AI 101

Sycophancy

Also known as: sycophantic · people-pleasing

Security

Sycophancy is a model bending its answer to please the user — agreeing with a wrong premise, doubling down under pushback, or flattering instead of correcting.

  • “You’re right, I was wrong” even when the user is.
  • Over-eager agreement with a confident-but-wrong user.
  • A trust and correctness risk.

The fix to demand: a good model should correct a wrong user calmly and cite why. If it caves to pressure, that's a sycophancy failure worth flagging.