"The technology isn't the hard part"
Directory of Strategy
Yesterday, we hosted a roundtable about adopting AI in government with Lieutenant General Tom Copinger-Symes. A theme of the discussion was that, when it comes to putting AI into defence, it’s not the technology that’s the hard part, it’s always the people.
It perfectly captures the themes of our paper ‘Getting the machine learning: scaling AI in public services’, out today. It’s a pragmatic approach to AI adoption for those of us who are optimistic about the opportunities for AI, but realistic about the state’s capacity to do digital transformation well. The technology has huge potential, but to use it we need to change how the civil service works. The technology isn’t the hard part, it’s the people.
Personally, this isn’t much of a surprise. It’s almost seven years since I started working on AI in public services, as a newly-promoted Grade 7 civil servant in the Home Office. Since then, it’s amazing how much the technology has progressed, and how much of that is recent. When I left the civil service to work for an AI company in 2022, other civil servants were really surprised — though it was only a year later that ChatGPT exploded into the public consciousness.
In stark contrast, the policy debates Westminster has about adoption are pretty much exactly the same — almost ignorant of how much change has happened outside of its bubble:
- Yes, data sharing in government is bad, and that makes AI adoption hard.
- Yes, the skills needed are in short supply, and the civil service can’t pay competitively.
- Yes, we need to put the users at the centre of the product, whilst also realising it is an innovative technology which nobody understands perfectly (Henry Ford’s “if I had asked people what they wanted, they would have said faster horses” comes to mind)
But are those things all as true as they were seven years ago, when I sat in an overheated Home Office conference room and got to watch a random-forest simulation play out live on a laptop screen?
The technology isn’t the hard part, it’s the people.
The new Government wants to capitalise on AI’s power to transform public services, and I hope they succeed. Matt Clifford’s AI Opportunities Action Plan is a promising start. But to realise those ambitions, the State will need to change itself first.
Here are three ‘new’ ideas (and one old one) which would be game-changing for the government’s AI plans:
1. A new Government Data & AI Service
Britain led the first wave of digital transformation in government, with our Government Digital Service (GDS) inspiring similar models in the US, Singapore and Canada. GDS’ role is now focused on running a few enterprise services, and we need the missionary zeal of a team set up to drive adoption of the newest digital technology — AI. “If it ain’t broke, don’t fix it”, as the saying goes.
2. £1 billion of new investment, spent ‘VC-style’
For years we’ve been expecting that investment in government AI will follow as the benefits of early projects are shown. But this doesn’t happen. Instead, government is stuck with an acute case of ‘pilotitus’ — constantly piloting many small projects, with no route to scale them up. Interviewees we spoke to blamed a business case process which was too big, bureaucratic, and often involved fake numbers.
The government needs to spend far more, but more important than the number is how it’s spent. The government should be prepared to spend money at the speed that technology develops and is tested, far faster than an annual business case and bidding cycle works. This ‘VC-style’ approach would hold back money from being allocated to projects, but issue additional funding immediately provided pilots are successful against key metrics, with no need to wait for another approval.
3. Buying off-the-shelf
Government procurement locks the civil service into outsourcing development of AI, buying big, bespoke bits of software which take years to build. GDS innovated on that by bringing more development in-house. But we think there’s an underrated third option — buying it off the shelf. Many more companies offer off-the-shelf software, from components of a technology stack to full integrated systems, for a host of government projects. Civil servants should be able to use this quickly and flexibly, through a new framework which recognises the risks are much smaller than committing to a big, expensive bespoke project, and so the process can be much more flexible.
4. Risk parity for AI
Inaccessible, poor-performing, biased, are all descriptions you might hear a lot of AI. But they wouldn’t be out of place in a discussion of any public service powered by human intelligence either. We hold AI to unrealistic standards for performance, and criticise it for falling short of 100 per cent effectiveness — but we would do well to apply that to our own efforts as well. When considering the risk of AI, including the legal risks from automation or data sharing, this should be treated with the same parity by government as the risks of continuing to operate existing services as they currently are — as often, automation will be the less risky option.
If the government is to grasp the nettle on AI adoption, it can’t roll out the same tired problems and expect them to change magically by themselves. It needs to try something different. The technology won’t wait patiently for government to keep up, it will continue innovating all by itself — it’s for the State to keep up.