Back to AIBriefs
AnalysisPolicy

Marcus: LLMs 'not built for safety', guardrails insufficient

Gary Marcus argues that next-token predictors cannot be made safe due to inherent architecture flaws. He and Katie Conrad warned since January 2024 that guardrails thread a needle between too restrictive and too permissive. Marcus suggests curtailing LLMs until a better technology is found.

Jun 17, 6:05 PM
Marcus: LLMs 'not built for safety', guardrails insufficient — AIBriefs