The myth of how policy is made
There is a familiar myth about government: that policy is shaped in quiet offices by officials poring over papers, weighing evidence and advising ministers with care. That image is already out of date.
Across Whitehall and beyond, artificial intelligence has entered the policy process – not as a distant future prospect, but as a practical tool reshaping how decisions are made. The question is no longer whether AI belongs in government. It is already there. The real question is whether it is being used wisely.
Faster government, but not necessarily better
The appeal is obvious. Policy-making runs on information: consultation responses, impact assessments, research papers, stakeholder submissions. Increasingly, AI is being used to process that material at speed. Tools developed inside government can now summarise lengthy documents, group thousands of public consultation responses and even help draft policy submissions in a fraction of the time it would take a human team.
For overstretched departments, the efficiency gains are hard to ignore. Generative AI promises to cut through administrative backlog, freeing civil servants to focus on strategy rather than paperwork. In theory, it could produce a more responsive state – one that listens faster and reacts sooner.
But speed is not the same as wisdom.
From judgement to data-driven policy
The growing role of AI in policy-making reflects a broader shift: from government guided by principle to government shaped by data. Increasingly, policymakers are using AI not just to summarise information, but to analyse it – identifying patterns, predicting outcomes and evaluating what works. The OECD has pointed to AI’s potential to transform policy evaluation, allowing governments to analyse vast datasets in real time and test the likely impact of policies before they are even implemented.
This is a profound change. Policy has traditionally been slow, iterative and often reactive. AI raises the possibility of something different: a more dynamic system where policies are continuously assessed and adjusted as new evidence emerges. Done well, this could lead to smarter decisions and better outcomes.
But there is a danger in mistaking technological capability for political judgement.
The risks hidden in the data
AI systems are only as good as the data they are trained on, and public sector data is often patchy, outdated or incomplete. Feed flawed data into an algorithm and it will produce flawed conclusions – just faster and with greater confidence. Worse still, those conclusions may appear objective, masking the biases embedded within them.
There is also the problem of “automation bias”: the tendency for humans to trust machine-generated outputs without sufficient scrutiny. In a policy context, this is particularly risky. Decisions about welfare, policing or healthcare cannot be reduced to pattern recognition alone. They involve trade-offs, values and human consequences that no model can fully capture.
Even the most advanced systems struggle with nuance. They can summarise what has been written, but they are less able to grasp what is not – the informal dynamics of institutions, the political realities behind decisions, the lived experience of those affected by policy. These are precisely the elements that often determine whether a policy succeeds or fails.
Regulation is catching up
And yet, for all these limitations, the direction of travel is clear. Governments are not pulling back from AI; they are doubling down. Across Europe, regulators are moving towards stricter oversight, with measures such as mandatory labelling of AI-generated content and greater transparency about how systems are used. The emphasis is shifting from abstract ethical principles to enforceable rules, particularly through frameworks such as the EU AI Act.
At the same time, there is a growing push for “AI literacy” within government. Civil servants are being asked not just to use these tools, but to understand their risks and limitations. The role of the policy-maker is changing: less time spent gathering information, more time interpreting, challenging and contextualising it.
A changing role for the policy-maker
This may ultimately be AI’s most significant impact. By automating the routine work of policy, it forces a rethink of what human expertise is for. If a machine can draft a briefing note in seconds, the value of the policy-maker lies not in writing it, but in knowing whether it is right.
That requires judgement – and judgement requires experience.
Here, there is a paradox. The very tasks AI excels at – synthesising evidence, producing first drafts – are the ones junior officials have traditionally used to learn their craft. If those tasks disappear, how will the next generation of policymakers develop the expertise needed to challenge the machine?
There are no easy answers. Some argue for preserving certain tasks as a form of training. Others suggest a more radical rethink, with greater emphasis on frontline experience or mentorship. What is clear is that the adoption of AI is not just a technical shift, but an institutional one.
The illusion of objectivity
For all the talk of innovation, the fundamentals of policy-making remain unchanged. Governments still need to decide whose interests to prioritise, which risks to take and what trade-offs to accept. AI can inform those choices, but it cannot make them.
Nor should it.
The danger is not that AI will take over policy-making, but that it will subtly reshape it – privileging what can be measured over what matters, speeding up decisions without improving them, and creating an illusion of objectivity where judgement is still required.
Conclusion: who is really in charge?
Used carefully, AI could help build a more capable state: one that is better informed, more efficient and more responsive to citizens. Used carelessly, it risks entrenching bias, weakening accountability and obscuring the human responsibility at the heart of government.
The algorithm is already in the room. The challenge now is ensuring it does not end up in charge.