Why government needs resilient AI, not just more AI

For Paul Jenkinson, government departments wishing to bridge the gap between high expectations and limited budgets are presented with three options.

The first two are to hope for a sudden increase in funding, or to pray that the demands placed on the organisation lessen - both “unlikely” outcomes, the CEO of British AI company Whitespace argued. The third is to look to AI.

Jenkinson’s analysis follows years of helping government organisations, initially in defence, to develop AI solutions. However, the self-proclaimed “tech optimist” argues that government should do two things to realise AI's potential : build resilience into its technology, and develop specific systems to support decision-making in complex environments.

Resilience, in his view, is concerned with minimising the extent to which external factors can disrupt the functioning of a service or piece of technology. These could range from certain LLMs suddenly being withdrawn from circulation, to unpredictable staff behaviour, to data centres being compromised.

He contrasted this crucial idea of resilience with a narrow understanding of sovereignty, arguing that “it's not about where our headquarters are, it's about whether a government department can rely on [a certain] service or capability”.

“It’s actually quite a complex thing”, he added.

One step towards resilience is avoiding excessive dependencies on any singular model, provider, individual or piece of infrastructure, meaning that if there is a problem with one component, the system does not develop problems along with it.

Alongside this, he believes that building systems specifically to support decision making in complex environments would help leaders to make better choices. 

In his view, this is a task at which LLMs are not particularly adept, with the fact that they were trained on historical data limiting their usefulness when it comes to giving advice on determining the best course of action in complex and rapidly changing environments. 

“If you're going into a life or death game of chess, you don't go in with a large language model, you go in with a chess engine”, he said.

He added: “A large language model can describe where the pieces are on the board, it can very eloquently describe everything, but what it is not designed to do is tell you which move is going to make you win. And essentially, government is all about decisions.”

While key to building systems for particular environments, he also warns against “over-optimis[ing]” for a specific function, instead highlighting the importance of analysing situations holistically while balancing competing factors.

For Jenkinson, how government builds AI cannot be separated from the question of who it buys that technology from.

Working with British technology companies can help government develop systems that are better attuned to the UK’s particular institutional and cultural context, Jenkinson said, arguing that greater familiarity with how public services operate can help reduce miscommunication.

He believes, however, that government can be too inclined towards established suppliers, making it harder for smaller British companies to demonstrate what they can offer.

In this way, he argues that procurement contracts also make a greater difference to British SMEs than funding programmes alone. While British SMEs cannot match large AI companies in terms of financial resources, procurement gives companies the opportunity to prove and deploy their technology in real government environments.

The end goal, Jenkinson argues, is to improve government delivery for citizens.

He said: “This technology has the power to change everybody's lives, from children who are really struggling at school, the social care system, to healthcare. 

“It has massive, massive potential, but you need to bring in people who can harness it in an ethical and trusted way.”