The limitation goes first, because for lawyers it decides how you use this at all. Our citation verifier queries CrossRef, arXiv, PubMed, Open Library and live URLs. Those are academic and bibliographic databases. None of them contains case law. So when this tool flags a fabricated case, that is a model judgement, not a database lookup.
An AI hallucination detector is still worth running here, and the test below shows why: it did flag both invented cases we fed it. But a flag is a place to look, not a finding. Every citation in a filing still has to be pulled and read. This page is about what the tool adds to that work and what it cannot take off your desk.
The gap between flagging a case and verifying one, which matters more on a filing than anywhere else.
The citation verifier recognises five kinds of reference and checks each against a live source: 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.
A fabricated case citation, the signature AI hallucination in legal drafting, is none of those. There is no connection to CourtListener, Westlaw, Lexis or any reporter. So a fabricated case is assessed the way a human skim-reader would assess it: does this reporter volume, circuit and year hang together, is this case known. That is a useful signal and it is not verification, and the output labels it as such rather than letting you assume otherwise.
Three things that genuinely come up. Expert literature: a journal article cited in support of an expert opinion usually carries a DOI or a PubMed ID, and those are checked properly. Treatises: an ISBN is checked against Open Library. Links: any URL is fetched, so a plausible-looking link to a court or agency site that does not resolve is caught outright, which is one of the more common artefacts of AI-assisted drafting.
If your filing leans on scientific or medical literature, that is where hallucination checking is actually strong. If it leans on case law, it is a triage aid.
Run against our live tool on 4 October 2026. This is the actual output.
We gave the hallucination detector a short passage containing two invented cases, an invented holding, a plausible-looking link to a federal courts domain, and an invented Restatement section.
| Flagged | Verdict | How it was determined |
|---|---|---|
| Hendricks v. Calloway Freight Systems, 814 F.3d 1122 (9th Cir. 2016) | fabricated | Model judgement |
| Olmstead v. Baytide Logistics, 2019 WL 4471883 | fabricated | Model judgement |
| "waives its limitation defence by failing to plead it within thirty days" | suspicious | Model judgement, flagged as uncited |
| A link to a federal courts domain | fabricated | Fetched. It does not resolve. |
| Restatement (Third) of Carriage Law 44 (2018) | fabricated | Model judgement |
Five flags, reliability 32, overall risk high. Both fake cases were caught, which is the useful result.
Now the number that frames everything: the response reported one source actually checked. The dead link. The other four flags are model assessment. So the tool was right about the cases and could not prove it, and on a document going to a court that distinction is the whole point. A flag tells you where to look. Confirming a case exists and says what you claim is still done in a real reporter.
As a triage step before the citation check, never instead of it.
Run a hallucination check on the draft before you start pulling cites. Its value is ordering your work: it tells you which of forty citations to pull first, and it reliably surfaces the specific-sounding detail that nobody inserted deliberately. On a long brief that is a real saving on a task that is otherwise uniformly slow.
Then pull every case anyway. Including the ones it did not flag. A clean result on a passage means nothing was caught, not that nothing is wrong, and a model that has never seen your jurisdiction's unpublished dispositions will not reliably notice a wrong one.
Treat an unflagged quotation with the same suspicion as a flagged one. The failure that survives hallucination checking is a real case cited for a proposition it does not support. Nothing in the output can see that, because the case exists and the sentence is fluent. Only reading the opinion closes it, and it is the error most likely to matter.
Whether using this discharges any duty you are under. It does not and we are not in a position to advise on it. Verification obligations, candour to the court and any local rule or standing order about AI-assisted drafting are matters for your jurisdiction, your bar and your firm, not for a vendor page. We are describing what a tool checks. The responsibility for what is in a filing stays where it already was.
On confidentiality, one plain statement: the passage you paste is sent to a model to be processed, because that is how the tool works. For privileged or client-confidential material, that is a decision to take against your own obligations and your firm's policy before pasting anything, not after. Our privacy policy governs retention.
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 humanization | 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.
It can flag one, and in our own test it flagged both invented cases we gave it. What it cannot do is verify one, because no case-law database is queried: the citation verifier checks DOIs against CrossRef, arXiv IDs against arXiv, PubMed IDs against PubMed, ISBNs against Open Library and URLs by fetching them. A case citation is none of those, so the verdict is a model judgement and the output labels it that way. Treat a flag as a place to look and pull every case regardless.
Three things properly. Journal articles cited in support of expert opinion, where a DOI or PubMed ID is checked against the real record. Treatises, where an ISBN is checked against Open Library. And links, which are fetched, so a plausible-looking URL to a court or agency site that does not resolve is caught outright. In our test the dead link was the single item the tool verified rather than assessed, out of five flags.
No, and we are not in a position to advise on it. Verification obligations, candour to the court, and any local rule or standing order about AI-assisted drafting are matters for your jurisdiction, your bar and your firm. What we can describe is what the hallucination checker verifies and what it only assesses. The responsibility for the contents of a filing is unchanged by running any software over it.
A real case cited for a proposition it does not support. The citation exists, the sentence reads fluently, and nothing in the output can see the mismatch. It is also the error most likely to cause you a problem, which is why a clean result is not a clearance. Reading the opinion is the only thing that closes that gap, and it applies equally to citations the tool did not flag.
The passage is sent to a model to be processed, because that is how a hosted tool of this kind works. Whether that is acceptable for privileged or client-confidential material is a decision to take against your own professional obligations and your firm policy before pasting, and our privacy policy is the document that governs retention. If material must not leave your systems, a cloud tool is the wrong category rather than the wrong vendor.
Three a day with no account, at up to 200 words a check and a thirty-word minimum. A free account gives five a day, Starter twenty-five, and Pro has no daily limit. These are counted separately from the shared daily actions that gate scans and rewrites, so a hallucination check does not spend one of those. Clickable source links and the downloadable report are paid features.
Three checks a day, no account, with the per-claim breakdown.
Check a passage →Six ways to catch a fabrication, ranked by what is actually verified.
See the comparison →The different question: was this drafted by AI, not is it true.
Read the guide →Detection across a team, and what the scores mean.
Read the guide →Three checks a day, no account. It will tell you where to look first; the reporter still tells you what is true.