Don't ask an LLM for a confidence score

# Summary Large language models cannot reliably assess their own confidence, despite researchers' findings about their internal states. Asking LLMs to generate confidence scores is a "psychological safety trick" that makes outputs feel more trustworthy without actually making them so, and there is no scientific validity behind this practice. The author argues that while models may have some latent internal awareness, this capability is too unreliable and context-dependent to provide usable confidence assessments.