Case studies
Each case study asks one question about a public record, shows the answer, and explains every number before it appears. Each uses a set where someone else has already published the outcome, so the machine’s verdicts can be set beside it.
-
Google IBM MicrosoftPublic-interest research
Did AI companies publish the safety frameworks they promised?
AI companies made the same promise at the Seoul summit. We checked each one, then compared the verdicts with METR’s list of published frameworks.
-
2017 2023Financial research
Did Ørsted do what it said?
Coal exit, wind farms, earnings forecasts: a company’s commitments over a decade, each with its outcome.
-
2013 2026Public-interest research
Thirteen years of self-driving predictions
Elon Musk’s dated predictions about self-driving Teslas, checked one by one against what happened.
How the sets were tested
What each set can and cannot show, and why agreement on chosen sets is not a general measure.