Vinicius Covas; Jorge Alberto Hidalgo Toledo
An experiment with 100 multi-agent debate sessions examines linguistic adaptation under different social observation framings.
Manuscript available as a preprint. This record does not imply journal publication or peer review.
Controlled experiment: 100 multi-agent debate sessions, five conditions and 20 sessions per condition. The framing of social observation is varied.
Linguistic register changes across observation conditions. Lexical diversity and message length respond differently to context.
Observed adaptation does not demonstrate consciousness, intention or human understanding. Findings concern the design and systems examined, not all AI.
Unmodified original PDF. Source and authorship are listed in the citation. License: CC BY 4.0.
Study data and code: 100 debate sessions among AI agents. The deposit documents one incomplete session and exclusion criteria.
Vinicius Covas; Jorge Alberto Hidalgo Toledo ·
Covas, V., & Hidalgo Toledo, J. A. (2026). AI Knows When It’s Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.15034
Context, audiences and agent behavior.
Context, audiences and agent behavior.
Context, audiences and agent behavior.