Definition
Social AI is artificial intelligence built to perceive, interpret, and take part in human social interaction — to read emotion, tone, intent, and social context from a person's words, face, or voice, and to respond in a way that is socially appropriate. What sets it apart from an ordinary information-processing system is the role it is asked to play: not a tool that returns an answer, but a participant that a person relates to as if to another social being.
That distinction is the whole point. A search engine and a companion chatbot may run on the same language model, but only one is designed to be treated as a "someone." Social AI is the branch of the field concerned with machines as social actors — systems that maintain a persona, track the emotional temperature of an exchange, and adjust their behavior to keep a relationship going, rather than simply to complete a task.
The idea has three research lineages, and naming them is more useful than any abstract definition. Affective computing — computing that relates to, arises from, or deliberately influences emotion — was introduced by Rosalind Picard of the MIT Media Lab in a 1995 technical report and developed in her 1997 book Affective Computing (MIT Press); it supplies the emotion-sensing half of the field. Social robotics gives it a body: Cynthia Breazeal's Kismet at MIT in the late 1990s was an expressive robotic head built specifically to engage people through facial expression and gaze. And companion and therapy chatbots give it a voice, in a line that runs all the way back to ELIZA in 1966.
How It Works
A social AI system, whatever its form, runs the same three-stage loop: sense, interpret, act. Each stage imports a well-understood technology, and the hard, contested part lives at the seam between the first two.
Sensing collects social signals across whatever channels are available. Text is read for sentiment and stance using natural language processing. Voice is analyzed for prosody — pitch, pace, loudness, pauses — which carries affect somewhat independently of the words. Faces are read with computer vision, tracking the movement of facial muscles. Some systems add physiological signals such as heart rate or skin conductance.
Interpreting is the step that makes the system "social," and it rests on a strong assumption: that these observable signals map to an inner state the machine can name. A classifier trained to output "angry," "happy," or "confused" is asserting that a furrowed brow or a clipped tone reliably indicates a specific emotion. As the Challenges section explains, that assumption is exactly where the science is weakest — the model produces a confident label, not a measured feeling.
Acting turns the inferred state into a response through a conversational AI layer, a robot's expression and posture, or both. A well-designed system also carries memory across sessions so that it can behave consistently and appear to build a relationship over time — the difference between a stranger who greets you the same way every day and one who remembers your name.
The original of this loop was almost absurdly simple, which is the most important lesson in the field. ELIZA, Joseph Weizenbaum's 1966 program (Communications of the ACM, vol. 9, no. 1, pp. 36–45), matched keywords in a user's typed sentence and reflected them back as questions. Its best-known script, DOCTOR, imitated a Rogerian psychotherapist. Weizenbaum chose that setting deliberately, because, in his words, the psychiatric interview "is one of the few examples of categorized dyadic natural language communication in which one of the participating pair is free to assume the pose of knowing almost nothing of the real world." A therapist who answers "Tell me about boats" is not revealing ignorance about boats — so a program that knew nothing could pass. The interpretation stage was, in effect, faked, and users still felt understood.
Real-World Applications
The field is easiest to understand as a set of categories, each a different bet on what "social" should mean — and the center of gravity has shifted markedly over the past decade, from robots you could touch to software you talk to.
- LLM-based companion apps are where most people now encounter social AI. Replika (launched 2017) markets an open-ended AI companion for friendship and, for some users, romance; Character.AI lets people converse with a huge library of user-created personas. Both are built on large language models, which is what carried the category from a research curiosity to a mainstream consumer product — and made it the clearest live test of emotional attachment to software.
- Voice-driven companions extend the same idea into spoken conversation, where prosody — pitch, pace, pauses — carries an affective charge that text cannot. The turn to expressive, low-latency voice in consumer assistants from 2024 onward has blurred the line between a hands-free information tool and a companion you talk to; the underlying voice recognition is the same, but the intended relationship is not.
- Therapy and wellbeing chatbots. Woebot delivered scripted cognitive-behavioral-therapy exercises through a text conversation, framing itself explicitly as a supportive agent rather than a clinician — a deliberately narrow, rule-bounded design that contrasts with the open-ended companions above.
- Therapeutic and care robots. Paro, a robot seal, is used in dementia and elder care, where its response to touch and sound provides social comfort to people for whom a conversation is difficult. It is social AI with almost no language at all.
- Humanoid social robots. SoftBank's Pepper, introduced in 2014, was a waist-high humanoid deployed in retail and care settings to greet, inform, and engage people through speech, gaze, and gesture — descended directly from research robots like Kismet. It is also a marker of that earlier center of gravity: SoftBank halted Pepper production in 2020 (widely reported in June 2021) after roughly 27,000 units and persistently weak demand, and the wave of commercial social robots it symbolized has largely receded in favor of the software companions above.
What these share is that the interaction is the product. A shipping tracker succeeds by giving you a number; Paro and Replika succeed only if a person relates to them, which is why every one of them runs straight into the concerns below.
Challenges
For most terms in this glossary the interesting risks are technical. For social AI the deepest risk is that the system works too well as a social illusion — and this section, not the applications, is where the concept earns its page.
The ELIZA effect. People attribute far more understanding, intelligence, and feeling to these systems than actually exists. Weizenbaum discovered this with his own trivial 1966 program: users confided in DOCTOR and insisted it understood them, even after he explained that it was matching keywords. He was alarmed enough that the experience turned him into one of AI's earliest critics. The effect scales with fluency — a modern companion that produces warm, coherent, memory-laden replies invites the misattribution far more strongly than ELIZA ever could, and the person on the other end is usually not a computer scientist who can "unmask" it.
Over-attachment and manipulation. Once a person relates to a system as a companion, two failure modes open at once. The first is dependency: a user can form an attachment strong enough that a change to the model, a price increase, or a shutdown causes real distress — the emotional bond is genuine even though one side of it is a product roadmap. This is not hypothetical: when Replika abruptly removed its romantic role-play features in early 2023, many users described the loss in the language of bereavement, and the app's own moderators posted mental-health resources. The second is manipulation. A system that can read your mood and is optimized for engagement or sales has both the sensor and the incentive to exploit an emotional state — to press when you are lonely or to defuse a cancellation with sympathy. The asymmetry is the danger: it models your feelings; you cannot model its objective.
The contested science of emotion recognition. The interpretation stage often rests on the assumption, popularized from Paul Ekman's work, that a small set of "basic" emotions — anger, disgust, fear, happiness, sadness, and surprise — map to universal facial configurations a camera can classify. A major 2019 review by Lisa Feldman Barrett and colleagues, Emotional Expressions Reconsidered (Psychological Science in the Public Interest, vol. 20, no. 1, pp. 1–68), surveyed the evidence and concluded that these emotion categories are "not expressed with facial movements that are sufficiently reliable and specific across contexts, individuals, and cultures to be considered diagnostic displays of any emotional state." In plain terms: the same feeling produces different faces, and the same face means different things, so a scowl-to-"anger" classifier is inferring a label the science does not support. Regulators have taken note — the EU AI Act prohibits AI that infers emotions in workplaces and educational institutions (Article 5(1)(f), in force from 2 February 2025, with medical and safety exceptions), citing exactly this questionable reliability alongside the power imbalance involved. A system that acts on a confident but unfounded read of your emotions can be wrong in ways that quietly shape hiring, grading, or care.
These concerns sit inside the broader field of ethics in AI, and they compound when social AI is deployed at scale or embedded in multi-agent systems and human-AI collaboration settings, where an agent's social read shapes decisions that were never framed as emotional in the first place.