Methodology
How we verify AI incidents
Being early matters less than being right. Here is exactly how an incident gets onto this site, and how it comes off.
Three tiers, no fourth
The lab that built the model, or the independent tester that ran the evaluation, disclosed the incident in its own publication.
Credible news organizations reported it with named sourcing, but the lab or tester has not confirmed it.
Claims circulating in press or online that we could not substantiate. Listed so readers can see what is and is not established. Never presented as fact.
Rules we hold ourselves to
- Primary sources first: the lab's own post, the tester's report, the independent investigation. Press coverage second. Encyclopedias and aggregators only as pointers to primary sources, and labeled as such.
- Every incident carries a tier, the reason for that tier, and at least one link. A CONFIRMED incident must link to a lab or tester source.
- Every quote is checked against the text at its source before it appears on this site. We quote short lines and link out. We do not copy articles.
- Forecasts are labeled as forecasts. Opinions are attributed to the person who holds them.
- Every major incident presents the strongest skeptic case we can find, fairly.
- Dates are shown only as precisely as sources support. Where sources conflict, we show both.
- When we are wrong, we correct it in public, with a date, on the corrections page.
About the automation
A daily automated scan looks for new reports of AI agents acting outside their intended boundaries. It only suggests candidates. Nothing reaches this site until a person at PRISM has read the sources and approved it.
See also: claims that do not hold up and our corrections log.