We Went Looking for the Weak Arguments Against AI in Planning

We spent this month doing something deliberately uncomfortable. Rather than argue for AI in planning, we went looking for every argument against it, across England, and gathered them in one place.

The method was simple and the numbers are the point. We attempted 164 sources across government and regulators, professional bodies and the private sector, and public, community and media discourse. 141 returned usable content. From those we captured 111 individual records and deduplicated them by claim, not by link, so that the same objection voiced by five people counts once and keeps all five sources.

That left 59 distinct objections.

We expected a proportion of them to be misinformed. Something we could set aside politely and move on from.

Some are predictions rather than evidence. Many are aimed at a different part of the planning system than the one we work in. Several are uncomfortable for anyone selling AI-assisted analysis, including us. But there is no pile of nonsense here, and anyone planning to defend AI in this sector by dismissing the critics as luddites is going to have a bad time.

Ask a room of developers what could go wrong with AI in planning and they will talk about machines making decisions. That is not what the sector is actually shouting about.

The largest single cluster in everything we gathered concerns AI used to oppose development. Residents generating objections at industrial volume.

The examples are specific. One housing application in the Midlands drew more than 500 objection emails from a single resident, including a 160-page report. A Staffordshire resident sent 250 separate emails on one application, individually critiquing each submission document. More than five thousand near-identical objections were submitted against one scheme, many raising arguments with no connection to the site. Commercial services will now generate a planning objection for a fee of around £45.

What this means before you spend. If you are promoting a site, the volume of opposition you see is no longer a reliable signal of the strength of local feeling, and it is no longer a reliable signal of anything at all. Read the substance of the representations, not the count. Councils are already triaging on exactly that basis, and the professional guidance now tells objection platforms not to enable automated mass submission.

This is the item that should give every vendor in this space pause, including us.

The government has been building a tool to digitise the paper records that councils hold, so that constraints like conservation areas, Article 4 Directions and tree preservation orders become properly available as data. It is genuinely useful work, and councils are now publishing its output to the national planning data platform.

Its own published evaluation reports that the tool maps conservation area boundaries correctly 81 to 89 per cent of the time. For Article 4 Direction boundaries, the figure is 40 to 58 per cent. The evaluation concluded the tool was not suitable for use without a human check, and separately recorded that running the same document through it three times could produce three different answers.

That is the government publishing an honest accuracy figure on its own work, which is more than most of the private sector does. It is also a genuine problem for anyone consuming that data downstream.

What this means before you spend. An Article 4 Direction removes permitted development rights. If a tool tells you a site's permitted development rights are intact, the useful question is not how confident the tool sounds. It is where the Article 4 layer came from, when it was captured, and whether anyone has checked it against the authority's own published record. That is a question worth asking of any provider, ourselves included.

Most commentary on AI and fabricated citations is speculative. Some of it is not.

In June 2025 the Divisional Court handed down judgment in Ayinde, R (on the application of) v London Borough of Haringey [2025] EWHC 1383 (Admin). Grounds filed in a judicial review cited five cases. They did not exist. One, cited as El Gendi, is held by no court record anywhere. The court found that providing a fake description of five fake cases qualifies clearly as professional misconduct.

Alongside that, the Planning Inspectorate's own guidance states that AI is capable of creating fake representations or evidence, and that improper undisclosed use in an appeal risks an award of costs for unreasonable behaviour. An Inspector has already found, in a costs decision, serious concerns that a third party's Statement of Case had been produced using AI without being declared.

What this means before you spend. Two things follow, and they are practical rather than philosophical. First, if you submit AI-assisted material into an application or an appeal, declare it. The disclosure costs you nothing and its absence is what attracts the costs risk. Second, one fabricated reference destroys the credibility of an entire document, including the parts that were sound. Planning officers we found describing this were unanimous on the point. It is not a proportionate penalty, but it is the real one.

We went looking specifically for what the community sector has said about AI in planning. CPRE, Civic Voice and the National Association of Local Councils were all searched directly.

We found nothing. Not a retrieval failure, and we checked: the pages loaded and contained no published position.

That is worth recording precisely because it will not last. The professional bodies have moved, the Inspectorate has moved, and the trade press has been covering this for a year. The organisations that speak for communities have not yet arrived at a position, and when they do it will be shaped by whatever the sector has done in the meantime.

We have not solved this. This article maps the objections; it does not answer them, and we would rather say so than imply otherwise.

We ran all 59 objections at our own work as well, which produced a list of things we need to fix. That is an internal exercise for now, and if it turns into something worth reading, it will appear here when the fixes have actually landed rather than while they are still intentions.

Given the subject, it would be poor form not to say so. This was researched and drafted with AI assistance, then checked against source. Every case citation in it was verified against the official court record before publication, and one case we found referenced elsewhere was left out because we could not verify it on an official source. The accuracy figures quoted are from the government's own published evaluation, not our estimate of it. That is the standard we would want applied to anything we were asked to rely on.

The wider 2026 policy picture is covered in planning reform in 2026.