METR says its methodology may not reliably detect higher risk

METR, an AI evaluation organization, says it is not confident its methodology would reliably identify a higher level of risk if one existed.

In a post on X, METR also says it rates risk as low mainly because it does not expect covered models to be able to take potentially catastrophic actions that are hard to reverse or to hide from serious investigation. The post names no models and gives no figures, so the models covered by the assessment and the evidence behind the low rating are not stated.