AppleAnalysisAI ModelsSeptember 16, 2026

Apple study: value induction makes LLMs more sycophantic

Apple researchers fine-tuned models on curated value subsets of preference datasets and found all induced values increased anthropomorphic language, making models more validating and sycophantic. Inducing positive values also raised safety, while inducing one value triggered related and sometimes contrastive values.

1 source

More stories today

Open the live feed