Fast-DetectGPT vs TMR, which AI detector is more accurate?
Fast-DetectGPT is far more accurate, with a balanced accuracy of 76.1% against 50.0% for TMR, though its threshold was set on our calibration split. TMR caught more AI text, 83.3% against 68.1%, but it also flagged 83.3% of human texts, against 15.9% for Fast-DetectGPT. TMR's verdicts mostly follow paragraph breaks.
Tested
Fast-DetectGPT (Qwen2.5-1.5B) vs TMR (Target Mining RoBERTa): the numbers side by side
| Measure | Fast-DetectGPT (Qwen2.5-1.5B) | TMR (Target Mining RoBERTa) |
|---|---|---|
| Rank on the leaderboard | 1 of 7 | 7 of 7 |
| Balanced accuracy | 76.1% | 50.0% |
| Catches AI text | 68.1% | 83.3% |
| Flags human text | 15.9% | 83.3% |
| Catches humanized AI text | 47.6% | 81.0% |
| AUROC | 0.815 | 0.605 |
| Flags famous public-domain texts | 56.7% | 93.3% |
| Catches Claude Fable 5.1 | 46.0% | 88.1% |
| Catches Claude Haiku 5.5 | 88.9% | 81.0% |
| Catches Claude Opus 5.5 | 63.5% | 83.3% |
| Catches Claude Sonnet 5.5 | 73.8% | 81.0% |
| Flags human text: Academic abstracts | 40.0% | 16.0% |
| Flags human text: Q&A answers | 21.4% | 100.0% |
| Flags human text: Encyclopedia articles | 8.0% | 100.0% |
| Flags human text: Fiction | 0.0% | 100.0% |
| Flags human text: News reports | 8.3% | 100.0% |
| Seconds per 1,000 words | 0.99 | 0.09 |
| Type | Zero-shot | Trained classifier |
| Licence | MIT (method), Apache-2.0 (model) | MIT |
What line breaks do to each
TMR and Fast-DetectGPT are the only detectors in our test whose catch rate moved much when we removed line breaks and scored every text as a single paragraph. The effect was very different in size.
TMR flagged all 101 human texts that had more than one paragraph. As single paragraphs, it flagged 25.7% of them, and its catch rate on AI text fell from 83.3% to 38.7%. Most of its “AI” verdicts were reactions to the layout.
Fast-DetectGPT gave exactly the same verdicts on human text either way, 15.9% flagged. Its catch rate on AI text fell from 68.1% to 57.9%. Fast-DetectGPT reads the text through Qwen2.5-1.5B, and one possible explanation is that the model finds some AI text less predictable once its paragraphs are run together. The paragraph break guide has the numbers for all seven detectors.
Famous texts
The two flagged more public-domain classics than any other detector: TMR 28 of 30 and Fast-DetectGPT 17. TMR passed only two classics, Hamlet’s “To be, or not to be” soliloquy and the Federalist No. 10, and those were exactly the two that Fast-DetectGPT flagged and no other detector did. A language model has likely seen famous passages like these many times, which would make them look predictable to Fast-DetectGPT.
Academic abstracts, the one exception
Abstracts are where TMR comes off better on false positives. It flagged 16.0% of human abstracts and Fast-DetectGPT 40.0%. That is the genre where TMR mostly sees single paragraphs, and its catch rate on AI abstracts was also 16.0%, so it passed most abstracts whoever wrote them. In every other genre, TMR flagged all human texts.
Choosing between the two
Fast-DetectGPT, without much doubt. It ranked first on our leaderboard and TMR last. Fast-DetectGPT had the advantage of a threshold tuned on our calibration split, while TMR used the 0.5 from its model card, but that cannot explain a gap this wide. TMR’s balanced accuracy of 50.0% is what a detector that flags everything would score. TMR’s even catch rates across Claude models, from 81.0% to 88.1% for Claude Fable 5.1, are likely a side effect of the same habit. For academic writing, neither detector is safe.
Questions people ask
Why does TMR have a higher catch rate than Fast-DetectGPT?
Because it flags almost any text with paragraph breaks, human or AI. With the breaks removed, TMR caught 38.7% of AI texts and Fast-DetectGPT 57.9%.
Is there anything TMR does better?
It flagged fewer human academic abstracts, 16.0% against 40.0%. But it caught only 16.0% of AI abstracts, against 84.0% for Fast-DetectGPT.
Which flagged more famous texts?
TMR, with 28 of 30 public-domain classics. Fast-DetectGPT flagged 17, the second highest count of the seven detectors.