Can AI detectors catch Claude Sonnet 5.5?
Most of the time, but far from always. A majority of the seven free detectors we tested flagged 61.1% of Claude Sonnet 5.5 texts. Fast-DetectGPT caught 73.8% of them and Desklib 63.5%, which leaves more than a quarter of Sonnet texts passing as human even on the best detector.
Tested
A majority of the detectors (4 or more of 7) flagged77 of 126 Claude Sonnet 5.5 texts (61.1%).
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| TMR (Target Mining RoBERTa) | 81.0% | 102 of 126 | 73.2 to 86.9% |
| Fast-DetectGPT (Qwen2.5-1.5B) | 73.8% | 93 of 126 | 65.5 to 80.7% |
| Fakespot AI text detector | 69.0% | 87 of 126 | 60.5 to 76.5% |
| E5-small LoRA detector | 64.3% | 81 of 126 | 55.6 to 72.1% |
| Desklib AI text detector | 63.5% | 80 of 126 | 54.8 to 71.4% |
| MAGE | 45.2% | 57 of 126 | 36.8 to 53.9% |
| OpenAI RoBERTa detector | 11.9% | 15 of 126 | 7.3 to 18.7% |
How we tested Sonnet
Claude Sonnet 5.5 answered the same 150 prompts as the other models, one for each human text in our set, matching its topic, genre and length. The middle Sonnet text was 257.5 words long, almost the same as Opus. We report on the 126 texts outside the calibration split.
Results
Fast-DetectGPT caught 73.8% of Sonnet texts, its second best result after Haiku. Fakespot caught 69.0%, E5-small LoRA 64.3% and Desklib 63.5%. As with every model, the catch rates of Fakespot and E5-small LoRA come with high false positive rates on human writing, 34.9% and 50.0%.
MAGE caught 45.2%, the same as for Opus, and the OpenAI RoBERTa detector caught 11.9%. TMR’s 81.0% is not a real result, because it flags most human writing too.
Sonnet sits in the middle of the four Claude models we tested: easier to catch than Opus and Fable 5.1, harder than Haiku 5.5.
Questions people ask
Which detector catches Claude Sonnet 5.5 best?
Fast-DetectGPT, at 73.8% of Sonnet texts. TMR flagged more, 81.0%, but it flags 83.3% of human writing too, so its result means nothing.
Is Sonnet easier to detect than Opus?
Slightly. A majority of detectors caught 61.1% of Sonnet texts and 52.4% of Opus texts.