AI Detector Accuracy: Test Results
On 3,379 test texts the AI detector called 59% of AI texts "Probably AI" and did the same to 3.0% of human texts. It is strong on unedited chatbot output and weak on AI text that has been cleaned up.
Verdicts on human and AI text
Each row shows how the texts in a group were split across the three verdicts.
| Group | Texts | Probably AI | Can't tell | Probably not AI |
|---|---|---|---|---|
| All human text | 2,027 | 3.0% | 54.4% | 42.6% |
| Human, typed (essays, emails, forum answers) | 570 | 1.1% | 29.8% | 69.1% |
| Human, published (books, journals, Wikipedia, news) | 1,457 | 3.8% | 64.0% | 32.2% |
| Human, essays by English learners | 110 | 0.9% | 21.8% | 77.3% |
| AI text from all sources, ordinary prompts | 1,320 | 59.2% | 36.3% | 4.5% |
| AI, unedited chatbot output | 128 | 95.3% | 4.7% | 0.0% |
| AI, prompt asked it to sound human | 32 | 81.3% | 15.6% | 3.1% |
| AI, with dashes, curly quotes and Markdown removed | 1,320 | 44.8% | 49.0% | 6.2% |
Three things stand out.
- A "Probably AI" verdict is rarely wrong about human text. About 1 in 30 human texts received it, and 0.9% of the 110 essays by English learners.
- A "Probably not AI" verdict is weaker. 4.5% of AI texts received it, rising to 6.2% once the punctuation marks were removed.
- About half of everything is "Can't tell". A habit count cannot settle most cases, and the checker says so instead of guessing.
AI text by source
| Source | Texts | Probably AI | Can't tell | Probably not AI |
|---|---|---|---|---|
| Casual answers prompted and adjusted to match human habits | 280 | 11.1% | 76.1% | 12.9% |
| Unedited chatbot output, generated for this test | 128 | 95.3% | 4.7% | 0.0% |
| Academic and reference-style text from six 2026 models | 856 | 69.7% | 27.6% | 2.7% |
| Encyclopedic and academic paragraphs from five models | 56 | 55.4% | 42.9% | 1.8% |
The checker does best on what people actually paste: emails, essays, cover letters and posts straight from a chatbot. It does worst on a dataset whose AI answers were written to a casual register, with surface habits matched to the rate at which people show them.
AI text by model
Results for each of the 20 models in the test. Models listed with a provider prefix, such as openai/ or google/, come from the casual dataset whose answers were matched to human habits, which is why their rates are low.
| Model | Texts | Probably AI | Can't tell | Probably not AI |
|---|---|---|---|---|
| gemini-3.8-flash | 64 | 100.0% | 0.0% | 0.0% |
| gemini-3.1-pro-preview | 64 | 90.6% | 9.4% | 0.0% |
| claude-haiku-4-5 | 12 | 83.3% | 16.7% | 0.0% |
| gemini-3.6-flash | 150 | 78.0% | 22.0% | 0.0% |
| qwen3.6:35b | 150 | 77.3% | 22.0% | 0.7% |
| moonshotai/kimi-k3 | 150 | 70.7% | 26.7% | 2.7% |
| deepseek-v4-flash-0731 | 150 | 70.0% | 30.0% | 0.0% |
| luna-medium | 150 | 64.0% | 29.3% | 6.7% |
| sol-medium | 106 | 53.8% | 38.7% | 7.5% |
| gpt-4.1 | 12 | 50.0% | 50.0% | 0.0% |
| gpt-oss-120b | 12 | 50.0% | 50.0% | 0.0% |
| qwen-3.8-27b | 8 | 50.0% | 37.5% | 12.5% |
| claude-opus-5 | 12 | 41.7% | 58.3% | 0.0% |
| z-ai/glm-5.3 | 40 | 27.5% | 60.0% | 12.5% |
| openai/gpt-5.6-luna-pro | 40 | 17.5% | 75.0% | 7.5% |
| google/gemini-3.7-flash | 40 | 10.0% | 87.5% | 2.5% |
| deepseek/deepseek-v4-pro | 40 | 7.5% | 77.5% | 15.0% |
| nvidia/nemotron-3.5-lightning | 40 | 7.5% | 62.5% | 30.0% |
| qwen/qwen3.8-max | 40 | 7.5% | 80.0% | 12.5% |
| x-ai/grok-4.6 | 40 | 0.0% | 90.0% | 10.0% |
Human text by source
| Source | Texts | Probably AI | Can't tell | Probably not AI |
|---|---|---|---|---|
| arXiv papers | 140 | 2.9% | 55.0% | 42.1% |
| Work emails (2000 to 2002) | 160 | 0.6% | 17.5% | 81.9% |
| Books from Project Gutenberg | 140 | 1.4% | 42.1% | 56.4% |
| Public-domain literature | 100 | 4.0% | 75.0% | 21.0% |
| News articles (before 2020) | 93 | 2.2% | 64.5% | 33.3% |
| OpenStax textbooks | 140 | 2.9% | 65.0% | 32.1% |
| PLOS journal articles | 140 | 2.9% | 65.0% | 32.1% |
| PubMed Central articles | 140 | 3.6% | 63.6% | 32.9% |
| PubMed Central articles by non-native English authors | 183 | 4.4% | 59.6% | 36.1% |
| Stack Exchange answers | 140 | 2.1% | 37.1% | 60.7% |
| Student essays (US school grades 6 to 12) | 270 | 0.7% | 33.3% | 65.9% |
| Wikipedia and sister projects | 140 | 1.4% | 74.3% | 24.3% |
| Wikipedia (2021 revisions) | 155 | 9.0% | 80.6% | 10.3% |
By length
| Words | Texts | Probably AI | Can't tell | Probably not AI |
|---|---|---|---|---|
| 80-149, human | 568 | 3.7% | 51.1% | 45.2% |
| 80-149, AI | 251 | 49.0% | 47.4% | 3.6% |
| 150-249, human | 414 | 4.8% | 51.2% | 44.0% |
| 150-249, AI | 149 | 63.1% | 29.5% | 7.4% |
| 250-499, human | 569 | 1.9% | 53.1% | 45.0% |
| 250-499, AI | 598 | 54.5% | 40.0% | 5.5% |
| 500+, human | 476 | 1.9% | 62.8% | 35.3% |
| 500+, AI | 322 | 73.9% | 23.9% | 2.2% |
The test set
3,379 English texts of 80 to 1,000 words: 2,027 written by people and 1,352 by AI models.
- Human text comes from public datasets: the test split of rasbt/human-vs-ai-50k, the textsightai AI text detection benchmark 2026, PERSUADE 2.0 student essays, the Enron email corpus and CC-News.
- AI text comes from the same two benchmarks, from mild-rgb/eli5-human-vs-ai, and from 160 texts we generated with Gemini 3.8 Flash and Gemini 3.1 Pro using plain prompts such as "write a cover letter for a junior software developer position".
- Models covered (20): claude-haiku-4-5, claude-opus-5, deepseek-v4-flash-0731, deepseek/deepseek-v4-pro, gemini-3.1-pro-preview, gemini-3.6-flash, gemini-3.8-flash, google/gemini-3.7-flash, gpt-4.1, gpt-oss-120b, luna-medium, moonshotai/kimi-k3, nvidia/nemotron-3.5-lightning, openai/gpt-5.6-luna-pro, qwen-3.8-27b, qwen/qwen3.8-max, qwen3.6:35b, sol-medium, x-ai/grok-4.6, z-ai/glm-5.3.
How the figures were produced
The weights and cut-offs were set on four fifths of the texts and tested on the remaining fifth, five times over, so every figure above comes from texts the weights had not seen. The "Probably AI" cut-off is the score that 3% of human texts exceed. The "Probably not AI" cut-off is the score below which 9 in 10 texts are human when half of the texts checked are AI.
Limits you should know about
- The word lists were tuned on the same texts. We chose which overused words and phrases to count after looking at this test set, so expect results on new text to be somewhat lower than the figures on this page.
- Word processors and phones are underrepresented. Our typed human texts are mostly plain text. Text written in Word or on a phone can pick up automatic formatting that plain text does not have.
- Some groups are small. The unedited chatbot group has 128 texts from two models, and the group asked to sound human has 32.
- Models change. The habits counted here are those of models available in 2026. The lists need re-measuring as new models appear.
- English only. Nothing here was tested on other languages.
For what each sign is and how often it appears, see how the checker works.
Check a text for AI, free.
Open the AI detectorFigures from our test run of 2026-10-09.