Can AI detectors catch Claude Haiku 5.5?
Usually. Claude Haiku 5.5 was the easiest of the four Claude models to detect. A majority of the seven free detectors we tested flagged 77.0% of its texts, Fast-DetectGPT caught 88.9% and Desklib 74.6%. Even so, roughly one Haiku text in nine slipped past the best detector.
Tested
A majority of the detectors (4 or more of 7) flagged97 of 126 Claude Haiku 5.5 texts (77.0%).
Numbers behind this chart
| Item | Rate | Texts | 95% interval |
|---|---|---|---|
| Fast-DetectGPT (Qwen2.5-1.5B) | 88.9% | 112 of 126 | 82.2 to 93.3% |
| Fakespot AI text detector | 81.0% | 102 of 126 | 73.2 to 86.9% |
| TMR (Target Mining RoBERTa) | 81.0% | 102 of 126 | 73.2 to 86.9% |
| Desklib AI text detector | 74.6% | 94 of 126 | 66.4 to 81.4% |
| MAGE | 66.7% | 84 of 126 | 58.1 to 74.3% |
| E5-small LoRA detector | 66.7% | 84 of 126 | 58.1 to 74.3% |
| OpenAI RoBERTa detector | 20.6% | 26 of 126 | 14.5 to 28.5% |
How we tested Haiku
Claude Haiku 5.5 answered the same 150 prompts as the other models, matched to the topic, genre and length of the human texts. The middle Haiku text was 277 words long, a little longer than Opus and Sonnet. We report on the 126 texts outside the calibration split.
Results
Every detector except TMR caught Haiku more often than any other Claude model. Fast-DetectGPT caught 88.9%, Fakespot 81.0%, Desklib 74.6%, and MAGE and E5-small LoRA both 66.7%. Even the OpenAI RoBERTa detector, which misses almost all modern AI text, caught 20.6% of Haiku texts, its best result.
TMR caught 81.0%, but that number means little, because it also flags 83.3% of human writing.
If you are checking text that may come from a small, fast model, detectors have their best chance. They still miss a real share, and the false positive rates on human writing stay the same whatever model you worry about.
Questions people ask
Why is Claude Haiku easier to detect than Opus?
Our test does not show the cause, only that Haiku texts look more machine-like to every detector except TMR, which flags nearly everything. Fast-DetectGPT, for example, caught 88.9% of Haiku texts and 63.5% of Opus texts.
Which AI detector is best for Claude Haiku 5.5?
Fast-DetectGPT caught the most, 88.9%, and flagged 15.9% of human texts. Desklib caught fewer, 74.6%, but flagged only 6.3% of human texts.