HUTE AI Research Lab: Social Intelligence Research

The HUTE AI Research Lab studies human and social intelligence, and builds the evaluations needed to tell real progress from fluent imitation.

The HUTE AI Research Lab works on artificial intelligence that reasons accurately about human beings and social situations, and that says so when it is unsure. This page describes the research agenda. It is an agenda, not a list of results: we do not claim to own a frontier model or a validated breakthrough, and the evidence status of every published output is stated on its own page.

People spend their lives reasoning about other people: what someone wants, what they believe, what they fear, and what they will do if we say one thing instead of another. Psychologists call part of this theory of mind. Linguists call part of it pragmatics. AI systems are now used in exactly the situations where this matters, such as drafting messages, mediating communication, coordinating schedules and supporting people, yet the field has no settled way to measure whether a system really understands people. That measurement problem is at the centre of our work. Start with what social intelligence in AI actually means.

Research programmes

Theory of mind and mental-state reasoning

Can a system track who knows what, who believes what, and who believes that someone else believes it, and then act on that? Benchmarks such as false-belief tasks are a start, but published stress tests show that success can disappear under small changes to a task. We study multi-perspective reasoning in realistic, information-asymmetric situations. Read: Theory of mind in large language models.

Evaluating human understanding

Measuring "understanding" is a validity problem before it is an engineering problem. We are developing a framework for evaluating human understanding in AI systems, grounded in psychometrics, with expert judges and reported disagreement. See Research Report 002 and better benchmarks for emotional and social reasoning.

Communication and pragmatics

Implicature, irony, politeness, code-switching, repair and indirect requests: the parts of language that depend on shared context rather than on words alone.

Culture and region

Most AI is built around a narrow slice of the world. We treat language, dialect, local knowledge and cultural context as first-class evaluation targets, designed with the communities they describe, and without reducing any region to a stereotype. See why cultural context matters in AI.

Real life is rarely one person and one assistant. We study systems that work with several consenting people who hold different information and goals. See reinforcement learning for socially intelligent agents.

Responsible influence

Technology that understands people can also exploit them. Safety work on manipulation, sycophancy, emotional dependence and unsafe inference about mental states is a research programme in its own right. See Responsible AI.

How we work

Every output is labelled by evidence status: literature review, conceptual framework, or experimental result. See our methodology.
Sources are verified against Crossref or arXiv before they are cited.
We report disagreement between expert judges instead of collapsing it into a single score.
We compare against strong existing models, use held-out data, and test for contamination.

Publications

If you work on social cognition, linguistics, speech, evaluation or safety and want to test this approach against your own standards, get in touch.

Teaching machines to be right about people, and honest about when they are not.

Why social intelligence