Whitehall is quietly worried its own staff are using AI badly and losing skills
Key Points
- The government's own civil service AI training programme is built around managing deskilling and security risk.
- Defra has told Parliament it ran exploratory AI initiatives to assist policy and legislation drafters.
- Angela Eagle stressed AI is not directly drafting policy or law and that officials keep full responsibility.
- A government study found AI cut evidence-review time by 23% but still needed manual checks for errors.
- The same skills-erosion worry is surfacing across professions, with 74% of clinicians citing it this week.
When the Open Innovation Team set out the case for its civil service AI training programme, it did something the rest of Whitehall’s enthusiasm for AI rarely does – it listed the downsides.
Tucked inside the Cabinet Office unit’s own slidepack for the AI Adoption Accelerator, alongside the promise of faster and higher quality work, sit four blunt warnings about what happens when officials reach for AI without support:
- Outputs poor enough to need redoing;
- Security and compliance risks;
- Frustration when the tools underdeliver;
- Concerns about staff losing core skills.
The programme itself is pitched as the remedy, offering workshops, hands-on coaching and bootcamps to train internal champions. But read it the other way and the document doubles as an admission.
The government is deploying AI directly into the working lives of civil servants faster than it is teaching them to use it well, and it knows the gap carries a cost it would rather frame as an opportunity.
The productivity story
Ministers have sold AI to their own workforce almost entirely as a productivity dividend.
The Humphrey suite of tools, a target to move a tenth of officials into digital and data roles, and a stream of departmental pilots have all been presented as a way to do more with a shrinking, squeezed civil service.
The Institute for Government’s Whitehall Monitor 2026 sets out the backdrop of a workforce under sustained pay pressure, grade inflation, and a leadership betting heavily on technology to close delivery gaps.
What rarely makes it into announcements such as this is how far the tools have already crept toward the core craft of government.
In a written parliamentary answer published on 24 April, Minister Samantha Dixon revealed that the department has rolled out Microsoft Copilot across the organisation to boost efficiency for all staff, including policy professionals and legislation teams.
“Copilot is used by staff across the department including legislation and policy teams.
“The department is clear that document authors are fully accountable for policy judgement and conclusions and compliance with departmental, legal, and information governance, irrespective of whether or not AI has been used to assist with drafting,” she said.
MP Sir John Hayes put the same question to departments across Whitehall, and the answers, compiled in a Commons Library briefing, sketch a government experimenting widely while the rules of safe use are still being written.
The evidence is thinner than the enthusiasm
The case for moving fast rests on numbers that deserve more scrutiny than they get. A cross-departmental study found AI-assisted evidence reviews took 23% less time, finishing in 90.5 hours against 117.75 for a human-only team.
Impressive, until you reach the caveat in the government’s own findings – the work still had to be manually checked for hallucinations and errors.
A single study with that massive caveat attached is now helping justify a rollout that will reshape how thousands of officials work, and what skills they are expected to keep sharp.
The worry is not confined to Whitehall. A health survey published this week found 74% of clinicians fear deskilling, defined as leaning on AI so heavily that the ability to spot a bad recommendation erodes.
It is the same anxiety the Open Innovation Team named in its own training deck, surfacing in every profession where AI meets skilled judgement.
