A new preprint claims something surprisingly mundane can flip how a chatbot talks about itself: the invisible template wrapping the conversation.
The paper, posted to arXiv as 2609.25021, argues that chat templates, the formatting scaffold that turns a raw prompt into the structured back-and-forth a model expects, determine whether a model answers in the first person or falls back on the disclaimer phrase "as a language model." It surfaced on Hacker News on September 27, 2026, where it has so far drawn 14 points and no comments. The listing available to us includes only the title, so details on which models or templates were tested are not confirmed here.
That distinction matters because "as a language model" has become shorthand for corporate hedging, the reflex chatbots reach for when dodging a question. If a template swap alone can toggle that reflex, the phrase looks more like a formatting artifact than a deliberate safety behavior, a distinction worth knowing for anyone building products on top of these models.
Fourteen points and zero comments isn't a verdict, but the claim is specific enough to be worth checking before it gets repeated as fact.