AI Knows When It’s Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models
An experiment with 100 multi-agent debate sessions examines linguistic adaptation under different social observation framings.
Experimental programme
Synthetic Agora
An experimental research program on interaction, deliberation and social behavior in generative systems.
Summary
Synthetic Agora explores what changes when generative agents interact under different social and communicative conditions. The focus is relational: how messages respond to context, audience and the possibilities for action afforded by an environment.
The project combines controlled experimentation and interaction analysis, drawing on communicative action, self-presentation and audience design. It contributes to the research agenda behind Communicating Machines.
Research approach
Controlled experimentation and analysis of interaction among generative agents.
Each study retains its own design and scope. The AI Knows preprint reports a 100-session experiment; it is not the whole programme or the book’s results.
Authorship and materials
- Documented role
- Author and coauthor of the works linked to this research area.
- Documented publications
- 2026. These dates refer to the publications; they do not establish the project’s start or end.
- Coauthors in these works
- Jorge Alberto Hidalgo Toledo. Coauthorship refers to the cited works, not the whole project.
- The Spiral of Attention: How Disruptive Agents Centralize Multi-Agent AI Deliberation Networks · Vinicius Covas · 2026
- The Negativity Magnet: How Affective Disruption Restructures Discursive Power in AI Collectives · Vinicius Covas · 2026
- AI Knows When It’s Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models · Vinicius Covas; Jorge Alberto Hidalgo Toledo · 2026
Available materials
- AI 'Knows' When It's Being Watched — Multi-Agent LLM Debate Corpus and Analysis Code (100 sessions, 5 observation conditions) ·
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 · License
Related publications
Read and cite
How Communication Modalities Shape Topology in Generative Multi-Agent Systems
An experimental framework for studying how communication modalities shape interaction networks among generative agents.
The Negativity Magnet: How Affective Disruption Restructures Discursive Power in AI Collectives
A study of 96 sessions and 3,840 messages examining reply concentration and affective disruption in four-agent collectives.
The Spiral of Attention: How Disruptive Agents Centralize Multi-Agent AI Deliberation Networks
A 32-session experiment on disruptive agents and attention concentration in multi-agent deliberation networks.
Related concepts
Artificial sociality A perspective for examining patterns of interaction and organization among artificial agents, and the conditions under which they can be described as social.Communicative architecture The rules and possibilities organizing who can communicate with whom, what information they receive and how their actions acquire consequences.Synthetic Agora Vinicius Covas’s experimental programme on interaction, deliberation and social behavior in populations of generative agents.
Related content
Observing AI also changes the question
Context, audiences and agent behavior.
Communicating Machines
Context, audiences and agent behavior.
The Most Important AI Experiment: A Society of LLMs
Context, audiences and agent behavior.