As artificial intelligence (AI) moves from experimentation into everyday public service delivery, government organisations face a growing challenge: how can they demonstrate that their AI systems are safe, accountable and worthy of public trust?
The answer requires more than publishing ethical principles or introducing another set of guidelines, according to speakers at a Think Digital Partners webinar, Using AI Responsibly in the Public Sector.
Bringing together representatives from the Department for Environment, Food and Rural Affairs (Defra), the Government Digital Service (GDS) and SAS, the discussion explored how public bodies can translate responsible AI ambitions into practical safeguards.
James Blackshaw, AI strategy lead at Defra, said responsible AI encompasses everything from accountability and transparency to sustainability, resilience and privacy.
“What these all have in common is the idea of minimising the harm from AI and the potential for harm from AI,” he said.
But with generative and agentic AI introducing new risks, organisations must continually adapt their approaches.
Here are five key takeaways for public sector leaders.
1. Start by asking whether AI is actually the right solution
Before procuring or developing an AI system, establish whether the proposed use case is lawful, whether it will improve public outcomes and whether AI is the appropriate technology.
Antti Heino, principal advisor, AI and analytics at SAS, argued that understanding what an organisation is buying is fundamental to responsible procurement.
“Is the use case lawful? Does it improve public outcomes?” he asked.
Even where the answer is yes, organisations must consider whether AI is necessary before progressing to questions about accountability, transparency and safeguards.
Rhian Jones, head of data policy, ethics and strategy at GDS, said organisations must also understand the specific context in which AI will operate.
“Is the AI use safe in this context? Who can be negatively impacted? Do the risks outweigh the benefits? And could we adopt a fairer, more sustainable, more transparent approach?”
These questions become particularly important as AI moves beyond generating content to taking autonomous actions.
Blackshaw warned that agentic systems may have access to sensitive internal infrastructure, placing greater emphasis on reliability, transparency and security.
2. Make procurement professionals part of your AI governance strategy
Responsible AI isn’t solely the responsibility of technology teams. Procurement professionals also have an important role in ensuring suppliers meet government’s expectations.
Blackshaw described procurement as a powerful mechanism for embedding responsible AI, particularly given the public sector’s reliance on privately developed technology.
He pointed to work at Defra on government buying standards, including efforts to embed sustainability requirements relating to how AI models are developed and where they are hosted.
“We need to include our procurement professionals in our overall responsible AI strategy,” he said.
That means ensuring procurement teams understand the risks, know which questions to ask and are prepared to challenge suppliers rather than accepting inadequate explanations.
Jones agreed that transparency, accountability, fairness, safety and public trust must be considered from the outset of procurement.
Crucially, she said, responsible AI principles apply equally to systems developed internally and those purchased from external providers.
3. Know what AI is being used across your organisation
How can public bodies govern AI effectively if they do not have a clear picture of where it is already being used?
Heino recommended establishing centralised inventories of AI models and use cases, giving organisations visibility across departments and teams.
“Having a centralised visibility on it reduces risk of duplicated efforts, but also gives leaders […] a clear view into what kind of risks they should be managing,” he said.
If you liked this content…
This would help identify duplicated work, clarify which risks need managing and enable governance processes to be tailored to individual applications.
He argued that organisations should combine these inventories with practical checklists and risk-based controls, rather than imposing the same governance requirements on every AI project.
The approach could also help organisations reuse existing models and capabilities instead of developing similar solutions independently.
Heino also recommended developing a synthetic data strategy to support AI testing and development while potentially reducing exposure of sensitive personal information.
4. Give people practical guidance – and invite sceptics into the conversation
Government has established a growing body of AI guidance, but translating it into everyday working practices remains a challenge.
Jones outlined how GDS is combining mandatory standards, guidance and capability building to support responsible adoption.
She said 150 transparency records had been published on GOV.UK at the time of the webinar, as GDS implements the mandatory rollout of the Algorithmic Transparency Recording Standard across government departments and arm’s-length bodies.
GDS also has a data and AI ethics community of practice with 500 members across government and the wider public sector.
However, Blackshaw argued that guidance cannot be developed by technical specialists alone.
“The person who builds an AI system is likely not going to be the best person to judge what the responsible use of that system looks like,” he said.
He recommended involving people who are sceptical about AI, arguing that their challenges can expose weaknesses in proposed frameworks and safeguards.
Defra has sought to make its gen AI guidance more accessible through flowcharts, interactive materials and a set of 12 principles that employees can follow.
Blackshaw also emphasised the importance of giving employees time to experiment with general-purpose AI tools, supported by resources such as prompt libraries and examples of successful applications.
5. Keep testing AI systems – and make someone accountable
One important question came from the audience: does government have sufficient assurance that it is using AI responsibly before asking the public to trust AI-enabled services?
Blackshaw argued that organisations need clear ownership of AI systems, extending from individual services through to senior leadership.
“We can’t simply rely on AI to make these decisions and assume that they’re accurate. We need to take ownership of what those outputs are,” he said.
That requires clearly defined responsibilities, escalation routes and continuous monitoring of accuracy, model drift and environmental impacts.
He also stressed the importance of understanding people’s concerns and experiences, rather than relying exclusively on technical measurements.
“We need to red team relentlessly,” he added, calling for organisations to continually test how AI systems could fail as their capabilities evolve.
Jones said transparency is equally important. The Algorithmic Transparency Recording Standard enables public bodies to explain how algorithmic tools are designed, governed and used in decision-making.
Those records can also support collaboration, allowing organisations to learn from deployments elsewhere in government.
For Heino, this combination of governance, transparency and reusable approaches is essential if public bodies are to move beyond isolated AI experiments.
“Trust is what enables these innovations to scale,” he concluded.
Watch the full webinar here.
