How accurate are AI detectors? We tested 7 free ones on 900 texts
No free AI detector we tested is reliable enough to judge someone's work. The best two, Fast-DetectGPT and Desklib, reached 76.1% and 76.0% balanced accuracy. Desklib flagged the fewest human texts (6.3%) but caught only 58.3% of AI text, and 117 of 126 human texts were flagged by at least one detector.
Tested
Which AI detector is the most accurate?
| # | Detector | Balanced accuracy | Catches AI text | Flags human text | AUROC | Seconds per 1,000 words |
|---|---|---|---|---|---|---|
| 1 | Fast-DetectGPT (Qwen2.5-1.5B) | 76.1% | 68.1% | 15.9% | 0.815 | 0.99 |
| 2 | Desklib AI text detector | 76.0% | 58.3% | 6.3% | 0.877 | 0.57 |
| 3 | Fakespot AI text detector | 65.7% | 66.3% | 34.9% | 0.741 | 0.07 |
| 4 | MAGE | 63.7% | 46.4% | 19.0% | 0.666 | 0.39 |
| 5 | E5-small LoRA detector | 55.6% | 61.1% | 50.0% | 0.582 | 0.03 |
| 6 | OpenAI RoBERTa detector | 52.4% | 9.5% | 4.8% | 0.579 | 0.07 |
| 7 | TMR (Target Mining RoBERTa) | 50.0% | 83.3% | 83.3% | 0.605 | 0.09 |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 68.1% | 343 of 504 | 63.9 to 72.0% |
| Desklib AI text detector | 58.3% | 294 of 504 | 54.0 to 62.6% |
| Fakespot AI text detector | 66.3% | 334 of 504 | 62.0 to 70.3% |
| MAGE | 46.4% | 234 of 504 | 42.1 to 50.8% |
| E5-small LoRA detector | 61.1% | 308 of 504 | 56.8 to 65.3% |
| OpenAI RoBERTa detector | 9.5% | 48 of 504 | 7.3 to 12.4% |
| TMR (Target Mining RoBERTa) | 83.3% | 420 of 504 | 79.8 to 86.3% |
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 15.9% | 20 of 126 | 10.5 to 23.2% |
| Desklib AI text detector | 6.3% | 8 of 126 | 3.3 to 12.0% |
| Fakespot AI text detector | 34.9% | 44 of 126 | 27.2 to 43.6% |
| MAGE | 19.0% | 24 of 126 | 13.1 to 26.8% |
| E5-small LoRA detector | 50.0% | 63 of 126 | 41.4 to 58.6% |
| OpenAI RoBERTa detector | 4.8% | 6 of 126 | 2.2 to 10.0% |
| TMR (Target Mining RoBERTa) | 83.3% | 105 of 126 | 75.9 to 88.8% |
Which AI models do detectors catch?
| Detector | Claude Fable 5.1 | Claude Haiku 5.5 | Claude Opus 5.5 | Claude Sonnet 5.5 | Humanized |
|---|---|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 46.0% | 88.9% | 63.5% | 73.8% | 47.6% |
| Desklib AI text detector | 32.5% | 74.6% | 62.7% | 63.5% | 32.5% |
| Fakespot AI text detector | 50.0% | 81.0% | 65.1% | 69.0% | 47.6% |
| MAGE | 28.6% | 66.7% | 45.2% | 45.2% | 23.8% |
| E5-small LoRA detector | 49.2% | 66.7% | 64.3% | 64.3% | 47.6% |
| OpenAI RoBERTa detector | 3.2% | 20.6% | 2.4% | 11.9% | 0.0% |
| TMR (Target Mining RoBERTa) | 88.1% | 81.0% | 83.3% | 81.0% | 81.0% |
What we tested
We collected 150 human-written texts from five kinds of writing and asked four Claude models to write 600 texts on the same topics at the same length. We also rewrote 150 of the AI texts with a humanizer, a tool that removes typical AI phrasing, to see whether light editing changes the result. That makes 900 texts. Every detector scored every text with the threshold its authors publish, and we report results on the 756 texts we did not use to tune anything.
All seven detectors are free, open-source models you can run on your own computer. Commercial services such as GPTZero and Turnitin are not in this test, because we only publish numbers we measured ourselves.
Two detectors are worth a look, and neither is good enough to accuse anyone
Fast-DetectGPT caught 68.1% of AI texts while flagging 15.9% of human texts. Desklib caught less, 58.3%, but flagged only 6.3% of human texts, the lowest false positive rate of any detector that catches a real share of AI writing. Their balanced accuracy is a near tie. Fast-DetectGPT also had an advantage the others did not: it ships without a threshold, so we set one on our calibration texts.
The OpenAI RoBERTa detector looks safe at first glance, with 4.8% false positives. It also caught only 9.5% of AI text. OpenAI built it in 2019 to spot GPT-2, and it does not recognise modern chatbot writing.
A high catch rate can hide a useless detector
TMR caught 83.3% of AI texts, more than any other detector. It also flagged 83.3% of human texts, so its catch rate tells you nothing. It flagged every human news report, encyclopedia article, forum answer and fiction passage in the test set, while most academic abstracts passed. Abstracts are usually a single paragraph, and that turned out to matter, as the paragraph break test shows. E5-small LoRA had the same problem in a milder form and flagged half of all human writing.
When you compare detectors, read the false positive rate first. A detector that flags everything will always catch the AI text.
Some Claude models are much harder to detect
Haiku 5.5 was the easiest model to catch: a majority of detectors flagged 77.0% of its texts. Fable 5.1 was the hardest, at 38.9%. Opus 5.5 and Sonnet 5.5 landed in between, at 52.4% and 61.1%. Each model has its own page, starting with Claude Opus 5.5.
Light editing also works against detectors. After our humanizer rewrite, Desklib’s catch rate on Opus texts fell from 62.7% to 32.5%. The humanized text results cover every detector.
If you are judging someone’s writing
A flag from one of these detectors does not show that a person used AI. Of the 126 human texts in our test set, 117 were flagged as AI by at least one detector. Real writing also comes with paragraph breaks, which is exactly what tripped TMR. The false positives page shows real pre-2022 writing that the detectors called AI.
More results
- Best AI detector for academic writing
- Best AI detector for encyclopedia articles
- Best AI detector for fiction
- Best AI detector for forum posts and Q&A answers
- Best AI detector for texts of 300 words or more
- Best AI detector for news articles
- Best AI detector for short texts under 200 words
- Famous texts flagged as AI
- Every pair of detectors compared
- Human writing flagged as AI
Questions people ask
Can AI detectors be trusted?
Not on their own. In our test even the most careful detector, Desklib, flagged 6.3% of human-written texts as AI, and the detector that caught the most AI text (TMR, 83.3%) flagged 83.3% of human texts too. A detector score is a reason to look closer, never proof.
Which free AI detector is the most accurate?
Fast-DetectGPT and Desklib tied at the top with about 76% balanced accuracy. Desklib is the safer choice if false accusations worry you, because it flagged 6.3% of human texts against Fast-DetectGPT's 15.9%.
Can AI detectors tell which AI wrote a text?
No. They only score how machine-like a text looks, and they miss some models far more than others. A majority of detectors caught 77.0% of Claude Haiku 5.5 texts but only 38.9% of Claude Fable 5.1 texts.
How did you know which texts were written by people?
Every human text was published before ChatGPT came out on 30 November 2022, on Wikipedia, Wikinews, Stack Exchange, arXiv or Project Gutenberg. The AI texts were written for this test by four Claude models, answering prompts on the same topics.