# Tool Benchmarks

> 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.

Independent tests of open-source AI text detectors. Last run: 30 October 2026, 756 test texts. Every number is downloadable; data licence CC BY 4.0. Full text of every page: https://toolbenchmarks.com/llms-full.txt

## Results

- [AI detector leaderboard](https://toolbenchmarks.com/ai-detectors/): detectors ranked by balanced accuracy
- [False positives](https://toolbenchmarks.com/false-positives/): human writing flagged as AI
- [Humanized AI text](https://toolbenchmarks.com/humanized-ai-text/): detection after editing
- [Fast-DetectGPT (Qwen2.5-1.5B)](https://toolbenchmarks.com/ai-detectors/fast-detectgpt/): catches 68.1% of AI text, flags 15.9% of human text
- [Desklib AI text detector](https://toolbenchmarks.com/ai-detectors/desklib/): catches 58.3% of AI text, flags 6.3% of human text
- [Fakespot AI text detector](https://toolbenchmarks.com/ai-detectors/fakespot/): catches 66.3% of AI text, flags 34.9% of human text
- [MAGE](https://toolbenchmarks.com/ai-detectors/mage/): catches 46.4% of AI text, flags 19.0% of human text
- [E5-small LoRA detector](https://toolbenchmarks.com/ai-detectors/e5-lora/): catches 61.1% of AI text, flags 50.0% of human text
- [OpenAI RoBERTa detector](https://toolbenchmarks.com/ai-detectors/openai-roberta/): catches 9.5% of AI text, flags 4.8% of human text
- [TMR (Target Mining RoBERTa)](https://toolbenchmarks.com/ai-detectors/tmr/): catches 83.3% of AI text, flags 83.3% of human text
- [Can AI detectors catch Claude Fable 5.1?](https://toolbenchmarks.com/can-ai-detectors-catch/claude-fable-5-1/): a majority of detectors caught 38.9%
- [Can AI detectors catch Claude Haiku 5.5?](https://toolbenchmarks.com/can-ai-detectors-catch/claude-haiku-5-5/): a majority of detectors caught 77.0%
- [Can AI detectors catch Claude Opus 5.5?](https://toolbenchmarks.com/can-ai-detectors-catch/claude-opus-5-5/): a majority of detectors caught 52.4%
- [Can AI detectors catch Claude Sonnet 5.5?](https://toolbenchmarks.com/can-ai-detectors-catch/claude-sonnet-5-5/): a majority of detectors caught 61.1%
- [Do AI detectors flag famous texts?](https://toolbenchmarks.com/famous-texts/): 30 of 30 public-domain classics were flagged by at least one detector
- [Best AI detector for academic writing (tested)](https://toolbenchmarks.com/best-ai-detector-for/academic-writing/)
- [Best AI detector for encyclopedia and Wikipedia-style articles (tested)](https://toolbenchmarks.com/best-ai-detector-for/encyclopedia-articles/)
- [Best AI detector for fiction and creative writing (tested)](https://toolbenchmarks.com/best-ai-detector-for/fiction/)
- [Best AI detector for forum posts and Q&A answers (tested)](https://toolbenchmarks.com/best-ai-detector-for/forum-posts/)
- [Best AI detector for long texts of 300 words or more (tested)](https://toolbenchmarks.com/best-ai-detector-for/long-texts/)
- [Best AI detector for news articles (tested)](https://toolbenchmarks.com/best-ai-detector-for/news-articles/)
- [Best AI detector for short texts under 200 words (tested)](https://toolbenchmarks.com/best-ai-detector-for/short-texts/)
- [Head-to-head comparisons](https://toolbenchmarks.com/compare/): every pair of detectors

## Guides

- [Can AI detectors detect Claude? Test results for four Claude models](https://toolbenchmarks.com/guides/can-ai-detectors-detect-claude/): We ran 504 texts from four Claude models through seven free AI detectors. A majority caught between 38.9% and 77.0% of them, depending on the model.
- [Falsely accused of using AI? What the detector data shows](https://toolbenchmarks.com/guides/falsely-accused-of-using-ai/): AI detectors flag real human writing. In our test, 117 of 126 human texts were flagged by at least one of seven detectors. How to respond with evidence.
- [How accurate are AI detectors? What our test of 900 texts shows](https://toolbenchmarks.com/guides/how-accurate-are-ai-detectors/): Free AI detectors caught between 9.5% and 68.1% of AI text in our test while flagging real human writing. What the numbers mean if you have to judge a text.
- [How AI detectors work, and why they flag human writing](https://toolbenchmarks.com/guides/how-ai-detectors-work/): AI detectors either learn the style of AI text or measure how predictable each word is. Both approaches flag predictable human writing. Examples from our tests.
- [Paragraph breaks can change an AI detector's verdict](https://toolbenchmarks.com/guides/paragraph-breaks-change-ai-detector-results/): We scored the same texts with and without line breaks. One detector went from flagging 100.0% of multi-paragraph human texts to 25.7%. Results for seven detectors.
- [Perplexity and burstiness in AI detection, explained](https://toolbenchmarks.com/guides/perplexity-and-burstiness/): Perplexity measures how predictable a text is to a language model; burstiness how much that varies. Why both flag human writing, with results from our tests.
- [How to run a free AI detector on your own computer](https://toolbenchmarks.com/guides/run-an-ai-detector-yourself/): Run Desklib, the most careful detector in our test, locally with Python. Working code, speed on a laptop, and how to read the score.
- [Should teachers use AI detectors? What our test data says](https://toolbenchmarks.com/guides/should-teachers-use-ai-detectors/): The most careful free detector we tested flagged 6.3% of human texts. What that means for grading, and how to use a detector without accusing an honest student.
- [What AUROC means for AI detectors, explained with real results](https://toolbenchmarks.com/guides/what-is-auroc/): AUROC measures how well a detector's scores separate AI from human text, from 0.5 (chance) to 1 (perfect). Our seven detectors scored between 0.579 and 0.877.

## Data and method

- [Methodology](https://toolbenchmarks.com/methodology/)
- [Data downloads](https://toolbenchmarks.com/data/): results.json, scores.csv, texts.jsonl