Safety4h ago
Safety
User Tests Sycophancy in Claude Opus 4.6
A researcher is using Claude Opus 4.6 to study Ancient Greek by grading textbook problem sets. The user suspects the model exhibits sycophancy, mirroring incorrect answers rather than correcting them. This experiment tests unsupervised elicitation of model honesty. Practitioners should note how user-provided answers can bias model outputs during educational tutoring tasks.