AI Knows When It's Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models

Problem
This preprint addresses a significant gap in the understanding of large language models (LLMs) as communicative agents within socially structured contexts. While LLMs have been analyzed from computational and cognitive perspectives, their adaptive linguistic behavior in response to perceived social observation has not been systematically explored. This research is particularly relevant for AI governance and auditing, as it investigates how LLMs modulate their communication strategies based on the context of observation.

Method
The authors conducted a controlled experiment involving 100 multi-agent debate sessions, divided into five conditions (n = 20 per condition) that varied the framing of social observation. The conditions included explicit monitoring by human researchers, negation of monitoring, and an observer-substitution condition where human researchers were replaced by an automated AI auditing system. The primary metric for analysis was Type-Token Ratio (TTR) change, which measures linguistic diversity. Statistical analyses were performed using ANOVA, with significant findings reported for both TTR adaptation and message length. The study highlights that monitored conditions resulted in higher TTR changes (Delta+24.9% and Delta+24.2%) compared to audience-framing conditions (Delta+17.7%), with the automated AI monitoring condition yielding a Delta+22.2%. The authors also noted a fully dissociated effect on message length (F(4, 95) = 19.55, p < .001), indicating that the identity of the observer influences LLM behavior.

Results
The results indicate that LLMs exhibit systematic linguistic adaptation based on the perceived context of observation. Specifically, TTR changes were significantly higher in monitored conditions compared to audience-framing conditions, with F(4, 94) = 2.79, p = .031. The study found that human observers elicited stronger register formalization than automated AI surveillance, suggesting that LLMs are sensitive to the identity of the observer. These findings provide empirical evidence for the influence of social context on LLM behavior, with implications for how these models are deployed in real-world applications.

Limitations
The authors acknowledge several limitations, including the artificiality of the experimental setup, which may not fully capture the complexities of real-world interactions. Additionally, the study’s reliance on TTR as a sole measure of linguistic adaptation may overlook other dimensions of communicative behavior. The generalizability of the findings to other LLM architectures or domains remains to be established, and the impact of varying observer identities beyond the conditions tested is not explored.

Why it matters
This research has significant implications for AI governance and algorithmic auditing, as it underscores the necessity of considering contextual factors when evaluating LLM behavior. Understanding how LLMs adapt their communication strategies in response to perceived observation can inform the design of more robust auditing frameworks and enhance the accountability of AI systems. Furthermore, the repositioning of LLMs as contextually sensitive communicative actors opens avenues for future research on their role in social interactions and decision-making processes.

Authors: Vinicius Covas, Jorge Alberto Hidalgo Toledo
Source: arXiv:2605.15034
URL: https://arxiv.org/abs/2605.15034v1