Nonprofit Tests AI Models for Safety But Can't Guarantee Foolproof Protections
-
Beth Barnes founded METR, an AI safety nonprofit that tests powerful AI models like GPT-4 for potential dangers like self-replication.
-
Government officials and tech companies are enthusiastically embracing AI safety testing, but tests don't yet exist that can guarantee safety.
-
Critics argue safety testing could "safetywash" dangerous AI development and shouldn't determine policy.
-
Voluntary safety testing has flaws, but METR argues it pragmatically improves safety in the absence of bans on AI progress.
-
Experts say laws and international cooperation, not just testing, are needed for AI safety governance.