There is one property that separates a useful AI hallucination detector from a confident-sounding one, and almost no comparison mentions it: does it check anything outside the model? Most tools in this category ask a language model whether some text looks wrong. That is a second opinion from the same kind of system that produced the error.
So this page compares approaches rather than brand names, because the approach is what determines whether a result means anything. Our own tool is one of the six and we are disclosing that it is ours. Below there is a live test with the actual output, so you can judge the claim instead of taking it.
Ranked by whether anything is checked outside the model that might have invented it.
| Approach | Checks citations against real databases | Finds fabricated stats and quotes | Speed |
|---|---|---|---|
| Manual fact-checking | Yes, you do it | Yes, if you have time | Hours |
| Checker with live citation lookup | Yes, automatically | AI-assessed, labelled as such | Seconds |
| LLM-only "hallucination checker" | No, it asks a model | AI-assessed | Seconds |
| Asking the chatbot to check itself | No | Unreliable, it defends its own output | Seconds |
| A generic AI detector | No | No, it answers a different question | Seconds |
| A plagiarism checker | No, it finds matches | No, invented text matches nothing | Minutes |
Manual fact-checking is still the most reliable hallucination check there is and we are not going to pretend otherwise. If you have the hours, checking every claim against a primary source beats any automated tool. Everything below is about narrowing where those hours go.
Two of the six are answering a different question entirely. A generic AI detector tells you whether text was probably machine-written, which is independent of whether it is true: a human can write something false and a model can state something correct. A plagiarism checker finds text that matches an existing source, and a fabricated citation matches nothing anywhere, which is precisely why it slips through. Reaching for either is a common and expensive mistake.
Asking the chatbot to check itself is the weakest option, because the model that invented a plausible citation will usually defend it when asked. You are sampling the same distribution twice.
Feature claims in this category change quickly and we have not re-verified competitors this week. Rather than publish a comparison that is wrong by next month, the table compares the approach, which is stable and which you can determine about any tool by asking it one question: when you tell me a citation is fake, what did you check? If the answer is "a language model thought so", you know what you have.
We ran this against our own live tool on 4 October 2026. This is the actual output.
We wrote a passage with five fabrications in it: an invented researcher, an invented university, an implausibly precise statistic, a journal that does not exist, and a fake DOI.
| Flagged | Verdict | How it was determined |
|---|---|---|
| DOI 10.1234/jwb.2019.887 | fabricated | CrossRef has no record of it |
| "Journal of Workplace Botany" | fabricated | AI-assessed, no database record |
| "Dr. Helena Marsh at the University of Brightwater" | fabricated | AI-assessed |
| "87.3% … by exactly 41 minutes" | fabricated | AI-assessed, false precision |
| "12,000 participants across 14 countries" | suspicious | AI-assessed |
The hallucination detector returned five flags, a reliability score of 32 and an overall risk of high, in about seven seconds. The part that matters is the right-hand column.
One of those five was verified against an external database. Four were assessed by a model. The response labels which is which, and so does the result on screen. The citation checks run live against CrossRef, arXiv, PubMed, Open Library and live URL lookups. The statistics, quotes and events are a model's judgement, and the tool says so in its own output rather than presenting all five findings as equally established.
That split is the honest version of this product category. A tool that reports every flag with the same confidence is hiding which of its findings are evidence and which are opinion. When you evaluate any of these, look for where that line is drawn.
Three is what the free tier gives you here, so spend them deliberately.
Check one: feed it a citation you know is fake. Invent a DOI, a journal and an author, the way we did above. Then read how the tool justifies the flag. "This appears fabricated" is a guess. "CrossRef has no record of this DOI" is a check. This single test separates the category.
Check two: feed it something true but obscure. A real finding from a real paper in a narrow field. The failure mode nobody advertises is false positives on correct material, which is worse than a miss: it sends you re-verifying work that was already sound, and it teaches you to ignore the flags. A tool that flags everything is useless in a different way from one that flags nothing.
Check three: run a hallucination check on your own real draft. Not a sample. The question is whether the output changes what you would do next, which is the only test that matters.
Ours runs with no account: 3 checks a day, up to 200 words each, minimum 30 words, resetting daily. You get the per-claim breakdown and the reliability score on the free tier. Source links and the exportable report are paid. The plan ladder is in the next section.
The daily count is the gate on 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.
Three limits that apply to every tool in the table, ours included.
None of them establish that a claim is true. A citation that exists in CrossRef is a citation that exists. It does not follow that the paper says what your text claims it says, and that gap is where the most damaging errors live: a real source attached to a claim it does not support. Only reading the source closes it.
None of them are complete. Treat the output as a prioritised list of what to check first, not a clearance. A clean result on a 200-word passage means nothing was caught, not that nothing is wrong, and the flags are weighted toward the fabrications that have detectable signatures.
None of them remove your responsibility. If you are filing something with a court, submitting to a journal, or publishing to readers, the claims are yours. A tool that narrows a two-hour check to twenty minutes is worth having; one that persuades you to skip the twenty minutes is worse than nothing.
Related, and a genuinely different question: whether text was written by AI rather than whether it is true. That is the AI detector, and the two tools answer questions that are easy to confuse and not related.
Judge it on one property: whether it checks anything outside a language model. Most tools in this category ask a model whether some text looks wrong, which is a second opinion from the same kind of system that produced the error. Ours verifies citations live against CrossRef, arXiv, PubMed, Open Library and direct URL lookups, and labels those findings separately from the ones a model assessed. We disclose that this is our own product, and the page gives you a three-check test to run against it.
Ask what it checked. Invent a DOI, a journal name and an author, then read how the tool justifies the flag. "This appears fabricated" is a model guessing; "CrossRef has no record of this DOI" is a lookup against a database. In our own test the fake DOI was caught by CrossRef while the invented statistic and researcher were flagged by model assessment, and the output says which is which. A tool that reports every finding with identical confidence is concealing that distinction.
No, and it is a common and costly mix-up. An AI detector estimates whether text was machine-written, which is independent of whether it is true: a person can write something false and a model can state something correct. A plagiarism checker is equally wrong for this job, because fabricated citations match no existing source, which is exactly why they survive a similarity check. Catching fabrications needs claim-level verification, not an authorship estimate.
Yes, with no account: three checks a day at up to 200 words each, minimum thirty words, resetting daily rather than expiring as a trial. The free tier includes the per-claim breakdown and the reliability score. Clickable source links and the downloadable report are paid features, and the current plan ladder is on the plans page.
Accurate enough to be worth running and not accurate enough to be a clearance. Citation lookups are close to definitive in one direction: if CrossRef has no record of a DOI, the DOI is not real. Judgements about statistics, quotes and events are model assessments and carry both misses and false positives. Treat any result as a prioritised list of what to verify first, and be as wary of a tool that flags everything as one that flags nothing, because the first teaches you to ignore it.
No, and this is the most important limit to understand. Confirming that a citation exists is not the same as confirming it says what you claim. The most damaging errors in AI-assisted writing are real sources attached to claims they do not support, and no automated checker closes that gap. Reading the source is the only thing that does. Our own output carries this caveat rather than leaving you to discover it.
Three checks a day, no account, with the per-claim breakdown.
Check a passage →Claim-by-claim verification with a credibility score.
Open the tool →The different question: was this written by AI, not is it true.
Open the detector →Build references once you have confirmed the source is real.
Open the tool →Three checks a day, no account. Invent a DOI and see whether the tool guesses or actually looks it up.