Whitehall and AI: how can government move from promising pilots to real results?

There are many reasons to be cautious about the advent of artificial intelligence (AI), but it could prove transformative for the public sector. This potential was identified by the previous government, with then prime minister Rishi Sunak personally associated with the technology, especially on an international front.

Labour, too, has shown an interest in deploying AI in public services, and its new secretary of state at the Department of Science, Innovation and Technology, Peter Kyle, presented improving public services as the third aim of his AI Action Plan. ⁶⁶ Department for Science, Innovation and Technology and HM Treasury, ‘AI expert to lead Action Plan to ensure UK reaps the benefits of Artificial Intelligence’, press release, 26 July 2024, www.gov.uk/government/news/ai-expert-to-lead-action-plan-to-ensure-uk-reaps-the-benefits-of-artificial-intelligence  The civil service is also experimenting with AI’s potential to improve its own work, which it will hope may offer some ‘quick wins’ through efficiency savings in a difficult fiscal context – and better serve the public at the same time. ⁶⁷ Department for Health and Social Care ‘Secretary of State makes economic growth a priority’, press release, 9 July 2024, www.gov.uk/government/news/secretary-of-state-makes-economic-growth-a-priority  

Whitehall is experimenting with many innovative uses of AI

To explore the role of AI in central government, the Institute for Government recently hosted two private roundtables, in partnership with Scott Logic, to investigate how it can improve the effectiveness of Whitehall’s core functions as well as the impact it may have on policy making and government communications. From these discussions and other materials we uncovered a range of established and nascent uses of AI – and new digital technologies more broadly – on Whitehall and beyond, motivated both by a desire to achieve efficiencies and by more open-ended curiosity as to AI’s potential. This should, however, be seen in the context of the NAO finding patchy adoption of AI in a recent survey: less than a quarter of government bodies it spoke to had an AI strategy. ⁶⁸ Comptroller and Auditor General, Use of artificial intelligence in government, session 2023-24, HC 612, National Audit Office, 15 March 2024, www.nao.org.uk/wp-content/uploads/2024/03/use-of-artificial-intelligence-in-government.pdf

Several tools have been developed to improve the effectiveness of existing processes in central government. The Government Communication Service (GCS) has developed Assist (previously the GCS ‘AI co-pilot’ ⁶⁹ Dunton J, ‘Government Communication Service works up innovation strategy’, Civil Service World, 24 April 2024, www.civilserviceworld.com/professions/article/government-communication-service-innovation-strategy ), which automates routine tasks to both improve content quality and deliver savings. The Government People Group developed an AI tool for optimising job descriptions, which gives advice to those wishing to advertise vacancies. The Royal Navy has also used AI to assist with recruitment by offering a new interface with potential applicants. ⁷⁰ Sutcliffe C, ’Why the Royal Navy is experimenting with an AI-powered recruitment bot’ The Drum, 17 April 2023, www.thedrum.com/news/2023/04/17/why-the-royal-navy-experimenting-with-ai-powered-recruitment-bot  Redbox, a tool created by the incubator for AI (i.AI), scans documents in ministerial red boxes and generates tailored summaries. ⁷¹ Incubator for Artificial Intelligence, ‘Redbox’, Incubator for Artificial Intelligence, retrieved 7 August 2024, https://ai.gov.uk/projects/redbox/   

i.AI has also been collaborating with Citizens Advice, a charity that provides one-to-one advice on complex issues, to produce a tool (named ‘Caddy’) that draws on the latest information from gov.uk and Citizens Advice to act as a customer service assistant. ⁷² Incubator for Artificial Intelligence ’Caddy’, Incubator for Artificial Intelligence, retrieved 7 August 2024, https://ai.gov.uk/projects/caddy/  The Government Digital Service (GDS) has also tested a chatbot called GOV.UK chat with a generally positive public response, although accuracy and reliability issues remain. ⁷³ Gregory M, Tosi A, McDonald S and Sewell R, ‘The findings of our first generative AI experiment: GOV.UK Chat’, blog, 18 January 2024, retrieved 29 August 2024, https://insidegovuk.blog.gov.uk/2024/01/18/the-findings-of-our-first-generative-ai-experiment-gov-uk-chat/

In the policy sphere, the Civil Service’s Policy Lab and Policy Profession Unit have experimented with using Pol.is, an open-source software that gathers collective intelligence and analyses voting results. ⁷⁴ Holden L, ‘Using collective intelligence in government: a view from the Policy Profession Unit’, blog, 13 December 2022, retrieved 7 August 2024, https://openpolicy.blog.gov.uk/2022/12/13/using-collective-intelligence-in-government-a-view-from-the-policy-profession-unit/  AI is also being used on the Hansard database to identify politicians who may be interested in specific policies, and the Office for National Statistics (ONS) is using AI to improve social statistics by assigning free text responses to pre-defined categories that make them easier to analyse. ⁷⁵ Data Science Campus, ‘ClassifAI – Exploring the use of Large Language Models (LLMs) to assign free text to commonly used classifications’, Data Science Campus, 9 July 2024, retrieved 7 August 2024, https://datasciencecampus.ons.gov.uk/classifai-exploring-the-use-of-large-language-models-llms-to-assign-free-text-to-commonly-used-classifications/

Several roundtable participants mentioned either currently using AI in public consultations or expressing a wish to do so – although mindful of the risks involved, including a potential for gaming by respondents. i.AI has created a large language model (LLM) to analyse the responses to consultations and extract recurring themes, and the tool is being tested in various departments. ⁷⁶ Incubator for Artificial Intelligence, ‘i.AI Consultation Analyser‘, Incubator for Artificial Intelligence, retrieved 7 August 2024, https://ai.gov.uk/consultations/  AI is also being used to extract insights from media and social media that would not ordinarily be input into consultation processes. Ministerial correspondence and public complaints to government departments are also being analysed by AI, and LLMs are being used to improve search experience on the ONS website. ⁷⁷ Data Science Campus, ‘Using large language models (LLMs) to improve website search experience with StatsChat’, Data Science For the Public Good, 9 July 2024, retrieved 7 August 2024, https://datasciencecampus

Legacy IT systems continue to hold back progress

Whitehall, and public services more broadly, have long been plagued by legacy IT systems that contain outdated, aging data that is difficult to extract, share and use, and can pose operational and security risks. For example, the Pension Services Computer System (PSCS) was built in 1988 and manages records of pension payments of around £104 billion. The main database of criminal records, the Police National Computer (PNC), is even older, dating to 1974. ⁷⁸ Comptroller and Auditor General, Digital transformation in government: addressing the barriers to efficiency, session 2022-23, HC 1171, National Audit Office, 10 March 2023 www.nao.org.uk/wp-content/uploads/2023/03/digital-transformation-in-government-summary.pdf

These legacy systems limit the quality of the data on which AI can be trained and on which they rely to function. Upgrading legacy IT systems is necessary to fully unlock the potential of AI in government. But this is expensive, long-term work that carries delivery risks – making it all too easy for decision makers to put off. Past efforts have also been hampered by lack of political will, funding deficits and coordination problems. ⁷⁹ Comptroller and Auditor General, Digital transformation in government: addressing the barriers to efficiency, session 2022-23, HC 1171, National Audit Office, 10 March 2023 www.nao.org.uk/wp-content/uploads/2023/03/digital-transformation-in-government-summary.pdf  One project is attempting to use AI to improve data sharing across government ⁸⁰ Incubator for Artificial Intelligence, ‘rAPId‘, Incubator for Artificial Intelligence, retrieved 7 August 2024, https://ai.gov.uk/projects/rapid/  – itself a wider problem that also relies on co-ordination between departments and agencies. ⁸¹ Freeguard G and Shepley P, Government data sharing during the pandemic, Institute for Government, 2023, www.instituteforgovernment.org.uk/publication/government-data-sharing-pandemic   

But it is not yet clear whether aiming to enable the full benefits of AI will be a catalyst for solving underlying data weaknesses, or whether AI will instead be used as a sticking plaster that can enable system upgrades to be further deferred.

A clear evaluation process is needed to select which pilots to scale up

The proliferation of pilots in Whitehall is a promising sign; creative experimentation is crucial for identifying areas where AI can help. But some pilots will inevitably fail, and both officials and politicians must be tolerant of this. The current principles-based guidance for civil servants using AI is appropriate to this experimentational phase – and perhaps beyond. ⁸² Central Digital and Data Office, Generative AI Framework for HMG, gov.uk, 18 January 2024, www.gov.uk/government/publications/generative-ai-framework-for-hmg  GDS has already adopted a ‘fail fast’ approach to developing its digital products. ⁸³ Worlidge J, Clyne R, Nye P and others, Whitehall Monitor 2024, Institute for Government, 2024, www.instituteforgovernment.org.uk/publication/whitehallmonitor-2024, page 87  But some consistent measure of success is needed to identify which tools are sufficiently useful to roll out more widely and which should be discontinued; better communication across government could enable projects to learn more quickly from others’ achievements – and mistakes. We understand that GCS’s Assist is the first tool powered by general purpose AI to be approved for use across government, so others may be able to learn from how this was achieved.

It is not yet clear who will be responsible for evaluating AI pilots and making the decision about whether to operationalise them further. The newly strengthened Department for Science, Innovation and Technology (DSIT), ⁸⁴ Department for Science, Innovation and Technology, ‘DSIT bolstered to better serve the British public through science and technology’, press release, 8 July 2024, www.gov.uk/government/news/dsit-bolstered-to-better-serve-the-british-public-through-science-and-technology  which brings together GDS, the Central Digital and Data Office (CDDO) and i.AI, could be a sensible locus for this – as well as a much-needed point of co-ordination for the range of activity underway. But it will need to develop and maintain the authority to perform this function as a department outside the traditional ‘centre’ of government, balanced by a willingness to partner with those areas of government that are already innovating, rather than seeking to crowd them out or control them.

A “human in the loop” is necessary – but may not be sufficient – to maintain public trust

Digital tools misfiring, particularly in ways that differ from mistakes a human might make, can be deeply damaging for public trust (recent examples include a facial recognition tool that makes more mistakes with Black faces and an algorithm that flags more potential sham marriages among certain nationalities). ⁸⁹ Stacey K, ‘UK officials use AI to decide on issues from benefits to marriage licences’, Guardian, 23 October 2023, www.theguardian.com/technology/2023/oct/23/uk-officials-use-ai-to-decide-on-issues-from-benefits-to-marriage-licences  In the private sector, a loss of trust could result in the displacement of one provider by another, but in government it may more likely lead to a range of such harms, or scepticism of AI being used to perform a function at all may preclude further innovation. Maintaining trust is both critical and difficult: ⁹⁰ Department for Science, Innovation and Technology, Introduction to AI assurance, gov.uk, 12 February 2024, www.gov.uk/government/publications/introduction-to-ai-assurance  research conducted by the Department for Transport suggests the public may hold AI systems to a higher level of accountability than humans. ⁹¹ Thinks Insight and Strategy and Department for Transport, Using AI in consultations and correspondence September 2023, Department for Transport, https://assets.publishing.service.gov.uk/media/654e6f078a2ed4000d720d12/using-ai-in-consultations-and-correspondence.pdf, page 8.   

A key means to maintain trust is for a “human in the loop” to retain the final say over key decisions, with AI acting as a “co-pilot” – producing first drafts of documents that are then edited or signed off by humans, for instance. Assist allows for this, requiring users to review outputs as well as offering pre-built prompts bespoke to common communications use cases (in addition to GCS’s broader generative AI policy for government communication). During our roundtables, we heard that AI was being piloted on activities that appeared less judgemental initially – such as processing data or drafting less politically sensitive documents – but that if successful its role could expand. The incentives on time-constrained humans may be to trust their AI tools more and more, so ensuring their involvement and responsibility remains substantive will be key.

AI could change Whitehall – but there is more to do

As well as improving outcomes, AI could be instrumental in generating efficiency savings in Whitehall. Such savings are much needed in the current fiscal context, and the enthusiasm we found across Whitehall for using AI to improve existing processes, as well as bringing new approaches to government communications and policy making, is welcome.

Nonetheless, roundtable participants cautioned that AI could also substantially change the experience of work on Whitehall. AI could automate routine tasks across government, freeing up civil servants’ time for tasks where a human has a comparative advantage – such as those involving evaluative judgement. Estimates of the potential change are fraught with assumptions, but researchers at the Alan Turing Institute, for example, estimate that 84% of complex but repetitive citizen-facing transactions could conceivably be automated. ⁹² Dungate J, ‘AI could help automate around 84% of repetitive service transactions across government’, press release, 20 March 2024, retrieved 29 August 2024, www.turing.ac.uk/news/ai-could-help-automate-around-84-repetitive-service-transactions-across-government  As it does this, however, AI could also lead to job losses and job intensification. Careful workforce planning will be required, involving unions and employees in decision making.

The government should seek to shape, rather than resist, the change AI will bring. It should continue to encourage the innovation that is taking place. Alongside this, it needs to develop clearer processes for sharing experience, evaluating pilots, and triaging which initiatives are and are not taken forward. To maintain public trust it must be transparent as to how humans remain involved in decisions – and, indeed, on the trade-offs that may be involved in the expansion of AI’s future role. These steps will be vital to ensuring that these innovations serve those they are intended to, as effectively as they can.