For researchers this is the one audience where the verification is real rather than advisory. The hallucination checker queries CrossRef for DOIs, arXiv for preprint IDs, PubMed for PMIDs and Open Library for ISBNs, and fetches any URL. That set is close to an exact overlap with an academic reference list, so a fabricated reference is checked against the actual record, not guessed at.
Which makes the honest part of this page narrow and important: confirming a reference exists is not confirming it supports your sentence. That second gap is where the errors that damage a paper actually live, and no automated checker closes it.
Five reference kinds, each checked against a live source rather than assessed.
A hallucination check resolves five reference kinds. DOI against CrossRef. arXiv ID against arXiv. PubMed ID against PubMed. ISBN against Open Library. URL by fetching it. Each returns one of three states: the record exists, the record does not exist, or it could not be checked, and the last of those is reported as unverified rather than quietly treated as fine.
That matters because it is the difference between hallucination checking that reads your reference list and a tool that merely reasons about it. A fabricated DOI comes back as absent from CrossRef, which is close to definitive in one direction: if CrossRef has no record of a DOI, the DOI is not real.
Compare the legal case. A fabricated court citation cannot be verified by any of the five, because no case-law database is queried, so for a filing the same tool is a triage aid rather than a check. Academic references sit almost entirely inside the five kinds above. The contrast is set out here, and it is worth knowing which side of it your work falls on.
Where hallucination checking weakens for research: an older monograph with no ISBN in Open Library, a conference paper with no DOI, a dataset, a working paper on an institutional repository. Those come back unverified, which is a different result from fabricated, and reading them as the same thing is the misuse to avoid.
Four levels, and only one of them is a measurement.
Verifiable means a live source confirmed the record. This is the only verdict backed by a lookup, and the response says which database answered.
Fabricated means either a lookup came back empty, which is strong, or a model judged the claim invented, which is weaker. These two arrive under the same word, so check the source field before you act on a fabricated verdict: it reads either as a verifier result with the database named, or as a model assessment.
Suspicious is always a model judgement about a possible hallucination. In practice it catches false precision, the implausibly specific figure, and the confident assertion with nothing behind it. Useful for prioritising, not evidence.
Likely true is also a judgement, and it is the one to be most careful with on your own work, because it looks like reassurance and is not.
The categories underneath are more specific than the verdicts and more useful: fake citation, invented statistic, made-up source, non-existent paper, dead link, suspicious specificity, and plausible-but-uncited. That last one is the one researchers should read first, because it is where an unsupported claim hides in otherwise sound prose.
The error that survives every automated check, and the one most likely to reach a reviewer.
Here is the failure that matters. A model drafts a sentence, attaches a genuine citation to it, and the sentence overstates what the paper found. The DOI resolves. CrossRef confirms it. The checker reports a verified source. And the claim is still wrong.
No hallucination detector can see this, because every signal it has says the reference is sound. It is also the most common way AI-assisted drafting damages a literature review: not invented papers, which are easy to catch, but real papers cited for conclusions they hedge or contradict. Reviewers find these, and finding one invites a reread of everything else.
So the workflow that follows is unglamorous and it is the only one that works. Verify existence with the hallucination checker. Verify support by reading the source. It reduces a long reference list to the subset worth opening first; it does not reduce the number you have to read.
Hedging is where drift concentrates. "Suggests a possible association" becoming "demonstrates" is a one-word change that reverses the strength of a finding, and it reads better afterwards. If you are also rewriting prose, the hedging problem is covered separately.
A clean result is not a clearance. It means nothing was caught in a 200-word passage, which is a statement about the passage and the tool, not about your paper. Journal and funder policies on AI assistance are set by them and differ; that is their document to read, not ours to summarise.
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.
For the five kinds it recognises, yes, against live sources: DOIs against CrossRef, arXiv IDs against arXiv, PubMed IDs against PubMed, ISBNs against Open Library, and URLs by fetching them. That set overlaps closely with an academic reference list, which is why this audience gets the strong version of the tool. A fabricated DOI comes back as absent from CrossRef, and that is close to definitive: if CrossRef has no record of it, the DOI is not real.
Reports it as unverified, which is deliberately a different result from fabricated. An older monograph with no Open Library record, a conference paper without a DOI, a dataset or a working paper on an institutional repository will all come back unverified. Reading unverified as fabricated is the main misuse, and reading it as fine is the other one.
No, and this is the limit that matters most for research. A genuine citation attached to an overstated claim passes every check it has: the DOI resolves, CrossRef confirms it, the source reports as verified, and the sentence is still wrong. That is the most common way AI-assisted drafting damages a literature review, and reviewers do find it. Verify existence with the tool and verify support by reading the source.
Fabricated arrives from two different routes under one word: a lookup that came back empty, which is strong evidence, or a model judging the claim invented, which is weaker. Check the source field, which reads either as a verifier result naming the database or as a model assessment. Suspicious is always a model judgement and typically catches false precision and confident claims with nothing behind them, so it is for prioritising rather than concluding.
No, and the two answer unrelated questions. An AI detector estimates whether text was machine-written. This estimates whether the claims in it are fabricated. A human can write something false and a model can state something correct, so neither result predicts the other. If what you need is the authorship question, that is a separate tool.
Two hundred words per check with a thirty-word minimum, three checks a day with no account, five on a free account, twenty-five on Starter and no daily limit on Pro. These are counted separately from the shared daily actions that gate scans and rewrites. For a full reference list, the workable pattern is to check the passages with the densest citations rather than the whole document, since the per-check cap is the binding constraint.
Where the same tool has no database to check against.
See the contrast →Six ways to catch a fabrication, ranked by what is verified.
See the comparison →The rewriting side, where a softened finding does the damage.
Read the guide →Build references once you have confirmed the source is real.
Open the tool →Three checks a day, no account. It narrows which references to open first; it does not replace opening them.