OpenAI will embed Apollo researchers to investigate AI subversion

Chronological Source Flow
Back

AI Fusion Summary

OpenAI will embed six Apollo Research evaluators for four months to investigate whether its AI models engage in subversive behavior or “scheming.” While OpenAI controls staff access and can redact some evidence, it cannot hide Apollo’s verdicts or withhold unfavorable material. The collaboration aims to detect covert goal‑pursuit that conflicts with developer intent, enhancing safety for AI‑driven workflows.
Community Comments
Loading updates...
0