Application of Generative AI in the Public Sector

The potential value of generative AI-based solutions in the public sector

The lack of specialist data analytics skills and employee capacity within most city and regional authorities is limiting their ability to leverage data they are already collecting. This hinders efforts to reduce congestion, tackle significant economic costs, and achieve ambitious net-zero targets.

Large language models (LLMs) are advanced generative AI systems trained on extensive text data to understand, generate, and analyse human language. Unlike traditional tools, which require specialised skills or focus on isolated datasets, LLM-based solutions can aggregate and analyse diverse data sources to offer insights tailored to locally specific needs. This unique capability breaks down data silos, improves data accessibility, and enables public sector organisations to make data-driven decisions, increase employee capacity and efficiency, enhance service delivery, and advance their social, economic, and environmental goals.

This briefing paper focuses on the use of LLMs in the context of transport related data, including traffic sensor counts, weather, air pollution sensor data and roadworks, highlighting lessons learnt and recommendations that can be applied across other uses of LLMs in the public sector.

Deploying LLM-based generative AI in the public sector

Deploying generative AI in the public sector presents key challenges, including algorithmic bias, lack of transparency, and reduced accountability. LLMs trained on large datasets can perpetuate societal biases and lead to discriminatory outcomes.

Therefore, the deployment of LLM-based generative AI in the UK public sector requires adherence to key principles, such as of transparency, fairness, accountability, and social benefit. The UK Government's "Model for Responsible Innovation" framework highlights the need for public sector bodies to assess the societal impact of AI systems at every stage of their development and deployment. This framework, underpinned by principles of ethics, underscores the importance of aligning AI technologies with public interest, maintaining public trust, and mitigating unintended harms. Furthermore, the complexity of these models often makes their decision processes opaque, which underscores the requirement for explainability to build trust, as highlighted in the Department for Science, Innovation, and Technology's (DSIT) guidance on AI ethics.

Policymakers must address these risks by ensuring that AI systems are trustworthy and explainable, to help address potential misinformation, and safeguard data privacy. Balancing innovation with the protection of public value is crucial for an ethically-informed integration of LLMs in delivering equitable and reliable public services.

TraffEase

The TRansport AI innovation CEntre (TRAICE) at Loughborough University is one of the UK’s largest transport research and innovation centres and the first to offer a fully interdisciplinary approach to address critical transport challenges. A recent project, TraffEase, is pioneering LLM-based transport data solutions to make advanced analytics accessible to non-experts through natural language query, enabling everyone to extract actionable insights from complex datasets. It is a top-10 finalist in the UK government’s leading prize for AI for public good, the Manchester Prize, initiated by DSIT. The technology has been test-deployed with Nottingham City Council and Transport for Greater Manchester, both of which have prioritised careful consideration for the responsible deployment of AI, guided initiatives such as the Greater Manchester Combined Authority AI Task and Finish Group. These efforts offer valuable use cases for the deployment of Generative AI in the public sector.

Ethical Considerations

The primary focus was to ensure that the systems produce trustworthy results by carefully evaluating the design of the TraffEase tool, its intended purpose, and the stakeholders involved. This approach aims to uphold—and even enhance—foundational values and principles, including internationally recognised human rights frameworks, the SDGs, and ethical principles such as fairness, privacy, and accountability.

Collaborative efforts with a team from the University of Cambridge guided the ethical dimensions involved. This collaboration leverages a funded project that aims to transform AI use in the public sector by prioritising ethics in digitalisation initiatives for creating public value by working closely with Local Authorities in England. In developing the TraffEase tool, we have drawn on a landscape review of UK government regulations,  policies, guidance, frameworks, bills and regulations conducted as part of the AI@CAM project to inform interviews with participating Local Authorities. These interviews sought to identify ethical risks about the tool, and which ethical considerations from the various UK government documents could shape how the tool is used in the future. Insights from interviews and discussions held with the Local Authorities involved in the TraffEase project have informed the policy recommendations we jointly propose below.

Policy Recommendations for Ethical Deployment of LLM-based Generative AI in the Public Sector

The research outlined above has enabled the joint team to offer a set of recommendations aimed at informing and advising local authorities seeking to use LLMs to improve data access and insights, increasing employee capacity and efficiency.

1. Establish clear governance structures to ensure ethical development and use

Prior to any attempt to develop and use an LLM-based tool in the public sector, it is crucial for public sector bodies (local authorities) to ensure that they have a robust internal governance structure. This is critical to ensure that any AI-based tools that would be developed adheres to established ethical principles, and the structure will provide a stage-gate mechanism to check the alignment of the tool in development and use. 

2. Human-Centred Design

LLM-based generative AI that integrate complex data, should aim to help public sector authorities make informed and citizen-focused decisions. They should be used as a decision-support tool, assisting human decision-makers rather than replacing them.

3. Comprehensive Documentation

Provide detailed user documentation that clearly explains the system’s capabilities, limitations, and intended context of use, helping to prevent accidental misuse.

4. Transparency and Explainability

Prioritise transparency for non-technical users by developing accessible explanations of how the system generates its outputs.

Include reference logs, highlighting where the returned data has come from, to foster trust and enable users to critically evaluate AI-generated insights.

5. Accountability and Validation

Conduct rigorous manual testing to ensure the system is reliable, accurate, and fit for purpose. Testing should focus on identifying bugs, verifying data clarity, and assessing alignment with specific use cases.

6. Active Risk Monitoring

Implement a robust feedback mechanism to allow users to report potential risks or issues. Regularly review this feedback and update a risk register to ensure that concerns are addressed promptly, enabling the system to evolve responsibly.