# Can AI detectors detect Claude? Test results for four Claude models

Source: https://toolbenchmarks.com/guides/can-ai-detectors-detect-claude/  
Tested: 30 October 2026. Publisher: Tool Benchmarks. Data licence: CC BY 4.0.

> Sometimes, and it depends heavily on the model. In our test of seven free detectors, a majority flagged 77.0% of Claude Haiku 5.5 texts, 61.1% of Claude Sonnet 5.5, 52.4% of Claude Opus 5.5 and 38.9% of Claude Fable 5.1. The best single detector, Fast-DetectGPT, caught 88.9% of Haiku texts and 46.0% of Fable texts.

## How we tested Claude

Four Claude models, Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, each answered the same 150 writing prompts. Every prompt asked for one kind of text, such as an encyclopedia article, a news report, a forum answer, an abstract or a passage of fiction, on the topic and at the length of a human text in our set. Nobody asked the models to sound human. We then ran every answer through seven free detectors and report the 126 texts per model that were not used for calibration.

## The smaller model is the easiest to catch

Haiku 5.5 was caught most often by every detector except TMR, which flags nearly everything with paragraph breaks. Fast-DetectGPT caught 88.9% of Haiku texts, Fakespot 81.0% and Desklib 74.6%.

Fable 5.1 sat at the other end. Fast-DetectGPT caught 46.0% of its texts and Desklib 32.5%, so on the two most careful detectors most Fable texts passed as human. Opus 5.5 and Sonnet 5.5 landed in between. The differences between models were larger than the differences between the two best detectors, so the model behind a text matters as much as the detector you pick. Each model has a page with its full results, for example [Claude Fable 5.1](/can-ai-detectors-catch/claude-fable-5-1/) and [Claude Haiku 5.5](/can-ai-detectors-catch/claude-haiku-5-5/).

## The old OpenAI detector barely notices Claude

The OpenAI RoBERTa detector, built in 2019 for GPT-2 and still free to download, caught 2.4% of Opus texts and 3.2% of Fable texts. Its best result was 20.6% on Haiku. A "human" verdict from it says nothing about Claude.

## Before you trust a flag

If you need to know whether Claude wrote something, a detector gives you a partial answer at best, and a much weaker one for some models than for others. Look at the [leaderboard](/ai-detectors/) for the false positive rates before trusting any flag, because the detectors that catch the most Claude text also tend to flag human writing.

## Questions people ask

### Can Turnitin or GPTZero detect Claude?

We have not tested commercial detectors, so we cannot say. Our results cover seven open-source detectors you can run yourself.

### Which Claude model is hardest to detect?

Claude Fable 5.1. A majority of detectors flagged 38.9% of its texts, and Desklib, the most careful detector that still catches a real share of AI text, caught 32.5%.

### Does editing Claude's output make it harder to detect?

Yes. After a humanizer rewrite, Desklib's catch rate on Claude Opus 5.5 texts fell from 62.7% to 32.5%.