Written evidence submitted by Ada Lovelace Institute (DCG0018)
Digital centre of government
Ada Lovelace Institute response to Science, Innovation and Technology Committee call for evidence
This document contains the Ada Lovelace Institute’s response to a call for evidence on the ‘digital centre of government’ from the Science, Innovation and Technology Committee. It synthesises lessons from Ada’s existing and forthcoming publications on AI in the public sector, including:
Digital government is a broad term which encompasses the use of AI in government but cannot be reduced to it. As such, the list of priorities we set out is not an exhaustive one: we would urge the Committee to take a broader view of digital government that looks beyond AI technologies.
For more information on any of the information contained in this response, please contact Ada’s Economic and Social Policy Lead Matt Davies (mdavies@adalovelaceinstitute.org) or our Associate Director (Society, Justice and Public Services) Imogen Parker (iparker@adalovelaceinstitute.org).
The Ada Lovelace Institute is an independent research institute with a mission to make data and AI work for people and society. This means making sure that the opportunities, benefits and privileges generated by data and AI are justly and equitably distributed.
We do this by:
Our work is currently spread across three research directorates:
The public sector is an important context for AI adoption, and getting ‘digital government’ right in 2025 necessarily means getting AI adoption right. Speaking to government and public service leaders last year, we found significant optimism that generative AI might be able to support tasks such as document analysis, policy design and answering public enquiries. We were encouraged to see the Government reiterate its plans to build on this optimism by making DSIT a ‘digital centre of government’, and welcome the decision of the Science, Innovation and Technology Committee to open an inquiry into the government’s plans.
Consolidating digital, data and AI expertise in one department is a positive step. However, we believe that the potential for widespread adoption of AI across the public sector is sharply limited by both an ‘information gap’ and a ‘governance gap’:
After providing some preliminary information on the current role of AI in the public sector, this response suggests steps towards fixing both gaps. We believe that creating a digital centre of government need not reinvent the wheel or disrupt existing work, but that it should address the information and governance challenges currently holding rollout back, and allow for a more coherent and joined-up approach to the roll-out of AI technologies across the public sector.
The National Audit Office’s 2023 survey found that just over a third (37%) of responding bodies were actively using AI, and a further 37% were actively piloting (25%) or planning (11%) use of AI. This is likely to be an underestimate for a number of reasons,[1] but it is difficult to assess the true extent of AI use across the public sector because there is currently no common resource or obligation to monitor where and how AI tools are being used, even within central Government.
Between June and July 2023, as part of our research on foundation models in the public sector, we spoke to people across central and local government, and public bodies and looked at use cases suggested by think-tanks and industry.[2] Drawing on those, we identified a list of potential use cases for AI (and specifically foundation models) in government:
These use cases could potentially benefit public services by improving service efficiency or productivity, or by improving the quality of a service. However, we believe that most estimates of the potential cost savings from AI use should be treated with caution. Some studies are compromised by flawed methodological approaches, and all are limited by our poor understanding of the impacts of existing AI use in the public sector. To enable the widespread use of AI across Government and public services, we need to first understand which AI interventions are likely to be most effective.
Improving monitoring and evaluation of AI in government and public services
Across our work, from looking at AI procurement to deployment of foundation models across the public sector, we have seen and heard that there’s a real problem with visibility across government about where AI projects are being piloted and deployed.[3][4] There remains a persistent, systemic deficit in understanding of where and how AI and data-driven systems are used in the public sector – both among the public and across the sector itself.[5] This reduces the ability for journalists, civil society, academia and ultimately the public to scrutinise and ultimately hold government accountable for its use of AI. Not only that, it hampers knowledge sharing across the public sector about what is working and delivering value for money.
We see this as a basic and foundational requirement for the successful and trustworthy implementation of AI in the public sector. Peter Kyle, the secretary of state for science and technology, has admitted the public sector ‘hasn’t taken seriously enough the need to be transparent in the way that the government uses algorithms’.[6]
Part of the problem is that there are a number of mechanisms for transparency that, taken individually or combined in the limited ways currently possible leave us far from ensuring that we are capable of scrutinising and evaluating automated decision making or AI systems in the public sector. These include impact assessments, procurement documents, open data, FOIs and standardised disclosure of data. [7]
Many of these mechanisms are not mandatory, and those that are still suffer from a lack of enforcement or not enough support for public authorities in ensuring best practice.[8] There is no systematic resource or obligation to monitor the full spectrum of where and how AI is being used in government. As we discuss below, the Algorithmic Transparency Reporting Standard (ATRS) represents a significant step forward in this regard but remains limited in scope. Consequently, neither the Government, nor the public, have a full understanding of what AI tools are being deployed where. The National Audit Office’s (NAO) survey of AI use in government is a useful snapshot in time of AI use,[9] but it is also symptomatic of the transparency problem that this data was not already easily available and required a custom survey by a government scrutiny body.
As a corollary of this, there is little independent analysis being done of the impact of AI tools on services and society. The analysis that does exist is rarely systematic or standardised enough to draw robust conclusions from, in part because there exists no coherent set of approaches to consider, understand and monitor the impacts of AI in the public sector. That leaves individual public sector procurers and users overly reliant on sales pitches from private providers, supplemented by the independent research that does exist from the media and civil society. Our research with representatives from local and central government representatives have indicated a strong demand for continuous monitoring and evaluation of AI implementations, to provide procurers and deployers in the public sector with the information they need to make appropriate decisions about when and how to adopt AI.[10]
An urgent priority for the digital centre of government should therefore be to develop a more detailed understanding of where and how AI is being used in public services, and what is currently working to improve outcomes and services for the public. The new unit should be tasked with an immediate review of the state of AI in Government – building on the NAO's report and other independent evidence – and with ongoing monitoring of AI deployment.
As part of this, it should continue efforts to roll out the ATRS across Government. The publication of the ATRS in 2021 was seen as a landmark achievement. It could have paved the way to improve transparency about government use of AI, and enabled better monitoring and evaluation, but progress stalled for several years before the previous government belatedly announced plans to make the ATRS mandatory in February 2024. [11] Recent months have seen a welcome uptick in the number of records uploaded to the government’s centrally-hosted repository, with 55 ATRS records published at time of writing and more expected to be brought online in the coming weeks and months.
This acceleration in ATRS activity should be commended, and represents real progress towards the goal of improved algorithmic transparency across government. That said, the limitations of the standard in its current form should also be recognised, and remedying these limitations ought to be an early focus for the digital centre of government. The ATRS records currently published on gov.uk are unlikely to be comprehensive when set against the scale of AI use across government, and key departments such as the Department of Work and Pensions are not yet featured. Government should consider whether to monitor whether the scope of the ATRS as currently designed is sufficient, given that it only covers a subset of AI used in the public sector. Further work needs to be done on ensuring adequate transparency for uses of AI that are not current captured: notably cross-departmental implementations, and informal uses of commercially available AI tools. Even for those tools already listed, government will need to work on an ongoing basis to ensure records are kept up-to-date as the tool is updated, refined or decommissioned.
Government should also consider the ways in which public sector leaders and procurers are affected by the information asymmetries prevalent in the AI market (see specific findings from our procurement research below). The regulation of AI firms in the private sector is often seen as a separate policy agenda from rolling out AI across public services, but the two are in fact deeply linked. Government teams often buy from “upstream” developers in the private sector and then adapt or build on their models for a range of “downstream” applications, and it is imperative that they understand both the effectiveness and the potential risks of the technologies they are building upon. Binding regulation on foundation model developers, and appropriate transparency requirements for the providers of other AI tools, will be an urgent first step towards providing the assurance that public sector leaders need to procure, deploy and use AI with confidence.
Beyond simple transparency, it is vital that government understands whether a given AI application is effective relative to other interventions (and many will not be) so that it can target resources at scaling up tools that are a net-positive and discontinuing harmful or ineffective interventions. Critically, these evaluations need to acknowledge the broader service and social context that AI systems are being deployed in, and take an iterative approach that remains sensitive to potential changes as models and systems are updated over time.
Our work on the early deployment of foundation models in the UK public sector found a need for continuous monitoring and evaluation, beyond initial development and procurement of foundation model applications. Local and central government representatives in our roundtables felt that this continuous monitoring and evaluation by public services, of both public and private applications, was needed to ensure foundation model systems operate as intended and discover when AI systems are failing to deliver results.[12]
Particularly in the context of using foundation models, e.g. GPT-4 or Gemini, our work on evaluations has shown the importance of contextual evaluations. While model evaluations can provide useful information about a model's overall capabilities and general risks, existing model evaluations have a number of issues. These evaluations often lack external validity, where results fail to generalise to real-world conditions, and results can vary significantly with small changes to the model or context. These evaluations also often overlook the sociotechnical context: how users interact with the model, the design of its interface, and its broader societal impacts, e.g. embedding systematic biases.[13]
Deployers of AI in the public sector should take a both systematic and iterative lens on evaluating deployments of AI, understanding both their immediate impact and how users, workers, and wider society changes and adapts to their use over time.
The government has recently published ‘Guidance on the Impact Evaluation of AI Interventions’.[14] This guidance builds of the government’s existing evaluation guidance, the Magenta Book, and has a welcome focus “on the systematic assessment of the outcomes of an intervention with the aim of establishing whether, to what extent, how and why an intervention has resulted in its intended impacts.” This guidance is a good start, but as the guidance itself notes, robust evaluation will also require drawing on the knowledge of evaluation experts in a given domain, e.g. social care or education.
Securing legitimacy for responsible AI rollout across the public sector
Our body of research suggests that AI tools in the public sector are more successful when they are trusted by the public and have social license to operate. A nationwide survey of the UK public carried out by Ada in conjunction with the Alan Turing Institute found that 62 per cent of respondents were in favour of ‘laws and regulations that prohibit certain uses of technologies and guide the use of all AI technologies.’[15] In general, the public would like to see regulation on the use of data and AI, [16] and expects organisations and governance structures deploying AI to be trustworthy. [17]
While government strategies and guidance on public sector AI have proliferated in recent years, our research suggests that there remains a significant governance gap which risks public backlash against AI use and the withdrawal of consent for data use. There is real precedent for this: anxiety around health data sharing around the General Practice Data for Planning and Research (GPDPR) led to a spike of people withdrawing consent for health data to be shared and used for research (levelling out at over 3 million).
Cross-cutting legislation on data protection and equality governs the development and deployment of AI in government, and specific applications of AI (such as the use of AI to make automated decisions) will trigger additional protections such as those found in Article 21 of the UK GDPR. There is a range of non-statutory guidance from central Government departments, regulators, and other organisations that is relevant to the procurement, deployment and use of AI in the public sector.
However, our research suggests that current regulation and guidance isn’t fit for purpose to offer clarity and confidence for public sector actors in deploying and procuring well-established AI technologies, and is certainly not comprehensive for more capable and novel systems such as foundation models.[18] Guidance needs to be reviewed regularly in order to keep pace with the development and deployment of these technologies, and streamlined to ensure that it can be implemented quickly and effectively in practice. Public sector actors need greater specificity and operationalizable advice than is currently available.
It is promising that in the case of Generative AI, the Government did publish initial guidance for public servants[19] in June 2023 and then followed that up with a more comprehensive Generative AI framework[20] once they had time to more fully assess and articulate practical considerations for the use of generative AI in government.
Guidance is a complement to, rather than a substitute for, ‘harder’ forms of regulation underpinned by statute. In particular, existing provisions in the UK GDPR provide important protections against harms such as biased or unaccountable algorithmic decision-making, including in the public sector. Far from being outmoded, independent legal analysis[21] carried out by law firm AWO suggests that these protections are vital to ensure the safe deployment of newer AI technologies like foundation model chatbots. As such, it will be critical to protect the fundamental accountability framework in UK GDPR.
The monitoring and enforcement of relevant legislation in the public sector can be patchy.[22] We have called for a clear plan for how the five cross-sectoral AI principles set out in the previous government’s white paper on AI regulation will be enacted in government, including who will be responsible for monitoring and enforcing them. If this government chooses to maintain and build on the previous government’s sector-based approach, then this will be crucial, and will likely require the introduction of a statutory duty mandating public sector bodies to implement the AI principles.
A vision for public sector AI use
Many of the challenges associated with AI in the public sector would be eased, if not solved entirely, by getting the basics right. In the interests of ensuring that AI procurement rollout – which is already happening, at pace, across large swathes of the public sector – proceeds in a way that works, and works for everyone, we think it is right that the digital centre of government prioritises fixing the information and governance gaps we have identified above. In the medium term, however, there is a pressing need for central government to provide a vision for public sector AI use to ensure its legitimacy over the longer term.
If there is a single overarching lesson from our research, it is that AI is irreducibly sociotechnical. It influences and is influenced by the social contexts in which it is deployed, often creating unintended and profound ripple effects. When used in government, AI technologies have the potential to transform for better or for worse the relationship between people and public services. Conversely, they can also risk entrenching existing business-as-usual approaches to delivery at the expense of fresh thinking and systemic solutions.
One upshot of this is that rolling out AI across the public sector should not be seen as a ‘quick win’. While there is optimism about the potential for cost savings and productivity gains, there is little evidence of them working in practice yet and existing studies claiming imminent savings for the exchequer should be viewed with caution.
Moreover, quantitative improvements may need to be weighed up against qualitative changes in the experience of those who deliver and use public services. In some cases AI interventions may need to be paired with other investments (such as in skills training, or in new posts) to yield the desired benefits. In others, AI interventions may not be desirable: AI only works when applied to the right kinds of problems, complementing skilled public service professionals rather than replacing them. For some jobs – such as frontline social workers, or wellbeing councillors – automation may not be desirable at all. In these instances, approaches such as ‘ringfencing’ may be desirable to safeguard particular tasks from automation.[23]
These aren’t calculations that can be made on a balance sheet. They are fundamentally values-based questions about how and for whom the state is run. The answers to these questions ought to be informed not by hype or vested interests but by political leadership, informed by independent expertise and democratic input. The task is nothing less than to develop a new consensus vision for the state and public services in the AI era – and as the proposed home of specialist expertise for digital, data and AI within government, the digital centre of government is well placed to support this work.
Fixing the ‘governance’ and ‘information’ gaps that currently exist will provide the foundations for this, by providing Government with a clearer-eyed view of what AI could deliver for the public. But we believe that the views of the public – particularly those of the diverse groups of people who use and work in public services – should be at the heart of this process. This not an idealistic or ideological appeal to the wisdom of the masses, but rather a pragmatic acknowledgement that those delivering and using services are likely to have important perspectives on how public sector automation should proceed.
Deliberative and participatory approaches to policymaking should be integrated into the work of the digital centre of government to help ministers and senior officials understand how frontline professionals envisage their roles changing, and what members of the public expect and want from their services.[24] These could be complemented with more traditional modes of input such as consulting worker representatives and implementing user oversight groups.[25] The government has already acknowledged that ‘the public sector cannot deliver next-generation public services alone’[26]: institutionalising deliberative and participatory approaches will help to ensure the vision for the digital centre of government is developed hand in glove with frontline workers and the British public.
03 March 2025
[1] The NAO explicitly excluded AI ‘embedded’ in other software packages (for example automatic email spam filtering) and ‘ad-hoc’ use of AI (such as the informal use of ChatGPT and Bard by Government employees).
[2] See Appendix 1: Examples of potential public sector applications of foundation models, https://www.adalovelaceinstitute.org/evidence-review/foundation-models-public-sector/#appendix-1-examples-of-potential-public-sector-applications-of-foundation-models-59
[3] https://www.adalovelaceinstitute.org/report/spending-wisely-procurement/
[4] https://www.adalovelaceinstitute.org/policy-briefing/foundation-models-public-sector/
[5] https://www.adalovelaceinstitute.org/blog/meaningful-transparency-and-invisible-algorithms/
[6] https://www.theguardian.com/technology/2024/nov/28/uk-government-failing-to-list-use-of-ai-on-mandatory-register
[7] https://www.adalovelaceinstitute.org/wp-content/uploads/2020/10/Transparency-mechanisms-explainer-1.pdf
[8] https://www.adalovelaceinstitute.org/wp-content/uploads/2020/10/Transparency-mechanisms-explainer-1.pdf
[9] https://www.nao.org.uk/reports/use-of-artificial-intelligence-in-government/
[10] https://www.adalovelaceinstitute.org/project/procurement-ai-local-government/
[11] https://www.gov.uk/government/consultations/ai-regulation-a-pro-innovation-approach-policy-proposals/outcome/a-pro-innovation-approach-to-ai-regulation-government-response#summary-of-consultation-evidence-and-government-response
[12] https://www.adalovelaceinstitute.org/evidence-review/foundation-models-public-sector/#continuous-monitoring-and-evaluation-45
[13] https://www.adalovelaceinstitute.org/wp-content/uploads/2024/09/Ada-Lovelace-Institute-Under-the-radar-230924.pdf
[14] https://www.gov.uk/government/publications/the-magenta-book/guidance-on-the-impact-evaluation-of-ai-interventions-html
[15] Ada Lovelace Institute, ‘How Do People Feel about AI?’ (n 1) 41.
[16] Ada Lovelace Institute, ‘What Do the Public Think about AI?’ (n 22) 19.
[17] Ada Lovelace Institute, ‘The Rule of Trust’ (n 12) 4.
[18] https://www.adalovelaceinstitute.org/evidence-review/foundation-models-public-sector/
[19] https://www.gov.uk/government/publications/guidance-to-civil-servants-on-use-of-generative-ai#full-publication-update-history
[20] https://www.gov.uk/government/publications/generative-ai-framework-for-hmg/generative-ai-framework-for-hmg-html
[21] https://www.awo.agency/blog/awo-analysis-shows-gaps-in-effective-protection-from-ai-harms/
[22] https://www.adalovelaceinstitute.org/report/regulating-ai-in-the-uk/
[23] https://www.ippr.org/articles/transformed-by-ai
[24] https://www.adalovelaceinstitute.org/evidence-review/what-do-the-public-think-about-ai/#finding-7-there-is-a-significant-body-of-evidence-that-demonstrates-ways-to-meaningfully-involve-the-public-in-decision-making-25
[25] https://www.tuc.org.uk/research-analysis/reports/ai-bill-project
[26] https://www.gov.uk/government/publications/a-blueprint-for-modern-digital-government/a-blueprint-for-modern-digital-government-html