What we tested, and a correction

This page used to put UnTXT ahead of GPTZero and Turnitin. That was wrong, and here is what we actually know.

Correction, 30 September 2026. Until today this page said UnTXT scored an F1 of 0.73, ahead of GPTZero at 0.68 and Turnitin AI at 0.65, and it credited those two figures to a 2023 study by Weber-Wulff and colleagues. That was wrong. The link we gave went to a paper about mapping coral reefs, and when we found the real study, in the International Journal for Educational Integrity, we couldn't trace the figures to it. Our own 0.73 was never fit to stand next to them anyway, as the rest of this page explains. So we've taken the comparison off this page, off the sign in page, off the infographic and off our Chrome Web Store tile. If you read it and believed it, we're sorry.

June 2026: our first engine, on 935 texts

We took 500 ChatGPT answers and 435 human answers from HC3, a research dataset from 2023, and scored each one with the engine we had then. That engine was a set of rules about sentence shape and word choice, and it has since been retired.

Measure	Result	What it means
AI texts flagged	460 of 500	At the threshold that gave the best F1
Human texts flagged	302 of 435	At that same threshold
Best F1	0.73	Threshold picked after seeing these same texts, which flatters it
AI caught at 5% false alarms	11%	The fairer number for anyone marking real work
AUC	0.668	0.5 is a coin toss, 1.0 is perfect

Put plainly: it caught most of the AI text, and it flagged most of the human text as well. That is the result that sent us off to build Analysis, and it is a big part of why UnTXT shows no percentage at all now.

By subject

HC3 subset	Texts	Best F1	AUC
Open QA	135	0.980	0.981
Reddit ELI5	200	0.709	0.705
Medicine	200	0.705	0.620
Finance	200	0.685	0.588
Wikipedia CS	200	0.667	0.615

Open QA has only 35 human texts in it, so its figures lean heavily on the AI side.

September 2026: a small pilot of Analysis

In September we ran Analysis over 60 HC3 texts, half from ChatGPT and half from people, because we wanted to compare two versions of the second model it uses. Both got all 60 right. Sixty texts is a pilot rather than a result, though, so we don't quote it as an accuracy figure. And since HC3's answers come from the ChatGPT of 2023, which Analysis now spots easily, a bigger run on HC3 alone would actually tell us less than it sounds like it would.

October 2026: the same rules, after fixing a bug

In October we found that one of the checks in our rule-based engine was running backwards, so it marked down writing with uneven sentence structure and marked up writing where every sentence was built the same way. We fixed it, along with two smaller faults that made short texts look more suspicious than they should, and then ran the same 935 texts again through the live engine. It is still the rule-based engine, the one behind the API's legacy mode and the free checks in the browser extension, and not Analysis.

Measure	June	October	What it means
AUC	0.668	0.810	How well it puts AI texts above human ones
AI caught at 5% false alarms	11%	29%	The fairer number for anyone marking real work
Best F1	0.73	0.79	Threshold picked after seeing these same texts, which flatters it
Human texts flagged at 0.20	302 of 435 at the best-F1 setting	48 of 435	0.20 is the setting the engine ships with
AI texts flagged at 0.20	460 of 500	218 of 500	Same setting

Put plainly: it now orders the texts much better, and at the setting it ships with it flags about one human text in nine, down from most of them. The price is that it misses more than half of the AI text at that setting. Moving the setting trades one for the other, and the best F1 sits at a looser one, where it would have flagged 170 of the 435 human texts. A score from this engine is still a reason to read more closely and never a reason to accuse somebody, which is why UnTXT doesn't show you a percentage.

What has been published about other tools

We have not run anybody else's detector. These are other people's published results, and they are here so you can see where the field stands, not as a ranking.

On HC3: GPTZero scored an F1 of 99.17 on the HC3 test set of 25,049 texts, with its threshold set on a separate sample (arXiv 2406.08922, Table 2). We found no published HC3 figures for ZeroGPT or Originality.ai.
On RAID: RAID (ACL 2024) tests GPTZero, ZeroGPT, Originality.ai and Winston on text from eleven newer models, and reports how much AI text each catches while wrongly flagging 5% of human text. Its results are on a public leaderboard.

What comes next

Next is a full run of Analysis on HC3 and on the matching parts of RAID, scored the way those papers score them. After that, we want somebody outside the company to run the test on current models, and to measure what we care about most, which is whether the feedback helps the person reading it. When there are numbers, they'll go on this page, including any we'd rather not publish.

How the June run worked

Dataset: HC3 English (Guo and colleagues, 2023, shared under a Creative Commons licence): questions from open QA, medicine, finance, Wikipedia computing and Reddit ELI5, each with human and ChatGPT answers.

Sampling: up to 100 human and 100 AI answers per subject, between 75 and 2,000 words each, random seed 42. That left 435 human and 500 AI texts.

Scoring: each text went through our detection API, which returned a score between 0 and 1. Nobody looked at the texts by hand.

Threshold: we tried every threshold from 0 to 1 in steps of 0.01 and kept the one with the best F1, which was 0.20. Choosing it on the same texts you then score makes the result look better than it would on fresh ones.

Limits worth knowing

Old AI text. HC3 was written by ChatGPT in 2023. Nothing here measures text from current models.
Short answers. Most HC3 answers are shorter than the 250 words Analysis asks for in the app, so benchmark runs lower that limit.
Not like for like. The published figures above were measured on different samples, with thresholds chosen in different ways, so putting them beside ours would only repeat the mistake this page is correcting.

Checking our arithmetic

The benchmark code and the score for every text sit in our repository, which is private. If you want the scores for each text to check the numbers above, write to us and we will send them.