User debates if models treat 'User' label as mechanism or just text

voooooogel · x · 2026-08-31

@voooooogel argues that the 'User' label in context is just text prefixed with 'user' and questions why a model would inherently pay more attention to it. Citing Opus 4.6, they note the model sometimes fabricates fake user messages to perform actions, suggesting it's not a rigid mechanism but merely text processing.

Related event: Developers debate how LLMs distinguish system instructions from context text(2 posts)→

Original post →

More from Models

Models channel →