Many systems using AI make remarkable claims of accuracy. Regularly you see claims of 95, 98 or even 99% accuracy. Yet, if this was the case then why are AI systems regularly breaking, why do you frequently need a new model and why are not all the data analysts and AI experts redundant. Also, why does frequently the user experience not live up to expectations. It suggests many times the accuracy bandied around is an illusion or a sale pitch out of context. Yes, sometimes, AI can achieve remarkable feats. Feats we could previously have not imagined and indeed even sometimes feats of remarkable accuracy. However it is not always the case and when it goes wrong it can be disastrous. Here, I will present current understanding of why these systems are not preforming as expected, some thoughts on how to spot AI systems that are less likely to not perform as expected and some of our recent work on developing a system to evaluate the AI system for its performance even if you do not have access to the core algorithms.
Kevin Maynard is a co-director at the Institute of Ethical AI at Oxford Brookes University. The Institute is working with the Centre for Data Ethics and Innovation to develop systems that can evaluate and regulate AI systems. This will contribute to the development of systems that set standards of performance for emerging AI systems particularly those which evaluate humans. Other work is focusing on tools which can constrain the behaviour of AI systems to perform within expected boundaries and be able to learn from multiple databases.