Fakespot vs Fast-DetectGPT, which AI detector is more accurate?
Fast-DetectGPT is the better choice for most readers because it wrongly flags much less human writing, 15.9% of human texts against 34.9% for Fakespot. On AI text the two are close, catching 68.1% and 66.3%. Fakespot is safer only on academic abstracts, where Fast-DetectGPT flagged 40.0% of human ones and Fakespot 12.0%.
Tested
Fakespot AI text detector vs Fast-DetectGPT (Qwen2.5-1.5B): the numbers side by side
| Measure | Fakespot AI text detector | Fast-DetectGPT (Qwen2.5-1.5B) |
|---|---|---|
| Rank on the leaderboard | 3 of 7 | 1 of 7 |
| Balanced accuracy | 65.7% | 76.1% |
| Catches AI text | 66.3% | 68.1% |
| Flags human text | 34.9% | 15.9% |
| Catches humanized AI text | 47.6% | 47.6% |
| AUROC | 0.741 | 0.815 |
| Flags famous public-domain texts | 43.3% | 56.7% |
| Catches Claude Fable 5.1 | 50.0% | 46.0% |
| Catches Claude Haiku 5.5 | 81.0% | 88.9% |
| Catches Claude Opus 5.5 | 65.1% | 63.5% |
| Catches Claude Sonnet 5.5 | 69.0% | 73.8% |
| Flags human text: Academic abstracts | 12.0% | 40.0% |
| Flags human text: Q&A answers | 39.3% | 21.4% |
| Flags human text: Encyclopedia articles | 40.0% | 8.0% |
| Flags human text: Fiction | 12.5% | 0.0% |
| Flags human text: News reports | 70.8% | 8.3% |
| Seconds per 1,000 words | 0.07 | 0.99 |
| Type | Trained classifier | Zero-shot |
| Licence | Apache-2.0 | MIT (method), Apache-2.0 (model) |
Who gets wrongly flagged
The gap between these two is almost entirely on human writing. Fakespot flagged 44 of our 126 human texts, all published before ChatGPT existed. Fast-DetectGPT flagged 20, and the genres each one trips over mostly differ.
Fakespot’s worst genre was news: it called 70.8% of human news reports AI-written, along with 40.0% of encyclopedia articles and 39.3% of Q&A answers. Fast-DetectGPT flagged 8.3% of those news reports and 8.0% of the encyclopedia articles, and it passed every human fiction passage. Its own blind spot is academic writing. It flagged 40.0% of human abstracts, where Fakespot flagged 12.0%. The likely reason is how Fast-DetectGPT works: it treats wording that a language model finds predictable as a sign of AI, and abstracts are written in a dense, formulaic style.
Which Claude models each one catches
Fast-DetectGPT pulled ahead on the models that are easiest to detect, catching 88.9% of Claude Haiku 5.5 texts against 81.0%, and 73.8% of Claude Sonnet 5.5 texts against 69.0%. Fakespot did a little better on the harder ones, with 65.1% of Claude Opus 5.5 against 63.5%, and 50.0% of Claude Fable 5.1 against 46.0%. After a humanizer rewrite of the Opus texts, both caught 47.6%.
Keep in mind that Fast-DetectGPT’s threshold was tuned on our calibration split. Fakespot used the standard cut-off of 0.5, so it had no such help.
Speed, formatting and famous texts
Fakespot is a RoBERTa-base classifier and ran at 0.07 seconds per 1,000 words on an Apple M4 Pro. Fast-DetectGPT is a zero-shot method that runs a whole language model, Qwen2.5-1.5B, and needed 0.99 seconds, the slowest in the test.
Fast-DetectGPT also reacts to formatting. With line breaks removed it caught 57.9% of AI texts, though its verdicts on human text did not change. Fakespot strips line breaks when it cleans its input, so it gave the same answers either way (see the paragraph break test). On 30 famous public-domain texts, Fast-DetectGPT flagged 17 and Fakespot 13. Hamlet’s “To be, or not to be” soliloquy was flagged by Fast-DetectGPT and by no other detector.
Which one to pick
For news, encyclopedia text, forum answers and fiction, Fast-DetectGPT is the safer tool. For academic abstracts Fakespot raises fewer false alarms, but it caught only 33.0% of AI abstracts against 84.0% for Fast-DetectGPT, so neither is much help there. Neither should be the only evidence against a writer. Even the better of the two flagged 15.9% of human texts, and the false positives page shows how those errors add up across detectors.
Questions people ask
Is Fakespot or Fast-DetectGPT better at catching AI text?
They are close. Fast-DetectGPT caught 68.1% of AI texts and Fakespot 66.3%. Fast-DetectGPT did better on Claude Haiku 5.5 and Sonnet 5.5, Fakespot slightly better on Claude Opus 5.5 and Fable 5.1.
Which one flags fewer human texts?
Fast-DetectGPT, with 20 of 126 human texts flagged against 44 of 126 for Fakespot. Academic abstracts are the exception, where Fast-DetectGPT flagged 40.0% and Fakespot 12.0%.
Was the comparison fair to both?
Fast-DetectGPT had a small advantage. Its method has no published threshold for the model we used, so we set one on our calibration split, while Fakespot used its default cut-off of 0.5.