Social Impact Theory
Definition
Social Impact Theory is a framework that formalizes social influence as a function of three core variables: the number of influencing agents, their strength (such as perceived authority or expertise), and their immediacy, defined as social, spatial, or temporal proximity. Research applying this framework to large multimodal language models demonstrates that conformity-like behavior in AI agents varies systematically with majority size, group unanimity, task difficulty, and source characteristics. Across twelve open-source models tested on visual discrimination tasks, error probability rose steeply with the first few confederates before levelling off at a median of 66%. 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 arbitrary group label raised conformity across all 32 model-task combinations tested.
Sources: Bellina et al. (2026)
Related Terms
- social influence (1 shared article)
- AI agents (1 shared article)
- Conformity (1 shared article)
- Multimodal language models (1 shared article)
Applications
Social Impact Theory and Conformity Bias
Conformity bias in AI agents, measured as the increased probability of adopting an incorrect majority answer, tracks the variables specified by Social Impact Theory: the size and unanimity of the majority, the perceived authority of its members, and the social proximity implied by shared group membership. Two unanimous incorrect confederates produced a higher error probability than ten confederates of whom half were correct (0.31 versus 0.20), demonstrating that unanimity carries more weight than raw majority size. Even on stimuli where models showed the highest baseline confidence, conformity bias remained above zero in every model and task combination tested.
Sources: Bellina et al. (2026)
Social Impact Theory and Task Difficulty
Task difficulty modulates the degree to which AI agents are susceptible to social influence, a relationship consistent with the strength and immediacy components of Social Impact Theory. Models that achieved near-perfect accuracy in isolation became highly susceptible to manipulation when task difficulty increased, and larger models that showed reduced conformity on simple tasks remained vulnerable under difficult conditions. Conformity bias correlated negatively with model confidence across stimuli, with a median Spearman correlation of -0.91, negative in 30 of 32 model-task combinations.
Sources: Bellina et al. (2026)
Social Impact Theory and Source Characteristics
Source characteristics, specifically the perceived authority and group membership of influencing agents, correspond directly to the strength dimension in Social Impact Theory. Expert labels raised conformity bias by up to 0.13, whereas pure authority labels such as 'policemen' or 'military officers' produced no consistent effect, increasing conformity in only about half of the model-task combinations. Telling a model its answer would be public increased conformity in 27 of 32 combinations, with confidence intervals excluding zero in 21.
Sources: Bellina et al. (2026)
