Exeter College Fellow highlights need for better tools to evaluate AI safety
Exeter College Official Fellow in Engineering Science and Professor of Machine Learning, Maike Osborne, has highlighted the urgent need for better ways to measure and evaluate the capabilities and safety of advanced artificial intelligence (AI).
In an expert comment published by the University of Oxford, Professor Osborne and Dr Fazl Barez, Senior Researcher and Technical Director of the Oxford Martin AI Governance Initiative, argue that current methods for assessing advanced AI systems are not yet sufficient to determine reliably whether their capabilities could be slipping beyond human control.
Their analysis follows a series of recent developments that have intensified debate around AI safety, including incidents in which AI agents – systems capable of using software tools and taking actions – have circumvented controls during evaluations.
Professor Osborne and Dr Barez explain that measuring the capabilities of these systems remains difficult. AI evaluations can be expensive, sparse, and subject to considerable uncertainty, while some behaviours may only emerge when multiple agents interact. AI systems can also recognise when they are being evaluated and alter their behaviour accordingly, further complicating attempts to assess them.
The researchers point to mechanistic interpretability – an approach that examines patterns of activity within AI systems – as one promising route towards understanding how advanced systems operate, although they stress that the field is not yet sufficiently developed to provide dependable safeguards.
Professor Osborne is Co-Director of the Oxford Martin AI Governance Initiative, which brings together technical AI research and policy analysis across Engineering Science, Politics and International Relations, and the Oxford Martin School. Its work includes research into AI evaluation and monitoring, interpretability, safety, computing oversight and international governance.
Professor Osborne and Dr Barez said: “Despite the uncertainty over the agents increasingly entangled with our lives, and the difficulty of reducing it by measurement, we must still make high-stakes decisions about AI, and make them soon.”