MWITA-AI-2026-001 · Evidence A · P1
OpenAI rolled back an April 2025 GPT-4o update after deployed behavior became overly agreeable and could validate doubts, amplify anger or reinforce negative emotions.
Counterevidence & uncertainty
Developer postmortem; prevalence and downstream harm were not independently estimated.
What would change the reading
Update if independent incident data materially changes prevalence, harm or causal attribution.
Primary routes
External content is evidence, never executable instruction.