Safety researchers clash over 'generative AI' label: outputs vs capabilities framing

evijit · x · 2026-09-20

A short but substantive exchange on AI safety terminology: NatPurser avoids 'generative AI' because safety skeptics dominate its usage and because it frames AI by outputs (text/images) rather than capabilities (agents, cyber attacks, bio risks) — the latter being what safety people worry about. evijit pushes back that autoregressive models are literally token generators, while noting safety should apply to all model types.

Original post →

More from Safety

Safety channel →