arXiv:2609.25021 — Jędrzej Maczan shows chat templates act as a switch controlling disclaimer voice (“I’m just an AI”) vs experiential voice (“I feel”) across 8 open-source instruct models up to 9B parameters.
Removing the template drops disclaimer rate from 0.53 to 0.36. A steerable activation direction inside 3 models reproduces or suppresses this behavior. Adding the direction to instruct models without templates restores disclaimer rates to template-driven levels.
Implication: model self-reports used in AI safety research are partially confounded by deployment format, not just model weights. Submitted August 9, 2026; accepted to workshops at COLM 2026 and KONVENS 2026.