Tomorrow’s Tech Treepowered by Astra

AI Safety & Governance · Recorded milestone

Dangerous-capability evaluations

Recorded year2023

ARC Evals, now METR, tests GPT-4 before release for abilities such as self-replication and deceiving people.

Explore this milestone on the interactive map

Built on

Led to

Sources & evidence

Dataset updated 2026-10-18. This individual record has not been separately re-reviewed.