ProofAI transparency report
A detector can be wrong, and you should be able to see how often, what we changed and what went wrong. This page is updated every month with our measured error rates, disputes received, model changes and incidents.
How to dispute a result
Open your result
Run the check on proofai.in as usual. Under every result there is a "Think this is wrong?" button.
Tell us what is wrong
Add an optional note. The report stores the verdict, score, type of check and a length range, plus your note. It does not store the content you checked.
We review it
Reports are counted on this page every month and used to find weak spots in the detector. A report does not change a past result, so keep your drafts and version history as your own evidence.
Need to talk to a person?
For a case that affects a student or applicant, contact us and we will look at it directly.
Current measured error rates
Taken from our own dated benchmarks. Full method, confidence intervals, dataset and weak spots are on the accuracy page. False positive means human work wrongly flagged as AI; false negative means AI work missed.
| What we measured | False positives (human flagged as AI) | False negatives (AI missed) | Tested on |
|---|---|---|---|
| AI image detection (109 images) | 0.0% (0/52) | 26.3% (15/57) | 2026-10-16 |
| AI text, our held-out set | 3.7% (5/135) | 10.0% | 2026-10-17, after the engine fix |
| AI text, Stanford non-native (TOEFL) essays | 28.6% (26/91) | not applicable | 2026-10-17 |
| AI text, Stanford native college essays | 2.9% (2/70) | not applicable | 2026-10-17 |
| AI text, formal human abstracts | 17.2% (25/145) | not applicable | 2026-10-17 |
| AI text, Stanford AI essays | not applicable | 7.9% (35/443) | 2026-10-17 |
These are benchmark rates on our test sets, not guarantees for your documents. Non-native English writing and formal human writing are still flagged too often. We missed three of our five text pass marks, and say so on the accuracy page.
Disputes received
Counts of "Think this is wrong?" reports. The feedback button is new, so we do not have a full month of data yet. We will not publish partial or estimated numbers.
| Month | Reports received | Status |
|---|---|---|
| October 2026 | Not yet published | Report #1 due 2026-11-16 (one month of feedback data) |
Each monthly report will show how many reports we received, split by text and image checks, and what we changed because of them. Reports never include the content you checked.
Model changes
| Date | Change | Why |
|---|---|---|
| 2026-10-17 | Text detector: primary model replaced and score recalibrated. Scores of 35 to 65 are now shown as unclear. | The previous primary model sometimes timed out and the timeout was scored as 0, which made human writing look like AI. See the incident below. |
| 2026-10-16 | First published image benchmark (109 images, 12+ generators). | To measure the image detector in the open. No image model change was made. |
Incidents
| Date | What happened | What we did |
|---|---|---|
| 2026-10-16 | Our benchmark found that the 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 the next day, calibrated the score on separate data and re-tested on our held-out set and the public Stanford set. Human texts wrongly flagged fell from 57.8% to 3.7% on our held-out set. Non-native essays are still flagged at 28.6%, and that is not fixed. |
Text results before the 2026-10-17 fix could have been wrong in this direction (we found the problem on 2026-10-16). If a text result affected you, please contact us.
Our commitment
We update this page every month, on or around the 16th, with the same four sections. If a month has no disputes or incidents, we say so. Benchmarks are re-run every quarter and after every major model change, and each run is dated.
- Only measured numbers appear here. If we have not measured something, we say "not yet measured".
- Bad results stay on the page. We do not remove them when they are fixed; they move into the dated log.
- A score is a probability, not proof. Do not use it alone against a student or candidate.
Frequently Asked Questions
A monthly page with our current false-positive and false-negative rates, the disputes we received, dated model changes and incidents.
Report #1 is due 2026-11-16, after one month of "Think this is wrong?" feedback. Until then we do not publish dispute numbers.
No. A report is used to improve the detector and is counted in the monthly report. Keep drafts and version history as your own evidence.
No. A dispute report holds the verdict, score, type of check, a length range and your optional note, not the checked content.
Related tools and guides
Last updated: 17 October 2026
Think a result is wrong?
Use the "Think this is wrong?" button under any result, or contact us.
Read the full benchmark