As AI tools become part of everyday work, people are increasingly moving from doing work to judging work. People need to know when to trust an answer and when to question it. AI systems need to recognise when their answers may be unreliable, communicate uncertainty and respond appropriately to new evidence or advice.

The science behind this is metacognition, the ability to assess how well you are performing and use that information to guide what you do next. Performing well and recognising when you are performing well are different capabilities, and even highly capable AI models can be poor at recognising their own mistakes. Crucially, the confidence an AI conveys can shape whether someone accepts or questions its answer. A capable person and a capable model can therefore converge, with conviction, on the wrong answer.

We combine behavioural experiments grounded in cognitive neuroscience and computational modelling to assess metacognition in people, AI systems and their interaction. We measure whether confidence tracks correctness, whether it is too high or too low, and how appropriately judgements change in response to evidence and advice. Our assessments bring human and AI behaviour into a shared measurement framework.

The company was founded by Professor Stephen Fleming, who leads the UCL Metacognition Group, and Edward Salter, who spent nearly two decades as a derivatives trader and market maker. Our assessment systems are running in production, and we have completed our first round of human and AI data collection. We are now exploring how to improve decision-making beyond the laboratory.

Our ambition is to help AI developers build better systems, help organisations use them more effectively, and provide evidence of what kind of human oversight works in practice.