Loading field map
How do we make sense of the people around us? Do we infer their hidden mental states, rely on learned scripts and stereotypes, or directly perceive their intentions through embodied interaction? This question has driven social cognition research for decades, and the answers have shifted dramatically. The history of the subfield is organized around a persistent tension: between approaches that treat social understanding as a form of inference—a process of reasoning about unobservable mental states—and those that emphasize direct, embodied interaction as the primary mode of social knowing. This tension has produced a sequence of frameworks, each with distinctive commitments about the nature of social knowledge, the role of automatic versus controlled processing, and the computational architecture of the social mind.
The first major frameworks emerged in the 1960s and 1970s, and they set the terms of the debate. Attribution Theory (1960–1990) treated social understanding as a rational, quasi-scientific process. People, on this view, act like intuitive scientists: they observe behavior, infer its causes (attributing it to dispositions or situations), and use those attributions to predict future actions. The framework drew heavily on the work of Fritz Heider and later Harold Kelley, who proposed that people make causal attributions using principles analogous to statistical analysis—covariation, consistency, and consensus. Attribution Theory was inferentialist through and through: social cognition was a matter of reasoning from observable behavior to unobservable causes.
At nearly the same time, Social Schema Theory (1970–1990) offered a very different picture. Instead of deliberate reasoning, it emphasized pre-existing mental structures—schemas—that organize social knowledge and guide perception, memory, and judgment automatically. A schema for "professor" or "restaurant" contains expectations about typical behaviors, roles, and settings. When we encounter a person, we activate relevant schemas without conscious effort, and these schemas shape what we notice and remember. Where Attribution Theory saw a rational scientist, Schema Theory saw a cognitive miser relying on mental shortcuts. The two frameworks coexisted, but they pulled in opposite directions: one stressed controlled, inferential processes; the other stressed automatic, knowledge-driven ones.
By the 1970s and 1980s, researchers began to synthesize these competing emphases into dual-process models. The Heuristics and Biases Program (1970–Present), pioneered by Daniel Kahneman and Amos Tversky, showed that people often rely on simple mental shortcuts—heuristics—rather than full-blown rational analysis. In social contexts, heuristics like the availability heuristic (judging frequency by ease of recall) or the representativeness heuristic (judging category membership by similarity to a prototype) produce systematic biases. This program did not reject inference altogether; rather, it argued that inference is often shallow and heuristic-driven, not the deliberate causal reasoning of Attribution Theory. The Heuristics and Biases Program remains active today, especially in behavioral economics and judgment and decision-making research, where it continues to document the gap between normative rationality and actual human cognition.
The Elaboration Likelihood Model (ELM, 1980–Present), developed by Richard Petty and John Cacioppo, applied dual-process thinking specifically to persuasion. ELM distinguishes two routes to attitude change: a central route, involving careful, effortful elaboration of arguments, and a peripheral route, relying on superficial cues like source attractiveness or message length. The model predicts that when people are motivated and able to think, they take the central route; otherwise, they default to the peripheral route. ELM absorbed the insights of both Attribution Theory (central route as deliberate inference) and Schema Theory (peripheral route as automatic schema activation), but it narrowed the focus to a specific domain—persuasion—and provided a testable mechanism for when each route dominates. ELM remains influential in social psychology and marketing, where its predictions about message design and audience engagement are still actively tested.
The 1980s and 1990s saw a surge of interest in the computational basis of social cognition. Theory of Mind (ToM, 1980–Present) emerged from developmental psychology and primatology, driven by the question of when and how children come to understand that others have beliefs, desires, and intentions that may differ from their own. The framework posits a dedicated cognitive module—or a set of modules—that computes representations of others' mental states. ToM is deeply inferentialist: social understanding is achieved by forming and reasoning about propositional attitudes (e.g., "Sally believes the marble is in the basket"). The framework's signature method is the false-belief task, which tests whether a child can attribute a mistaken belief to another agent. ToM remains a major research program in developmental psychology, comparative cognition, and autism research, where deficits in mental-state attribution are a central explanatory construct.
At the same time, Connectionist Models (1980–Present) offered a radically different computational account. Instead of modular, rule-based representations, connectionist networks learn social categories and behaviors through distributed patterns of activation across many simple units, adjusted by experience. A connectionist model of stereotyping, for example, might learn to associate certain traits with social groups not through explicit rules but through statistical regularities in the input. This framework challenged ToM's modularity: social knowledge, on the connectionist view, is not stored in a dedicated module but emerges from domain-general learning mechanisms. Connectionist Models also challenged the inferentialist assumption that social cognition requires explicit representations of mental states; instead, social behavior can be driven by learned associations that never rise to the level of propositional belief. The rivalry between ToM and Connectionist Models was a central debate of the 1990s, with ToM emphasizing innate, specialized architecture and Connectionist Models emphasizing learned, distributed representations. Both frameworks remain active today, though they have increasingly informed each other: some researchers now ask how connectionist learning might give rise to ToM-like abilities.
By the 1990s, a more radical challenge emerged. Embodied and Enactive Approaches (1990–Present) rejected the very idea that social cognition is primarily a matter of internal representation or inference. Drawing on phenomenology (especially Merleau-Ponty) and ecological psychology (Gibson), these approaches argue that social understanding is grounded in the body's interactions with the environment. We do not infer another person's emotions from their facial expression; we directly perceive their anger in the tensing of their jaw and the narrowing of their eyes. Social cognition, on this view, is not a spectator sport but a participatory activity: we understand others by engaging with them, not by forming models of their minds. Embodied approaches differ sharply from both ToM and Connectionist Models. Where ToM posits a mental representation of another's belief, embodied approaches say that understanding is enacted in joint action, not represented in the head. Where Connectionist Models still treat social knowledge as stored patterns in a network, embodied approaches locate social cognition in the dynamic coupling of agent and environment. This framework has been especially influential in robotics (socially situated robots) and in studies of infant interaction, where pre-verbal infants coordinate with caregivers without apparent mental-state attribution.
The most recent major framework, Predictive Processing (2000–Present), attempts to reconcile the inferentialist and interactionist strands. Predictive Processing (also called predictive coding or the Bayesian brain hypothesis) proposes that the brain is fundamentally a prediction engine: it generates top-down predictions about sensory input and updates those predictions based on prediction error. Applied to social cognition, this framework suggests that we understand others by predicting their behavior using a hierarchical generative model. When predictions fail, we update the model—a process that can be described as Bayesian inference. Crucially, Predictive Processing incorporates both inference and interaction. The inferential side is clear: the brain infers the causes of sensory input (including other people's actions) by minimizing prediction error. But the framework also accommodates embodied interaction through active inference: we can act on the world to make it match our predictions, rather than passively updating our models. In social contexts, this means that we sometimes shape others' behavior to fit our expectations, rather than revising our expectations to fit their behavior.
Predictive Processing differs from earlier frameworks in several ways. Compared to Heuristics and Biases, it offers a unified computational principle (prediction error minimization) rather than a catalog of shortcuts. Compared to ToM, it is domain-general: the same predictive machinery applies to social and non-social domains, challenging ToM's modularity. Compared to Embodied Approaches, it retains a central role for internal models (the generative model) while also explaining how action and interaction serve prediction. Predictive Processing has rapidly become one of the most active frameworks in cognitive science, with applications to social cognition, autism, schizophrenia, and social interaction. It is not yet a settled paradigm, but it has absorbed elements of many earlier frameworks: the Bayesian inference of Attribution Theory, the automaticity of Schema Theory, the dual-process distinctions of ELM, and the embodied emphasis of Enactive Approaches.
Today, social cognition research is a pluralistic field. No single framework has achieved dominance, but several remain highly active. Theory of Mind continues to drive research on developmental trajectories, cross-cultural variation, and neural correlates (especially the temporoparietal junction and medial prefrontal cortex). Connectionist Models have evolved into deep learning approaches that can simulate complex social behaviors, though they are less central to mainstream social cognition than in the 1990s. Heuristics and Biases remains a major force in behavioral economics and applied judgment research. ELM is still widely used in persuasion and communication studies. Embodied and Enactive Approaches have gained traction in philosophy of mind, developmental psychology, and human-computer interaction, though they remain a minority voice in experimental social psychology.
The leading frameworks today—Theory of Mind, Predictive Processing, and Embodied Approaches—agree on at least one point: social cognition is not a single, monolithic process. They disagree, however, on what the fundamental architecture looks like. ToM advocates argue for specialized, possibly innate modules for mental-state reasoning. Predictive Processing advocates argue for a domain-general prediction engine that learns social regularities through experience. Embodied advocates argue that the whole framework of internal models and representations is misguided; social understanding is enacted, not computed. This disagreement is not merely theoretical; it shapes experimental design, clinical interventions (e.g., for autism), and the design of social robots. The field's history suggests that the tension between inference and interaction will continue to drive new frameworks, and that the most productive future work may lie in specifying the conditions under which each mode of social understanding operates.
From Attribution Theory's intuitive scientist to Predictive Processing's hierarchical prediction engine, social cognition has moved through a series of frameworks that have progressively complicated the picture of how we understand others. The early frameworks set up a debate between deliberate inference and automatic schema activation. Dual-process models synthesized these into a single framework. The representationalist peak of ToM and Connectionist Models debated the architecture of social knowledge. Embodied Approaches challenged the very need for internal representations. Predictive Processing now offers a synthesis that incorporates both inference and interaction. The field remains open, with active frameworks coexisting and competing, and the central question—how do we understand others?—continues to generate new answers.