Social Intelligence
Human communication carries far more information than the words themselves.
AI can understand what we say. Understanding people is a different problem
We notice when someone hesitates, becomes more engaged, loses confidence, changes their tone, looks away, or reacts differently than we expected. We often interpret these changes without consciously thinking about them. We take what we see and hear with what we already know about the person,the situation and the interaction, and continuously form a picture of what might be happening and how we should respond. This is social intelligence.
Social intelligence is not a single ability, but a continuous process of perception, inference and adaptation. We observe what happens, compare it with what we expected, update our understanding of the person and the interaction, and change what we do next. Context is essential: the same behaviour can mean entirely different things in different situations , which is why a pause can be hesitation in one interaction and completely ordinary turn-taking in another.
For AI, this creates a challenge that is considerably larger than recognising expressions, analysing voices or detecting individual signals. A system needs to move from observing what someone does to forming hypotheses about what they might want, expect or intend
Bringing social intelligence into machines
Our goal is to bring these same principles into AI, contributing to the development of artificial social intelligence.
Social signal perception represents the first step. A socially intelligent AI system needs to sense the many cues people produce through their voice, facial expressions, words, timing and behaviour, and understand what those cues might signal. This interpretation depends on the context, the person, their personality or cultural background, and what has happened in the interaction so far. This is why different cues like a pause, a change in tone or a facial movement do not have fixed meaning on their own.
Building artificial social intelligence
Once a perceptual layer is built, the next step is to interpret what the various social signals might mean. If someone appears skeptical, uncertain or disengaged, the system needs to reason about possible explanations. This is a capability that psychologists call Theory of Mind and consists in the ability to understand what that person may believe, expect, intend or want, and how those states relate to what is happening in the conversation. These interpretations should remain hypotheses that can change as new information arrives during the interaction.
That understanding then needs to guide what the AI does next. Sometimes the right response may be to explain something differently or ask a question, while at other times it may be better to stop talking, give the other person space, or simply wait. Social intelligence therefore involves not only deciding what to say, but also when to speak, when not to speak, and how to manage the flow of an interaction.
Memory connects all of these abilities over time. During a conversation, the system needs to keep track of what has already happened and how the interaction is evolving. Across conversations, it may also draw on what it has learned about the person before, while updating that understanding as new interactions provide new evidence.
Artificial social intelligence is therefore not a single model or capability. It is the combination of perception, interpretation, memory and adaptive behaviour, working together continuously as an interaction unfolds.
Where we are
We are currently building the foundation.
Our work so far has focused on social signal detection, developing the perceptual layer that allows AI to pick up information beyond the literal content of speech. Around that foundation, we have started exploring how signals can be connected to context and interaction, through early versions of our Conversation Intelligence Framework and Interaction Feedback systems.
The next step is to move from detecting what is happening towards understanding what might be happening next. That means developing systems that can reason about , maintain context across an interaction, model relationships and perspectives, and use that understanding to decide how to respond.
We do not believe this necessarily means recreating every component of human cognition inside a machine. Our goal is more practical: to discover and build out the components required for an AI system to behave in ways that humans themselves recognise as socially intelligent.
Building AI that can navigate people
The first generation of AI learned to process information. The next generation learned to generate it. The next challenge is understanding the people behind it.
That means moving beyond language and towards the richer, continuous stream of signals, context, expectations and relationships that make human interaction work. We are starting with the ability to perceive those signals, and building towards systems that can turn perception into inference, inference into understanding, and understanding into action. That is the path towards artificial social intelligence.
In our own words
