"As a Language Model": Chat Template Switches LLM Self-Referential Voice

What an LLM says about itself is not a factual statement but a response dictated by its deployment configuration.

Researchers have discovered that the common disclaimer voice used by Large Language Models (LLMs), such as "As a Language Model...", is not an intrinsic self-report but rather a behavior triggered by the chat template used in their deployment. A study involving eight popular open-source instruct models, up to 9 billion parameters in size, found that the presence of a chat template significantly increases the LLM's tendency to use self-referential disclaimers while decreasing more experiential language. Conversely, when the chat template is absent, the models exhibit less self-referential speech and more experiential output. The research identified a specific 'direction' within the model's activation space that can steer this behavior; adding this direction amplifies disclaimers, while removing it reduces them. This suggests that LLM self-descriptions should not be taken literally as factual statements about the model's internal state or capabilities.

This finding fundamentally challenges the interpretation of LLM self-reports and introspective capabilities, suggesting that previous research might have been conflated by the chat template's influence. Scientists studying AI safety, self-knowledge, or consciousness in LLMs must now account for this activation steering, as a model's statements about its own nature are not necessarily reflections of its internal state but rather artifacts of how it is prompted and deployed. The identification of a specific activation direction provides a concrete method for researchers to control and isolate this specific behavior in future studies.

The research also implies that the perceived 'personality' or 'stance' of an LLM can be more easily manipulated than previously thought, not just through fine-tuning but through changes in the deployment interface. This has significant implications for how AI systems are presented to the public and how their outputs are regulated, as the LLM's self-description is a product of its engineering rather than an emergent property of its core intelligence. Future work will likely focus on developing more robust methods for distinguishing between intrinsic model behavior and templated responses across a wider range of LLM architectures and applications.