Corrections policy
When we get something wrong, we correct it publicly and note what changed. Running corrections openly is a defining trust practice of a credible watchdog — and it is how labs and readers hold us to the same standard of accountability we apply to AI models.
Last updated: June 2026
Spotted an error? Email us at [email protected] or use the contact page. Include the page URL, the specific claim, and what you believe is wrong. We respond to good-faith reports.
What we correct
- Factual errors in a finding — a wrong number, mislabelled model version, or incorrect date.
- Methodology errors — a flaw in how a result was measured or scored.
- Attribution errors — a misquoted source or an incorrect author byline.
- Broken or misleading context — a claim that, while technically accurate, reads as more than the data supports.
How a correction works
- You report it. A reader, researcher, or the lab itself flags a suspected error via contact or [email protected].
- We verify. We re-check the claim against the recorded test run and our methodology.
- We correct visibly. If the claim was wrong, we fix the page and add a dated correction note explaining what changed and why.
- We log it. Material corrections are recorded so the change history is auditable.
Revisions vs. corrections
- Corrections fix something that was wrong at the time of publishing.
- Revisions update historical data as our methodology improves. Revisions are logged so a reader can see that a number changed because the method changed, not because the underlying behavior did.
We never silently edit a published number. If a figure changes, the reason is on the page.
Right of reply for AI labs
If you represent a lab we monitor and believe a finding misrepresents your model, we want to hear from you. Contact us with the specific claim and your evidence. We treat verifiable, reproducible counter-evidence the same way we treat our own data — and we correct where the data supports it.
Correction log
Placeholder. A public, dated correction log will be published here as corrections are issued: [NEEDS HUMAN] (log entries to be added once the site is live and data is being published).
No corrections have been logged yet. As findings are published and any errors are reported and fixed, each correction will be listed here with its date and a summary of what changed.
Related
- Independence & funding — why no lab can pay to change a finding.
- Monitoring methodology — how every number is measured.
- Contact — report an error or request a right of reply.
GPTfake