Why Government is ready for AI

by Richard Stirling and Emma Truswell

The story of public sector reform has been one of big promises, big investment, good intentions, and often disappointing results.  Over the last 40 years we have seen a huge drive for efficiency - from the initial time and motion studies in the 70s, efficiency reviews in the 80s, New Public Management in the 90s, operational efficiency in the 2000s and now digital transformation in 2010s. Each has struggled to realise the full potential promised because public services are difficult to reduce to simple rules. Lives are complex and helping people when they most need help requires sound judgment and a sophisticated understanding of how to separate cause and effect, manage expectations among citizens, and get the politics right within government.

The potential of Artificial Intelligence comes from its ability to deal with the complexity of real life. It offers is the beginnings of computer programmes that can make judgments - rather than simply following preset rules.

A conventional programme takes inputs it has been given, follows a designed process, and ends with an output that is assessed by a human brain to help make decisions to create a certain outcome.  Recent strides in Artificial Intelligence allow the human brain to define the desired outcome, and to leave the rest to the computer programme. The programme can then work out the inputs it needs, the process it wants to follow, which outputs are useful, and how to apply those to make a decision.

The current talk about AI is focussed on machine learning - where how to best achieve an outcome is worked out iteratively by the machine as it learns. While these techniques are still in their infancy in public services, the underlying technology is advancing rapidly, such as being used to identify cancer in a clinical setting. However the most striking example is self driving cars e.g. this video showing how Tesla’s system identifies risks in real time:

This shows how sophisticated this technology can feel, and also gives insights into how the car is ‘thinking’. The car picks out features, tries to classify them and then react accordingly. As it encounters a new object or a new scenario it learns from how it dealt with it before - learning from its mistakes.

On top of all the usual requirements for change in the public sector, machine learning needs a few things to be effective (and thanks to previous waves of e-government reform the public sector has services with them):

  • Defined outcomes - so that the machine knows whether it has been successful or not (this is one reason why early testing for Deepmind was on computer games)

  • Robust data infrastructure - so there is data on which the machine can operate

  • Large transaction volumes - so that there are lots of examples to learn from and experiment with.

These possibilities raises many important questions: about where are the best places to try out artificial intelligence in government, about the ethics and accountability of computer decision making, about the future of jobs, about security and the democratic process, about the role of the civil service and what kinds of leaders we will need, to name a few. Oxford Insights looks forward to publishing on these questions and more over the coming months.

If you’d like to work with us on these or other topics, we’d love to hear from you.

AIRichard Stirling