When an LLM decision assistant makes a bad call, what went wrong?

A post sharing a new paper says people use LLMs as decision assistants, including to estimate risk and recommend treatment. It asks whether a bad call means the model formed incorrect beliefs or failed to optimize its action under a given utility—a distinction the post says matters for aligning these assistants with our goals.