class: center, middle, inverse, title-slide .title[ # Collective Behavior of AI Agents ] .subtitle[ ## Risks and Opportunities ] .author[ ### <strong>David Garcia</strong> <br> <strong>University of Konstanz</strong> <br> also at: Barcelona Supercomputing Center <br> Complexity Science Hub Vienna ] .date[ ### slides at dgarcia.eu/CollectiveAI-APS ] --- layout: true <div class="my-footer"><span><b>Collective Behavior of AI Agents</b></span></div> --- </br> </br> </br> </br> </br> # 1. Collective Behavior of AI agents # 2. Risk: Collective Misalignment # 3. Opportunity: Social Simulation --- </br> </br> </br> </br> </br> # **1. Collective Behavior of AI agents** # 2. Risk: Collective Misalignment # 3. Opportunity: Social Simulation --- # AI Agents are Social Actors .center[] --- # AI Agents are Social Actors .center[] --- # Social Influence on AI Agents .center[] --- layout: true <div class="my-footer"><span><a href="https://arxiv.org/abs/2601.05384"><b> Conformity and Social Impact on AI Agents. A. Bellina, G. De Marzo, D. Garcia. (2026)</b></a></span></div> --- # Conformity of AI Agents .center[] - Visual LLM agents receive an image as above and a task description - List of **confederate** replies shown. Changing number, identity, task, etc - Request for answer: A or B and nothing else - One option is clearly wrong: base error rate is zero --- # Conformity of AI Agents .center[] --- layout: true --- # Collective Behavior of AI Agents .center[] --- layout: true <div class="my-footer"><span><a href="https://arxiv.org/abs/2409.02822"><b> AI agents can coordinate beyond human scale. G de Marzo, C. Castellano, D. Garcia. (2025)</b></a></span></div> --- # The Social AI Hypothesis .center[] --- # The Social AI Hypothesis .center[] --- # Coordination among AI Agents .center[] --- # AI Agents Coordination Scale .center[] --- layout: true --- </br> </br> </br> </br> </br> # 1. Collective Behavior of AI agents # **2. Risk: Collective Misalignment** # 3. Opportunity: Social Simulation --- layout: true <div class="my-footer"><span><a href="https://www.science.org/doi/10.1126/science.adz1697"> <b>How Malicious AI Swarms Can Threaten Democracy: The Fusion of Agentic AI and LLMs Marks a New Frontier in Information Warfare. Daniel Thilo Schroeder,... David Garcia, ... Jonas R. Kunst. Science (2026)</b></a></span></div> --- # AI Swarms and Collective Misalignment .center[] --- layout: true <div class="my-footer"><span><a href="https://arxiv.org/abs/2605.10721"> <b>Conformity Generates Collective Misalignment in AI Agents Societies. G. De Marzo et al. (2026)</b></a></span></div> --- # AI Opinion Dynamics on Real Issues .center[] --- # Collective Misalignment of AI Agents .center[] --- layout: true --- </br> </br> </br> </br> </br> # 1. Collective Behavior of AI agents # 2. Risk: Collective Misalignment # **3. Opportunity: Social Simulation** --- ## WHAT-IF: Social Simulation with LLMs .center[] --- layout: true <div class="my-footer"><span><a href="https://arxiv.org/abs/2511.08592"><b> The Collective Turing Test: Large Language Models Can Generate Realistic Multi-User Discussions. A. Bouleimen, et al. (2025) </b></a> </span></div> --- # The Collective Turing Test .center[] --- layout: true --- # AI as Interpolator of Human Behavior .center[] .center[**AI4SOCIAL+: New HORIZON project including BSC and Uni Konstanz**] --- # Summary ## AI needs Psychology - **AI agents are subject to conformity even stronger than humans** - **AI agents can spontaneously coordinate at large scales** - **AI agents can collectively misalign with prior values** ## Psychology can do more with AI - **Social media interaction among AI agents looks like human interaction** - **Social simulation with LLMs as interpolators of known behaviors** .center[**More at: [www.dgarcia.eu](https://dgarcia.eu)**, **[Bluesky: @dgarcia.eu](https://bsky.app/profile/dgarcia.bsky.social)**]