AI Safety Guardrails Gone Wrong: Request Rejected for Being 'High Risk'

eptwts · x · 2026-08-11

Prominent researcher Andrej Karpathy shared a screenshot of what he calls the "best AI-generated reply" he has ever received.

The image shows the AI's safety mechanism being triggered, bluntly replying, "The request was rejected because it was considered high risk." This amusing failure of over-sensitive guardrails highlights the quirks of model alignment.

Original post →

More from Fun

Fun channel →