How accurate is ProofAI?
We test ProofAI the way you use it and publish the results, including the weak spots. A detector score is a probability, not proof: use it as one signal among several.
AI image detection: results
109 images: 52 real photos and 57 AI-generated images from 12+ generators, scored by the production image detector on 2026-10-16.
| Measure | Result |
|---|---|
| Accuracy | 86.2% (94/109; 95% CI 78.5–91.5%) |
| Real photos wrongly flagged as AI (false-positive rate) | 0.0% (0/52; 95% CI 0–6.9%) |
| AI images missed (false-negative rate) | 26.3% (15/57; 95% CI 16.6–39.0%) |
| AUC (ranking quality, 1.0 = perfect) | 0.949 |
| Results in the unclear band (35–65%) | 9.2% of images |
| Median time per image | 7.8 s |
Confidence intervals are Wilson 95% intervals. Small samples give wide intervals; treat them as estimates, not guarantees.
By image generator
How many AI images from each generator were flagged as AI. Most groups are small (n = 3), so read them as indications, not rates.
| Generator | Images (n) | Flagged as AI | Average AI probability |
|---|---|---|---|
| Real photo (Wikimedia Commons) | 52 | 0/52 | 19.3% |
| Gemini 3.1 Flash Image (via OmniRoute) | 13 | 5/13 | 54.9% |
| Stable Diffusion 1.x (DiffusionDB) | 6 | 6/6 | 81.0% |
| Adobe Firefly | 3 | 1/3 | 46.5% |
| Bing Image Creator (DALL-E 3) | 3 | 3/3 | 88.0% |
| DALL-E (OpenAI) | 3 | 3/3 | 85.8% |
| Flux (Black Forest Labs) | 3 | 3/3 | 82.3% |
| GPT Image / ChatGPT (OpenAI) | 3 | 2/3 | 55.3% |
| Gemini / Imagen (Google) | 3 | 3/3 | 82.7% |
| Grok (xAI) | 3 | 3/3 | 85.7% |
| Imagen (Google) | 3 | 2/3 | 64.7% |
| Meta AI | 3 | 3/3 | 91.1% |
| Midjourney | 3 | 3/3 | 82.5% |
| Stable Diffusion | 3 | 3/3 | 86.2% |
| StyleGAN (thispersondoesnotexist.com) | 3 | 0/3 | 17.2% |
| Pollinations.ai (sana) | 2 | 2/2 | 82.9% |
Where image detection is weak
Photorealistic faces from GAN models (such as StyleGAN), some Adobe Firefly images and photoreal images from newer Gemini image models were often scored as real photos. Heavily compressed images (screenshots, forwarded images) lose detail and give less reliable results.
AI text detection: results after the 17 Oct 2026 engine fix
On 2026-10-16 our text detector wrongly flagged 62% of formal human writing as AI. The cause was a model that timed out and was scored as 0. We replaced the primary model, calibrated the score on separate data, and re-tested twice: on our own held-out set and on the public Stanford "GPT detectors are biased" set (749 texts, run once, nothing tuned on it).
| Measure | Before the fix | After the fix |
|---|---|---|
| Our held-out set: human texts flagged as AI (test + fresh, n=135) | 57.8% (78/135) | 3.7% (5/135; 95% CI 1.6–8.4%) |
| Our held-out set: AI texts missed (n=90) | 14.4% | 10.0% (95% CI 5.4–17.9%) |
| Our held-out set: AUC | 0.722 | 0.988 |
| Stanford: non-native (TOEFL) essays flagged as AI | 68.1% (62/91) | 28.6% (26/91; 95% CI 20.3–38.6%) |
| Stanford: native college essays flagged as AI | 1.4% (1/70) | 2.9% (2/70) |
| Stanford: gap between non-native and native | +66.7 points | +25.7 points |
| Stanford: formal human abstracts (CS224N) flagged as AI | 82.8% (120/145) | 17.2% (25/145; 95% CI 12.0–24.2%) |
| Stanford: AI essays caught (n=443) | 94.4% | 92.1% (408/443) |
| Stanford: AUC | 0.763 | 0.953 |
AI means a score of 65 or more; 35–65 is unclear. Our pass marks were: non-native flagged under 10%, gap under 5 points, formal abstracts flagged under 10%, AI caught over 70%, AUC over 0.90. We met the last two and missed the first three, so non-native and formal human writing is still flagged too often. Stanford data: Liang et al. 2023 (MIT licence). Download scores only (no texts): Stanford text results (CSV).
Where text detection is still weak
Do not use a text score as proof against a student or candidate.
- Non-native English writers: 28.6% of TOEFL essays were still flagged. This is the biggest open problem, and it is why a high text score should only start a conversation.
- Formal human writing such as academic abstracts: 17.2% flagged.
- AI text written by Claude was missed more often (22.6% on our held-out set) than text from OpenAI or Google models, and edited or paraphrased AI text is caught less often.
- Not measured yet: casual writing, code, and non-English text. Short texts can still produce false positives.
- The detector model we use was trained on data that is not published, so we cannot rule out overlap with public texts.
How we test
Real samples written or taken before 2022
Human texts from Wikipedia (CC BY-SA), arXiv abstracts (CC0) and Project Gutenberg (public domain); real photos from Wikimedia Commons quality images, all from before modern AI tools.
AI samples from many generators
AI images published online with open licences (Wikimedia Commons AI categories, DiffusionDB), a few from public generators, and AI texts from GPT, Claude and Gemini models matched by topic and length.
Scored the way you use ProofAI
The same detection code that runs on proofai.in, with nothing tuned on the test samples. Failed calls are excluded and reported, never counted as right or wrong.
Re-tested every quarter
Each run is dated. We publish the dataset list, method and every result, including the bad ones.
What this test does not cover yet
Every benchmark has gaps. These are ours:
- Hindi, Gujarati, Hinglish and other Indian languages (planned).
- Student essays beyond the Stanford TOEFL and college-essay sets, and other non-native writing samples.
- Mixed documents that are partly human and partly AI.
- Video (video detection is coming soon).
Frequently Asked Questions
On our 2026-10-16 image benchmark (109 images), ProofAI was 86.2% accurate and flagged 0.0% of real photos as AI. For text, 28.6% of non-native TOEFL essays were still wrongly flagged after our 17 Oct fix, and we publish that too.
Human work wrongly labelled as AI. It is the number that matters most when a result could affect a student or job applicant, which is why we publish it separately.
Yes. The image list with sources, licences and results is available as a CSV file below.
Every quarter, and after every major model change. Each run is dated on this page.
Related tools and guides
- Stanford text results (CSV)
- Image results (CSV)
- Why human writing gets flagged
- AI image detector
- Transparency report
- Privacy policy
Last updated: 17 October 2026
Try it on images you know
The best way to judge any detector is to test it on images and texts whose origin you know.
Check an image