Last week saw a major shake-up of the UK’s science, technology and digital policy landscape – as the Department for Science, Innovation and Technology was stood down and its responsibilities shared among other departments. Much of the mainstream discussion around these changes seems to have focused on the wisdom of a machinery of government change at this point and the elevation of AI policy to have a dedicated minister attending cabinet.
As chair of the AI Task Force at the Royal Statistical Society, I have been struck by how these changes – and the response to them – seem to be driven by a view of AI as a standalone policy area. For statisticians, AI is a fundamentally statistical technology – it is built on statistical pattern recognition. This means that the successful application of AI in any domain – as with statistics – depends on the quality of the digital infrastructure, data assets and governance arrangements on which it is built.
From this perspective, separating AI policy (split between the Cabinet Office and the Department for Business, Innovation, Science and Trade) from digital and data policy (in the Department for Culture, Media and Sport) is concerning. The RSS has been making the case that statistics, especially statistical approaches to evaluation, should form a central part of how AI is deployed and regulated. This means thinking about AI in the context of the broader digital and data environment – there is a risk of serious issues arising if this is missed.
The RSS makes this argument in our recent paper AI Regulation Needs Statistics, and the risk is made most clear by looking at specific case studies. Take the use of AI in healthcare and how that increases the risk of deanonymisation. As AI models and systems become more powerful, they become increasingly capable of identifying patterns and connections in anonymised datasets in ways that a human simply never could. This means, as the Information Commissioner’s Office has recognised, that it is no longer enough to simply treat anonymisation as a feature of a dataset – instead, it is a complex interaction between the dataset, the AI models that are used on it and the wider digital and data ecosystem.
This example indicates how questions around data security, which AI models can be used on which datasets and in which environments are all closely connected. Separating the parts of government dealing with AI policy and digital and data policy and regulation into different departments risks making effective decisions in a complicated environment harder than it already is. To help the government and civil servants make the best possible decisions about where AI is used and how it is regulated, we need strong data foundations and investment in statistical capability.
"It is no longer enough to simply treat anonymisation as a feature of a dataset – instead, it is a complex interaction between the dataset, the AI models that are used on it and the wider digital and data ecosystem"
The new prime minister’s desire to make the most of AI – through ensuring it is widely adopted in the public sector and that innovation is supported is understandable and welcome. But we cannot lose sight of the importance of doing this in a trustworthy and effective way – if we are to see sustainable benefits from AI, it will be because people trust how it is governed and used. This is going to depend on developing AI policy on solid foundations, which means ensuring a close relationship between AI policy and digital and data governance.
Digital and data policy will be going to DCMS, removed from the centres of gravity for AI decision-making. If government does not remain mindful of the connections and interdependencies between these policy areas – and active in cultivating links between them – there is a risk that the UK’s AI strategy becomes untethered from the digital and data foundations upon which it rests. This would make the government’s laudable ambitions for AI-driven growth and public service transformation much harder to realise.
Donna Phillips is chair of the Royal Statistical Society's AI Task Force