Best AI detector for short texts under 200 words (tested)
For short texts under 200 words, Fast-DetectGPT had the best balanced accuracy, 77.9%, catching 74.2% of AI texts but flagging 7 of 38 human texts (18.4%). Desklib is the safer pick when a false flag matters, flagging 3 of 38 (7.9%) and catching 51.5%. Our shortest texts were 150 words, so these results do not cover comments or social posts.
Tested
AI detectors ranked for short texts under 200 words
| # | Detector | Balanced accuracy | Catches AI text | Flags human text |
|---|---|---|---|---|
| 1 | Fast-DetectGPT (Qwen2.5-1.5B) | 77.9% | 74.2% | 18.4% |
| 2 | Desklib AI text detector | 71.8% | 51.5% | 7.9% |
| 3 | Fakespot AI text detector | 54.5% | 43.3% | 34.2% |
| 4 | OpenAI RoBERTa detector | 53.6% | 7.2% | 0.0% |
| 5 | MAGE | 51.1% | 20.6% | 18.4% |
| 6 | E5-small LoRA detector | 45.7% | 30.9% | 39.5% |
| 7 | TMR (Target Mining RoBERTa) | 38.5% | 40.2% | 63.2% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 18.4% | 7 of 38 | 9.2 to 33.4% |
| Desklib AI text detector | 7.9% | 3 of 38 | 2.7 to 20.8% |
| Fakespot AI text detector | 34.2% | 13 of 38 | 21.2 to 50.1% |
| OpenAI RoBERTa detector | 0.0% | 0 of 38 | 0.0 to 9.2% |
| MAGE | 18.4% | 7 of 38 | 9.2 to 33.4% |
| E5-small LoRA detector | 39.5% | 15 of 38 | 25.6 to 55.3% |
| TMR (Target Mining RoBERTa) | 63.2% | 24 of 38 | 47.3 to 76.6% |
How short is short in this test
No text in our test was shorter than 150 words, so this bucket covers texts from 150 words up to 200. The test split holds 38 human texts and 97 AI texts of that length. Tweets, comments, product reviews and two-line emails are shorter still, and nothing here tells you how detectors handle them. With fewer words a detector has less evidence to work from, and we would expect results on very short text to be worse.
Which detectors cope with fewer words
Fast-DetectGPT held up best. Its balanced accuracy of 77.9% on short texts was as good as on longer ones, and it caught 74.2% of short AI texts. It also flagged 7 of 38 human texts, a rate of 18.4% with a 95% interval from 9.2% to 33.4%, and its threshold was tuned on our calibration split, which gave it an edge the other detectors lacked. Most of those false positives were academic abstracts, the genre it flags most, and abstracts are the largest group among our short human texts.
Desklib kept its false positives low, at 3 of 38, but caught only 51.5% of short AI texts, against 63.5% for texts of 200 to 299 words. The others fared worse. MAGE caught 20.6% of short AI texts while flagging 18.4% of short human ones, for a balanced accuracy of 51.1%. E5-small LoRA scored 45.7%, below chance. The OpenAI RoBERTa detector flagged no short human text but caught just 7.2% of AI ones.
TMR and single paragraphs
TMR is the odd case. On longer texts it flags nearly everything, human or AI. Under 200 words it flagged 63.2% of human texts and caught 40.2% of AI texts, a balanced accuracy of 38.5%, the lowest of any detector in any length bucket. The likely cause is formatting. TMR flags almost every text with more than one paragraph, and in our set most short AI texts were a single block, while nearly every text of 300 words or more had several paragraphs. Our paragraph break test shows the effect directly.
If you only have a short passage to judge, a flag from any of these tools is weak evidence. Compare with the results for longer texts or the full leaderboard.
Questions people ask
Are AI detectors less accurate on short text?
It depends on the detector. MAGE scored 51.1% balanced accuracy under 200 words and 71.7% at 300 words or more, while Fast-DetectGPT scored 77.9% and 77.4%. Desklib scored 71.8% on short texts.
Can an AI detector check a single paragraph?
Within limits. Our short texts start at 150 words, and many of them are a single paragraph. On those, Desklib flagged 3 of 38 human texts and caught 51.5% of AI ones.
Why did TMR score below chance on short texts?
It flagged 63.2% of short human texts but caught only 40.2% of short AI texts. A likely reason is that TMR reacts mostly to paragraph breaks, and in our set short texts were much more often a single paragraph than longer ones.