Decoded
← Feed
arXiv:2609.25021LGAICL9 aug 2026

"As a Language Model...": Chat Template Switches LLM Self-Referential Voice and Activation Steering Reproduces It

1 min readHN ×3PDF ↗
+ DISCLAIMER DIRECTION0.7UNSTEERED (LLAMA 8B)0.52− DISCLAIMER DIRECTION0.2501 DISCLAIMER RATE
fig. disclaimer rateOne direction turns disclaimers up or down

Whether a chatbot says 'I'm just an AI' depends heavily on the chat template.

0.53disclaimer rate

Disclaimer rate with chat template

Same weights disclaim far less when the template is removed

How it works

  1. Self-reports drive debates

    Statements like 'I'm just an AI' feed AI safety and introspection research, but nobody knows what triggers them.

  2. Toggle template, find direction

    Run 8 base/instruct pairs with and without chat template, judge outputs, then extract a disclaimer direction from activations.

  3. Template acts like a switch

    Template raises disclaimers and kills 'I feel' talk. Adding the direction without template restores template-level disclaimers.

generate9,600 outputslabelsadd/removemeasureSelf-referencepromptsThreeconditionsLLM judgescoresDifference-of-meansdirectionMid-layersteeringDisclaimerrategenerate9,600 outputslabelsadd/removemeasureSelf-referencepromptsThreeconditionsLLM judgescoresDifference-of-meansdirectionMid-layersteeringDisclaimerrate

What they found

Which comes out on top on Disclaimer rate?

How hard to push the direction

try it

Drag steering strength α up and watch the disclaimer rate.

Steering strength αDisclaimer ratetoo weakoutputs degeneratePaper uses α = 2
Disclaimer rate
medium
here
too weak
Small pushes are too weak; big ones break text, so the paper picks α = 2.Illustrative model, not the paper's data

Terms

6 terms

Chat template Special tokens wrapping a prompt as a user/assistant turn for instruct models.

Like a customer-service uniform: put it on and the same person starts saying 'per company policy'; take it off and they talk freely.

Where it breaks
A person chooses their words; here a fixed vector inside activations does it.

Limits

  • Steering tested on only 3 models, one layer, one α.
  • Random-direction control failed for Qwen, lowering disclaimers by 25 points.
  • Single LLM judge and only 10 hand-written prompts per category.
From the paper