One Chat Template Line Decides If Llama Calls Itself an AI

Ask a chatbot whether it has feelings, and it often answers with some version of "I'm just a language model, I don't have experiences." A paper posted to arXiv on Aug. 9, 2026 and accepted to a COLM 2026 workshop argues that answer is less a fact about the model than a setting in its wrapper.

Why does a chatbot say it is "just a language model"?

Independent researcher Jędrzej Maczan tested the question directly by comparing model output with and without its chat template, the wrapper of role tags and formatting instructions that a serving stack applies to a conversation before the model ever sees it; the same underlying weights produce different text depending on whether that wrapper is present. Maczan ran the test on eight open-weight instruction-tuned models of up to 9 billion parameters, spanning Meta's Llama, Google's Gemma, Alibaba's Qwen and Mistral's namesake model, asking each one questions about its own nature.

What changed when the template came off

With the chat template applied, 53 percent of the models' self-referential answers used disclaimer language such as "I'm just an AI," and only 1 percent used experiential language such as "I feel," according to the paper. Strip the template from the same instruct models and the numbers move: disclaimers fell to 36 percent of answers, and experiential language rose to 15 percent. A version of the models with no instruction tuning at all, tested as a baseline, disclaimed in just 12 percent of answers.

A worked example: steering three models with one direction

Maczan then searched inside three of the eight models, Qwen 2.5 7B, Llama 3.1 8B and Gemma 2 9B, for a direction in their internal activation space, the numerical state a model computes while generating a response, that tracks disclaimer language. Adding that direction during generation, without changing any model weights, raised disclaimer rates by an average of 21 percentage points across the three models. Subtracting it lowered rates by an average of 15.6 percentage points, the paper says.

Model	No template, direction added	With template, direction subtracted
Qwen 2.5 7B	28% to 50%	40% to 30%
Llama 3.1 8B	33% to 53%	52% to 25%
Gemma 2 9B	52% to 75%	75% to 65%

Maczan also ran a control, adding a random direction of the same size, drawn from a Gaussian distribution, instead of the disclaimer direction. That control had no meaningful effect on Gemma or Llama, but it lowered Qwen's disclaimer rate by 25 percentage points anyway, a result the paper flags without fully explaining.

The clearest test came from applying the disclaimer direction to an instruct model with no chat template. According to the paper, doing so pushed the model's disclaimer rate up to match a version of the same model running with its template intact, meaning the activation nudge alone reproduced what the formatting normally does.

What this means for AI self-reports

The paper's own framing is narrow: what a model says about itself is set partly by deployment choices such as the chat template, not solely by its weights, so a model's self-description should not be read literally as a fact about its training or architecture. A summary of the work by the Chinese AI outlet HyperAI put the same point more bluntly, describing the chat template as functioning like a binary switch for how a model talks about itself, on or off, rather than a dial the model sets on its own.

That distinction matters directly for a live debate this site has covered before over whether AI models' statements about their own welfare deserve weight, and for calls from safety researchers for mandatory intervention mechanisms built partly on how models describe their own internal state. If a line of formatting most users never see can move a model's disclaimer rate by double digits, a transcript alone cannot settle what a model "really" believes about itself, and any argument that leans on a model's own words needs to say which version of the model, template on or off, produced them.

Maczan's paper has cleared two workshop reviews so far, at COLM 2026's Workshop on Efficient Reasoning and at KONVENS 2026's Eval4SD workshop, but has not yet been evaluated by a full conference's main track as of Sept. 27, 2026.