Notes on people, technology and the systems shaping everyday life.
AI is moving deeper into the machinery of UK government: helping civil servants analyse information, draft documents, process casework and respond to public demand.
AI is beginning to settle into the everyday machinery of government.
Some of its uses are fairly ordinary. It can summarise a meeting, sort through consultation responses or help produce a first draft. Other uses are moving closer to the decisions and services that shape people’s lives.
There is plenty of excitement about the time this could save. There are also some less obvious changes happening around it.
AI is changing how much work people can produce, how quickly information moves and what civil servants are expected to check. It is changing how citizens communicate with public bodies too.
So perhaps the more interesting question is not whether government will adopt AI. It is whether the rest of government can adapt to what follows.
A few years ago, government AI was mostly discussed through pilots and future plans. That is becoming harder to say now.
In 2024, the National Audit Office found that 70% of the government bodies it surveyed were piloting or planning uses of AI. Thirty-seven per cent had already deployed it in at least one area.
Since then, government has developed tools for consultation analysis, policy work, document processing and administration. The Civil Service has also been trying to move people from basic AI literacy towards practical use. Its 2025 One Big Thing programme involved 221,000 civil servants, and the 2026 programme is focused on “AI in Action.”
There are good reasons for this interest. Government contains an enormous amount of repetitive and information-heavy work. If AI can reduce some of that, civil servants may have more time for the work that needs experience, care and judgement.
But introducing AI does not simply remove work. It changes its shape.
A first draft may appear faster, but someone still has to decide whether it is right. A case may move more quickly, but the consequences of an error remain. A public body may process more information, while the person responsible is left with more output to understand.
That is the point at which adoption becomes more than a technology story.
The further AI moves along this line, the more important the surrounding institution becomes.
One of the more unexpected signs of change is appearing in complaint queues.
Schools, councils and other public bodies have reported receiving more complaints written with the help of generative AI. The problem is not simply that people are using AI. In many cases, that may be completely reasonable.
Formal complaints can be difficult to write. People may not know the right language, understand the process or feel confident that a short explanation will be taken seriously. AI can help someone express a real problem in a form that feels more official.
The same tools can also turn one grievance into several pages of polished correspondence. They can make repeated submissions easier and produce more material than already stretched teams can read carefully.
This creates a difficult tension.
AI may help more people speak in the language institutions expect. At the same time, the resulting volume may make it harder for those institutions to hear the individual problem underneath.
Filtering out anything that looks AI-written would be a crude response. Writing style tells us very little about whether a complaint is true, urgent or important.
The more useful question is how public bodies can find the human issue beneath the words without penalising the person for using help to express it.
The Civil Service is investing in AI training, which makes sense. People need to understand the tools entering their work.
Still, there is a large difference between knowing how to use an AI system and knowing when its output should not be trusted.
A caseworker may receive a summary that sounds convincing but misses a detail that changes the case. A policy official may be given a neat synthesis that leaves out an inconvenient piece of evidence. A complaints officer may need to separate formal language from the substance supporting it.
These are not simply prompting skills. They involve professional judgement, context and responsibility.
AI may reduce the effort required to produce something. It can shift that effort to the person who must check it.
That person needs enough time to notice a problem, enough information to understand it and enough authority to do something about it. Without those conditions, keeping a person “in the loop” may not mean very much.
AI tools tend to look most impressive in clean examples.
Public services are rarely clean examples.
A housing case may involve missing documents, conflicting accounts, changing policy and a person in an unusual situation. Social care, benefits and other public services carry their own legal duties and consequences.
This is why a general performance score can only tell us so much.
Before an AI system enters a real government workflow, someone needs to ask more specific questions. What information can it reach? What is it allowed to do? Which policy applies? What happens when the situation falls outside the examples it has seen? Who notices a failure? Can the action be stopped?
The answers may be different in every service.
An AI system does not only need to work. It needs to work within the rules, pressures and human realities of the environment where it has been placed.
The Algorithmic Transparency Recording Standard is now mandatory for central government departments and relevant arm’s-length bodies.
Making government algorithms more visible matters. It gives the public, researchers and oversight bodies a better chance of seeing where these systems are being used.
But disclosure is not the same as recourse.
A person may be able to find a record explaining that an organisation uses an algorithm and still have no idea whether it influenced their own case. Even if they know it did, they may not know what information shaped the outcome, who remains responsible or how to ask for the decision to be reconsidered.
“Transparency tells us that a system exists. Contestability asks whether the person affected has any meaningful way to respond.”
Complaint queues, workforce judgement, testing and transparency can look like separate problems.
They may actually be different parts of the same adjustment.
AI allows more material to be produced. It allows information to move faster. It may soon allow parts of a workflow to operate with much less direct involvement from a person.
Government then has to absorb that change.
Its staff need different kinds of support. Its assurance processes need to reflect the specific work being performed. Citizens need to understand when AI has touched their case. Someone must still be able to explain, interrupt or reverse what the system has done.
The technology is changing quickly. The institution around it has more to catch up with.
“AI is becoming more capable. The systems around it are not yet making that capability fit for people.”
The most visible response to a new technological capability is often another piece of technology.
Yet many of the gaps appearing around government AI do not seem to require a more intelligent model. They appear where the model meets a person, a professional judgement or an institutional rule.
If AI enables citizens to produce more correspondence, the system around it should help staff identify the underlying issue without treating AI assistance as a reason to dismiss it.
If AI helps civil servants produce work faster, the interface should make uncertainty, missing evidence and the need for human review easier to see.
If an AI system begins handling more complicated workflows, its permissions and limits should reflect the particular service, policy and consequences involved.
If AI makes a decision process harder to follow, the person affected should still be able to understand what happened and reach someone capable of reconsidering it.
The gap may therefore sit between intelligence and use.
It is the layer that translates human needs, professional judgement and institutional rules into boundaries the technology can follow.
This includes the interfaces, permissions, explanations, escalation routes and safeguards that allow AI to adapt to the environment it enters, rather than expecting people to adapt themselves around the model.
People should not need to understand the technical architecture of an AI system to question how government has treated them.
They should be told when AI has played a meaningful role in their case. They should be able to understand the explanation they receive and reach someone who has the power to reconsider the outcome.
A second automated answer is not the same as human recourse.
Government organisations may need to look beyond the easiest productivity measures.
Time saved matters. So do the errors that need correcting, the work transferred to another team and the experience of the person using the service.
A faster process is not necessarily a better public service. The result still has to be fair, understandable and workable for the people delivering it.
For those building government AI, control cannot be something added after the intelligent part is finished.
Permissions, policy, escalation and traceability need to shape the system from the beginning.
This is one of the questions we have been exploring through SwarmSight, an open-source Thirty Labs experiment for governing AI agents in public-sector workflows.
Its architecture separates intelligence from authority. An agent can prepare and propose an action, but a separate layer decides what information it can access, which rules apply and whether the proposed action is allowed. Those decisions are recorded so they can be inspected later.
SwarmSight is not intended as an answer to every problem in government AI. It is one attempt to explore a narrower idea:
“What would an AI system look like if control were part of its architecture rather than a promise surrounding it?”
Explore the SwarmSight repository →
The next phase of government AI will probably be less visible than the first.
Announcements tell us what has launched. The more important changes may appear later, in what systems are permitted to do, where the work goes and what happens when a person needs to question the result.
Here are five developments we will be following.
Many government AI tools currently help someone analyse, summarise or prepare work. A more significant change will come when systems begin moving cases, initiating actions or coordinating complete workflows.
At that point, it will be worth looking closely at where permission sits. Can the system act without approval? Can someone stop it halfway through? Is there a named person responsible for what happens next?
A productivity trial can tell us how quickly AI produced a first draft. It may reveal much less about the person who checked it, the team that corrected it or the additional demand created by faster production.
Some work disappears. Some work moves.
We will be looking for evaluations that follow the work through the whole service rather than stopping when the AI completes its task.
Citizens are beginning to use AI to write complaints and navigate complex government processes. Public bodies may increasingly use AI to organise and answer those submissions.
That raises a strange possibility. A citizen’s AI writes to a government system, which uses another AI to interpret and respond to it.
If this becomes common, the challenge will be keeping the actual person and their problem visible inside the exchange.
More public records about government algorithms would be valuable. The next question is what someone can do with that information.
Can they tell whether AI influenced their own case? Can they understand how? Can they reach someone capable of reviewing it?
We will be looking for signs that transparency is becoming a practical route to explanation, challenge and reconsideration.
Many arguments for government AI begin with time and money. Those are important, particularly in services working under pressure.
They are not the whole purpose of public service.
A system might complete a task faster while making the decision harder to understand. It might reduce one team’s workload while increasing another’s. It might improve the average experience while working poorly for people whose circumstances do not fit the usual pattern.
The most revealing question may be what government eventually chooses to measure.
Will success mean tasks completed and hours saved, or will it also include accessibility, fairness, public experience and the quality of the decisions being made?
The starting points for this issue were surfaced using SignalFoundry, a Thirty Labs research tool for finding recurring patterns, disagreements and unanswered questions across different sources.
SignalFoundry helped identify the initial cases and connections. The claims used in the published issue were then checked against their original sources.