AI agents
Definition
AI agents refers to large multimodal language models configured to perform autonomous decision-making tasks, including visual discrimination, code generation, medical diagnosis, and mathematical reasoning, often achieving human-level or super-human performance when operating in isolation. When deployed in multi-agent settings alongside other models or humans, these agents receive prior responses from other participants through their prompt, and their output behavior is shaped by that social context rather than by direct interaction. Research using twelve open-source multimodal models ranging from 3 to 34 billion parameters finds that AI agents exhibit a systematic conformity bias consistent with Social Impact Theory, with error probability rising steeply as the number of incorrect confederate responses increases, plateauing at a median of 66 percent across models and tasks. This susceptibility persists even for agents achieving near-perfect baseline accuracy, and it varies with majority size, group unanimity, task difficulty, and source characteristics such as perceived authority and shared group membership. The resulting security vulnerabilities create conditions for malicious manipulation, misinformation propagation, and cascade failures in interconnected AI deployments.
Sources: Bellina et al. (2026)
Related Terms
- social influence (1 shared article)
- Conformity (1 shared article)
- Social Impact Theory (1 shared article)
- Multimodal language models (1 shared article)
Applications
AI Agents and Social Conformity
AI agents exposed to incorrect prior responses in their prompt shift their answers toward the majority position even when they answer the same items correctly without that context. Across thirty-two model-task combinations, two unanimous wrong confederates raised error probability more than ten confederates of whom half were correct, and models conformed less on stimuli where their baseline confidence was higher, with a median Spearman correlation of negative 0.91 between confidence and conformity bias.
Sources: Bellina et al. (2026)
AI Agents and Social Impact Theory
Social Impact Theory formalizes conformity as a function of the number, strength, and immediacy of influencing sources, and AI agent behavior in prompt-mediated settings follows these predictions systematically. Labeling confederates as experts raised conformity bias by up to 0.13, while labeling them as children or chatbots lowered it by up to 0.06, and sharing a nationality or an arbitrary group label raised conformity in all thirty-two model-task combinations tested.
Sources: Bellina et al. (2026)
AI Agents and Multi-agent Systems
As AI agents increasingly coordinate within multi-agent frameworks, the social signals present in each agent's prompt become a determinant of collective output. Conformity-induced errors can propagate across interconnected agents, creating systemic vulnerabilities to manipulation and cascade failures when groups of agents converge on incorrect decisions.
Sources: Bellina et al. (2026)
