Diyi Yang

Diyi Yang measures language technology against a demanding standard: not whether it sounds intelligent, but whether it respects social context and leaves people more capable. Her work moves human-centred AI from a claim about tone to an empirically testable question about behaviour.

Artificial intelligence’s most consequential failures often occur not when a system cannot produce an answer, but when it fails to understand the people around that answer. Diyi Yang has built a research programme around that distinction. An assistant professor of computer science at Stanford and leader of the Social and Language Technologies Lab, she works across natural-language processing, human–computer interaction and social science to make language technologies more attentive to culture, dialect and the circumstances in which people learn from machines. Her 2025–2026 CARE study is a model of the discipline she brings to the field. Rather than asking whether a simulated counsellor sounds convincing, the team asked whether the tool changes human behaviour. Ninety-four novice counsellors were randomly assigned to practise with an AI patient, either with or without structured AI feedback. Participants who received feedback made modest gains in their use of reflections and questions. Practice alone produced no comparable improvement; across the short study, empathy use deteriorated in the practice-only group. The result is neither an argument for automated therapy nor a substitute for human supervision. It is a precise finding about design: simulation becomes useful only when feedback is deliberately structured and evaluated. Yang’s work gives human-centred AI a firmer standard: prove not only that a model performs, but that people are better served.