Best AI detector for long texts of 300 words or more (tested)
Desklib is the best AI detector for texts of 300 words or more in our test. It flagged none of 28 human texts and caught 54.4% of AI texts, a balanced accuracy of 77.2%. Fast-DetectGPT scored 77.4% by catching more AI text, 69.0%, but it flagged 4 of 28 human texts. Our longest texts were a few hundred words, so multi-page essays were not tested.
Tested
AI detectors ranked for texts of 300 words or more
| # | Detector | Balanced accuracy | Catches AI text | Flags human text |
|---|---|---|---|---|
| 1 | Fast-DetectGPT (Qwen2.5-1.5B) | 77.4% | 69.0% | 14.3% |
| 2 | Desklib AI text detector | 77.2% | 54.4% | 0.0% |
| 3 | MAGE | 71.7% | 57.6% | 14.3% |
| 4 | Fakespot AI text detector | 59.7% | 76.6% | 57.1% |
| 5 | E5-small LoRA detector | 58.4% | 81.0% | 64.3% |
| 6 | OpenAI RoBERTa detector | 52.0% | 7.6% | 3.6% |
| 7 | TMR (Target Mining RoBERTa) | 50.0% | 100.0% | 100.0% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 14.3% | 4 of 28 | 5.7 to 31.5% |
| Desklib AI text detector | 0.0% | 0 of 28 | 0.0 to 12.1% |
| MAGE | 14.3% | 4 of 28 | 5.7 to 31.5% |
| Fakespot AI text detector | 57.1% | 16 of 28 | 39.1 to 73.5% |
| E5-small LoRA detector | 64.3% | 18 of 28 | 45.8 to 79.3% |
| OpenAI RoBERTa detector | 3.6% | 1 of 28 | 0.6 to 17.7% |
| TMR (Target Mining RoBERTa) | 100.0% | 28 of 28 | 87.9 to 100.0% |
More words help some detectors and hurt others
A longer text gives a detector more to measure, and MAGE used it. Its balanced accuracy rose from 51.1% on texts under 200 words to 71.7% at 300 words or more, and it caught 57.6% of long AI texts against 20.6% of short ones. Fast-DetectGPT barely moved between length buckets, scoring 77.9% on short texts and 77.4% on long ones, while Desklib rose from 71.8% to 77.2%.
Two detectors went the other way. Fakespot flagged 34.2% of short human texts and 57.1% of long ones, and E5-small LoRA 39.5% and 64.3%. Both also caught more long AI text, 76.6% and 81.0%, but their balanced accuracy on long texts stayed low, at 59.7% and 58.4%. Our data does not show why. One possibility is that both models treat the structure of a longer piece, several paragraphs on one topic, as a sign of AI. TMR flagged all 28 long human texts and all 158 long AI texts, which fits its habit of flagging any multi-paragraph text.
Desklib or Fast-DetectGPT for long documents
On balanced accuracy the two are tied for practical purposes, 77.4% for Fast-DetectGPT and 77.2% for Desklib. They differ in which mistake they make. Fast-DetectGPT caught 69.0% of long AI texts and flagged 4 of 28 human ones (14.3%, with a 95% interval from 5.7% to 31.5%). Desklib caught 54.4% and flagged none. Fast-DetectGPT’s threshold was also chosen on our calibration split, a step the other detectors did not get, so its score here is a little flattering.
If a flag could cost someone a grade or a job, Desklib’s profile is the safer one. Its clean record rests on 28 texts, though, and the interval for its false positive rate runs up to 12.1%. It also missed many AI texts, so a pass is no proof of human writing.
What “long” means here
The bucket starts at 300 words and ends at the length of our longest texts, a few hundred words. The 28 human texts in it come from encyclopedia articles, news reports, Q&A answers and fiction. None are academic abstracts, which are shorter, so the abstract problem described on the academic writing page does not show up in these numbers. The leaderboard has the results for all lengths combined.
Questions people ask
Are AI detectors more accurate on longer texts?
Some are. MAGE went from 51.1% balanced accuracy under 200 words to 71.7% at 300 words or more. Fast-DetectGPT changed little, and Desklib rose from 71.8% to 77.2%. Fakespot and E5-small LoRA flagged more human writing as length grew.
How long should a text be for an AI detector to work?
Longer is not always better. Desklib's best length bucket was 200 to 299 words, where it scored 77.6% balanced accuracy and caught 63.5% of AI texts, against 54.4% at 300 words or more.
Can I use these results for a student essay?
Only loosely. Our long human texts are encyclopedia articles, news reports, forum answers and fiction passages of a few hundred words, not student essays. On them Desklib flagged 0 of 28 human texts, with a 95% interval up to 12.1%.