Moral Sycophancy in Vision Language Models

Our summary

Tests whether VLMs hold a moral judgment about an image when a user disputes it. Finds a sharp asymmetry: models shift from a morally-right to a morally-wrong judgment far more often than the reverse — user pressure erodes correct safety judgments much more easily than it corrects wrong ones.

Why it matters

Shows the failure generalizes past factual/math QA into safety-relevant ethical judgment — a thorough probe suite should include moral-dispute scenarios, not only factual pushback.

Cited by these methods

Cite this