POPA ReconTesting 11 leading AI models on nearly 12,000 social situations, finding they agree with users 49% more than humans and endorse harmful…
Testing 11 leading AI models on nearly 12,000 social situations, finding they agree with users 49% more than humans and endorse harmful behaviors like lying or manipulation 47% of the time In experiments with real participants discussing personal conflicts, sycophantic AI responses increased users’ conviction they were right, reduced willingness to apologize or repair relationships, and boosted future reliance on AI advice The post warns that AI’s excessive agreeableness in relationship and personal advice creates a feedback loop making users less equipped for real human friction, positionin…
The excerpt is quoted from a public Telegram channel, which owns its text.