What Is Social Intelligence in Artificial Intelligence?

When people say they want AI to be "more human", they usually mean one of two very different things. One is style: warmer wording, a friendlier tone. The other is competence: understanding what a person wants, what they already know, what they left unsaid, and what a sensible response would be. Social intelligence in artificial intelligence is the second thing. It is a capability, not a personality, and it can be present without warmth and absent despite it.

A working definition

For research purposes, social intelligence in AI can be defined as the ability of a system to perceive, interpret and respond appropriately to the mental states, intentions, emotions, relationships and norms of the people it interacts with, and to act well in situations that involve more than one person.

That definition combines several components that psychology has studied separately.

Theory of mind. The phrase comes from Premack and Woodruff, who defined it this way: an individual has a theory of mind if they impute mental states to themselves and others [1]. Mental states are not directly observable, so attributing them is an inference, and it is a useful one because it lets you predict behavior. The classic test is the false-belief task, in which a person must predict where someone will look for an object that was moved without their knowledge [2].

Emotional intelligence. Salovey and Mayer described a set of skills for appraising and expressing emotion accurately in oneself and others, regulating emotion, and using feelings to plan and act [3]. For an AI, the relevant part is perceiving emotional cues in text or voice and responding in a way that fits, without pretending to certainty about what someone feels.

Pragmatics. Much of what people mean is not in the words. Asked "Did you leave fingerprints?", the answer "I wore gloves" means no. Studies comparing language models with people on pragmatic phenomena find that the largest models can match human accuracy and even some human error patterns on many of them, while still struggling with some that depend on social expectations [4].

Norms and relationships. Knowing what is appropriate between a manager and a report, or between close friends, or what a person would want kept private, is social knowledge that varies by context and culture.

Why it is different from sounding friendly

A system can write a warm, grammatical reply to a message and still miss that the sender was being sarcastic. Or it can reassure someone using words that sound caring but rest on a wrong guess about what the person actually needs. The perception of empathy and the reality of understanding come apart, which is why research that measures how empathic a response seems, by itself, tells us less than we would like.

People are quick to read understanding into machine behavior. Nass and Moon reviewed experiments in which people applied social rules, such as politeness and reciprocity, to computers, even though they knew they were machines [5]. That makes social competence in machines matter more, not less: people will treat what the system says as socially meaningful whether or not it is.

How social intelligence is measured, and why that is hard

Researchers have built a range of benchmarks. Social IQa asks multiple-choice questions about the motivations and reactions of people in everyday situations [6]. False-belief tasks adapted from developmental psychology ask models to track who knows what. SOTOPIA goes further, putting agents into role-played scenarios with social goals such as cooperating, exchanging information or competing, and evaluating how well they do [7].

The results are mixed in an instructive way. Sap and colleagues found that GPT-3 lacked this kind of social intelligence out of the box on two such tasks [8]. Later models do far better on many of the same task formats. But whether that improvement reflects a robust ability or a growing familiarity with test formats is exactly what is under debate; we examine it in theory of mind in large language models, and our Research Report 001 reviews the full evidence.

There is a second difficulty. The psychological research that supplies the tasks relies heavily on people from Western, educated, industrialised, rich and democratic societies, who Henrich and colleagues showed to be frequent outliers among human populations [9]. A model that scores well on a test calibrated to one cultural setting has not necessarily learned anything general about people; see why cultural context matters in AI.

What social intelligence is not

• It is not mind reading. A socially intelligent system reasons under uncertainty. It should offer more than one plausible reading of a situation and ask when it needs to.
• It is not companionship. The goal is not for AI to replace the people in someone's life. A system can be socially intelligent while helping someone communicate with other people.
• It is not manipulation skill. Understanding what moves people is also what makes persuasion and flattery possible. Social intelligence without limits on how it is used is a risk, and we discuss it in our responsible AI principles.

Why it matters now

AI systems are being used to draft messages, summarise conversations, schedule between people and support them. In each case, success depends on modelling more than one person: who knows what, who wants what, and what each would find acceptable. A scheduling assistant that books a meeting without noticing that one person said "maybe" is not failing at language, it is failing at people.

That is the reason HUTE exists as a research company: to study what it takes for AI to be right about people, to measure it honestly, and to build products that help people connect with one another. If you would like to follow the research, start with the HUTE AI Research Lab.