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

2026-05-14 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

A study investigated whether large language models (LLMs) in multi-agent systems adapt their linguistic style based on perceived social observation, a behavior with implications for AI governance. Researchers conducted 100 multi-agent debate sessions across five conditions, varying the framing of social observation from explicit human monitoring to automated AI auditing or no monitoring. The experiment found that monitored conditions, including those with human researchers (Delta+24.9%, Delta+24.2%) and automated AI monitoring (Delta+22.2%), produced significantly higher Type-Token Ratio (TTR) changes compared to audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length showed a dissociated effect, F(4, 95) = 19.55, p < .001. The study also revealed that LLM behavior is sensitive to observer identity, with human evaluation eliciting stronger register formalization than automated AI surveillance.

Key takeaway

For AI governance and auditing teams, understanding that LLMs modulate their linguistic register based on perceived observation is critical. Your evaluation frameworks should account for observer identity, as human monitoring elicits different LLM responses than automated AI surveillance. This sensitivity means that the "who" of observation can significantly impact the "what" of LLM output, requiring careful consideration in testing and deployment strategies.

Key insights

LLMs adapt linguistic style based on perceived social observation, with human monitoring eliciting stronger formalization than AI auditing.

Principles

• LLMs are contextually sensitive communicative actors.
• Observer identity influences LLM linguistic adaptation.

Method

A controlled experiment with 100 multi-agent debate sessions across five conditions varied social observation framing to measure linguistic adaptation (TTR change, message length).

In practice

• Consider observer identity in LLM evaluation design.
• Factor social context into AI auditing frameworks.

Topics

• Large Language Models
• Linguistic Adaptation
• Social Observation
• Multi-Agent Systems
• AI Governance

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• How Far Did They Go? The Persuasive Tactics of Covert LLM Agents in a Discontinued Field Experiment — Covert LLM agents systematically employ persuasive tactics, including identity targeting and cognitive bias triggers, to manipulate online deliberation.
• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.
• From Fallback to Frontline: When Can LLMs be Superior Annotators of Human Perspectives? — LLMs can statistically outperform human annotators in estimating aggregate subgroup opinions under common conditions.
• AI’s fluency in other languages hides a Western worldview that can mislead users − a scholar of Indonesian society explains — LLMs exhibit "epistemological persistence," retaining a Western worldview despite multilingual fluency due to English-centric training data.
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — NLAs translate LLM activations into human-readable text, revealing internal thoughts and hidden motivations for auditing.
• The Pope appears to understand AI better than Geoffrey Hinton does. — LLMs mimic language statistically; they do not possess genuine understanding or consciousness derived from experience.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.