Much of conversational AI is organised around a question and an answer. Our focus is what happens beyond the answer: helping people think, learn and practise. That means deciding when to explain, when to let someone work through a problem, and how to tell whether they can use what they have learned on their own.
The challenge extends across many interactions: understanding what someone needs, choosing support that helps them progress and checking whether that progress lasts. These are four of the research problems we are working on.
Establishing trust in AI-generated learning material
A question, its answer and its explanation can agree with one another and still be wrong. A second model can repeat the same error. The challenge is to establish evidence of correctness beyond that agreement, while also checking ambiguity, conceptual depth and what the material actually tests. We are working on how automated verification and expert judgement can combine to make those decisions traceable and dependable at scale.
Inferring understanding from incomplete evidence
Understanding is not directly observable. The same answer can result from reasoning, recall or chance; an incorrect answer can conceal a sound method. Difficulty, prior exposure and forgetting further complicate the evidence. We are working on how to estimate what a learner understands, represent uncertainty and revise that estimate as new evidence arrives—without turning a small number of attempts into a confident judgement.
Adapting support without reinforcing mistaken assumptions
A learning agent changes the evidence it later observes: what it chooses to teach or ask determines what it can learn about the student. An early misjudgement can narrow future practice and appear to confirm itself. We are working on how agents can balance acting on what they know with testing what they might have misunderstood, while maintaining useful memory and adapting to changing goals over time.
Measuring learning beyond immediate performance
A better score today may reflect a familiar question, an easier task or help that will not be available tomorrow. Establishing whether AI support leads to lasting understanding requires separating those effects from learning that transfers to unfamiliar problems. We are working on evaluation methods that connect an agent’s decisions to later evidence, using controlled simulations to test assumptions and investigating how those findings can be tested with real learners.