Two UK ministries are deploying a Google DeepMind system to triage routine home extensions and loft conversions. The rules governing who is responsible when the algorithm gets a decision wrong have not been written.
When a homeowner in England files a planning application for a loft conversion or a rear extension this year, the document is likely to be read first by an AI system built by Google DeepMind rather than by a human planning officer. The shift, announced by two UK central government departments in partnership with Google Cloud, turns a routine bureaucratic task into a quiet experiment in machine-made administrative decisions. The rules for who is responsible when the system gets it wrong have not been written.
The departments involved are the Ministry of Housing, Communities and Local Government (MHCLG) and the Department for Science, Innovation and Technology (DSIT). Together they have expanded two machine learning tools aimed at accelerating municipal planning workflows, according to a report on the rollout in Artificial Intelligence News. One, called Extract, is in deployment and is being used to read and structure planning documents. The other, an Augmented Planning Decisions (APD) prototype, is being tested with councils and is described as a tool to help officers weigh applications against local plan policy. Both are framed as ways to relieve planning officers of repetitive paperwork, but the wider question is what happens to the residents whose submissions are now summarized and triaged by software before a human ever sees them.
The stakes are larger than the tools themselves. Householder applications, the routine submissions covering extensions, loft conversions, garage conversions and similar small works, make up roughly 70% of all planning applications filed in the UK. The volume of these submissions, alongside the dense paperwork that surrounds them, is one of the main reasons local planning authorities carry backlogs that delay larger housing developments. If the front of that pipeline can be cleared faster, ministers argue, the bigger pipeline behind it can move as well. The UK central government has set a stated target of 1.5 million new homes by 2029, and any tool that can shorten decision times is being marketed as a contribution to that goal.
Google is presenting a specific number. Lila Ibrahim, Chief AI Readiness Officer at Google DeepMind, told the Google Cloud Summit London that the system can cut planning decision times by 50%. That figure is a vendor claim, not an independent finding. It has not, on the public record, been audited against a control group of councils still using conventional workflows, and the methodology behind it has not been published. Any reader weighing the impact should treat the number as the company's own benchmark rather than a government measure of effectiveness.
What changes in practice is the role of the planning officer. Under the new arrangement, a large language model first reads the application, pulls out the relevant facts, flags potential issues against local plan policies, and produces a structured summary for the officer. The officer's job shifts from being the first evaluator of the document to being a reviewer of the machine's evaluation, then a final arbiter on edge cases and appeals. Proponents frame this as a productivity gain. Critics frame this as a quiet delegation of discretion, in which the algorithm sets the agenda of what an officer looks at and what they do not.
The accountability question is the part that has not been designed in public. If a planning officer rejects an application today, the applicant has a clear route of appeal and a clear sense of who made the call. If an LLM triages the application, flags it as routine, and routes it for fast approval, and the approval later turns out to rest on a misread plan policy, the chain of responsibility is harder to describe. Was the mistake Google's, the council's, or the ministry's? Is the model a tool, an adviser, or a co-decision-maker? None of those questions has an answer in current UK planning law, and the departments involved have not published guidance on how the new layer fits with existing appeal rights or with the duty to record reasons for decisions.
There is also a thinner question about the model itself. The available reporting does not specify which Google model lineage is being used, whether it is a Gemini variant, a Vertex-deployed system, or something from DeepMind's research stack, and the councils piloting the system are not named. A homeowner whose application is rejected on the basis of a model's summary has, today, no obvious way to ask what the model was, how it was trained, or what data about them and their street was used to score their submission. Transparency expectations that the UK has applied to other algorithmic tools in the public sector, including the Algorithmic Transparency Recording Standard, do not yet appear to have been mapped onto planning AI in any public document.
What to watch next is concrete. The first test is whether MHCLG and DSIT publish deployment scope, named councils, named models, and a methodology behind the 50% claim before the tools are described as nationwide. The second is whether the Planning Inspectorate, which hears appeals, has been briefed on the role of the new layer in officer decisions, and whether appellants will be told when a model has been involved in the case against them. The third is whether any council using Extract or piloting APD will publish its own error rate, including rejections later overturned, against a comparable council that has not adopted the system. Until those three things happen, the headline number is the only number anyone outside Google has, and it is the company's own.