fabrications.
A plagiarism checker finds text matching an existing source. A fabricated citation matches nothing anywhere, which is exactly why it passes a similarity check untouched. This is the most expensive of the three confusions, because the check runs clean and feels like a verification.
A fact-checker is the closest neighbour and overlaps substantially. The practical difference is emphasis: hallucination checking is built around the failure modes of generated text, particularly citation fabrication, while general fact-checking addresses claims regardless of origin.
A citation that exists in CrossRef is a citation that exists. Whether the paper says what your sentence claims is a separate question, and the gap between the two is where the most damaging errors live: a real source attached to a claim it does not support. No automated checker closes that. Reading the source does.
So the honest framing of the whole category: treat the output as a prioritised list of what to check first, never as a clearance. A clean result means nothing was caught, not that nothing is wrong.
The daily count gates scans and rewrites; hallucination checking is counted separately.
| Free | Starter | Pro | |
|---|---|---|---|
| Price | $0 | $9.99/mo · $7.49 yearly | $19.99/mo · $14.99 yearly |
| Actions a day (shared) | 3 | 20 | Unlimited |
| Words per scan | 1,500 | 5,000 | 7,500 |
| Words per rewrite | 260 | 5,000 | 10,000 |
| Detection a month | No cap | No cap | No cap |
| Humanizer a month | No cap | 40,000 words | 80,000 words |
One action is one scan, one file upload or one rewrite, from a single shared daily count. That daily number is the gate that matters: there is no monthly ceiling on any plan, so what separates the tiers is how often you can work and how long a single piece can be, not how much you get through in a month.
Read the free row together rather than line by line. No monthly cap sounds unlimited and is not: three actions a day at up to 260 words a rewrite is roughly 780 rewritten words in a day. It is a real allowance for checking and fixing specific passages, and it is not a plan for processing a backlog.
Current prices, the Business and Enterprise tiers and regional pricing are on the plans page, which is the page we keep up to date.
The hallucination checker is not one of those shared actions. It runs on its own allowance: 3 checks a day with no account, 5 on a free account, 25 on Starter and no daily limit on Pro. So checking a passage here does not spend a scan or a rewrite, and the row above does not cap it.
A tool that reads AI-generated text and flags the parts that look invented. A hallucination is a confident statement that is false or unverifiable: a made-up statistic, a citation to a paper that does not exist, a wrong date, an event that never happened. Because the prose reads fluently these are hard to spot by reading, so the tool breaks the text into claims and reports which ones look fabricated, with categories like fake citation, invented statistic and suspicious specificity.
Most often a citation to something that does not exist, which is the single most common academic failure. The other signature is false precision: a model reaching for authority produces figures like "87.3% improved by exactly 41 minutes", and the implausible exactness is itself the signal. Also common are real-sounding journals and organisations that no database carries, and URLs that look right and do not resolve.
Two different ways, and the distinction is the most important thing to understand. Reference identifiers are checked against live sources: a DOI against CrossRef, an arXiv ID against arXiv, a PubMed ID against PubMed, an ISBN against Open Library, and a URL by fetching it. Everything else, including statistics, quotes and events, is assessed by a model. Only the first kind is a measurement, and a result that reports both with equal confidence is hiding which findings are evidence.
No, and mixing them up is common and costly. An AI detector estimates whether text was machine-written, which is independent of whether it is true. A hallucination detector estimates whether the claims are fabricated, regardless of who wrote them. A plagiarism checker is equally wrong for the job, because invented citations match no existing source and therefore sail through a similarity check.
Treat it as "nothing was caught", not "nothing is wrong". Model judgements carry both misses and false positives, the per-check size is limited, and the flags are weighted toward fabrications with detectable signatures. The error that survives most reliably is a genuine source attached to a claim it does not support, because every signal the tool has says that reference is sound. Only reading the source closes that gap.
Yes, with no account: three checks a day at up to 200 words each, minimum thirty words, resetting daily. The per-claim breakdown and the reliability score are included on the free tier; clickable source links and the downloadable report are paid. These checks are counted separately from the shared daily actions that gate scans and rewrites, so running one does not spend a scan.
Six ways to catch a fabrication, with a real test published.
See the comparison →Where the verification is real, and the gap it cannot close.
Read the guide →Why a fake case is judged rather than verified.
Read the guide →Three checks a day, no account.
Check a passage →Three checks a day, no account. Read how it justifies the flag: a guess and a database lookup look very different.