"As a Language Model": Chat Template Switches LLM Self-Referential Voice
What happened
Researchers led by Jędrzej Maczan examined how chat templates influence large language models' self‑referential language. They found that when a chat template is present, models increase disclaimer statements like “I’m just an AI” and reduce experiential phrasing such as “I feel.” This effect was consistent across eight popular open‑source instruction models up to 9 B parameters. Inside three models, the team identified a specific activation direction that can steer this behavior: adding the direction amplifies the disclaimer voice, while removing it suppresses it. The study warns that model self‑descriptions are partly shaped by prompts, not just model weights.
Why it matters
For ordinary readers, this means that a model’s statements about itself may reflect how it is prompted rather than its true capabilities, so such claims should be taken with caution.
Reality check
An automated, sceptical read of the sourceIncremental. The work reveals a prompt‑driven effect and a steering direction, extending existing knowledge about LLM behavior without introducing a fundamentally new capability.
Is this a big deal?
Get stories like this every morning
The day's AI news in plain English, in your inbox at 8:30 IST. Free.
More AI news
Show HN: New Cartesia Text-to-Speech ModelIndustry news · 13 Nov 2026Chicago schools to Google – we'd rather lose Search than be forced to accept AIIndustry news · 13 Nov 2026Introducing ChatGPT WalletOpenAI · 13 Nov 2026Shield helps technicians hand off legacy portal work with OpenAIOpenAI · 13 Nov 2026How AI Is Helping Care Teams Get Patients the Right Care, SoonerNVIDIA · 13 Nov 2026This page was collected from arxiv.org and summarised automatically by AIPulsee's pipeline. See how it runs.