Guide

Your AI does not know Section 21 is gone

Guillaume Black Chief Executive, Property Filter 29 July 2026

The example that should end the argument

Until 30 April 2026 a landlord in England could serve a Section 21 notice and recover a property without giving a reason. On 1 May 2026 that ended. Section 21 is gone, existing and new tenancies converted to assured periodic tenancies, fixed-term assured shorthold tenancies were removed, and every possession claim now has to run through Section 8 with evidenced grounds. Serving a Section 21 after abolition can attract a civil penalty from the local authority of up to £7,000 for a single breach.

Now consider what an AI assistant trained on the last fifteen years of British property writing has absorbed. Section 21 appears in tens of thousands of articles, forum posts, template letters and landlord guides. Section 21 abolition, as settled law rather than a pending bill, appears in a comparatively tiny amount of very recent text. Ask a general model about ending a tenancy and the weight of evidence points squarely at the wrong answer.

The failure is not that the model is stupid. It is that the model cannot tell you what it does not know. It has no reliable sense of where its own knowledge stops, so it will not say "check whether this changed in 2026". It will simply answer, well, in complete sentences, with the notice periods and the form number.

This has already gone wrong in public

Through 2025 and into 2026, the First-tier Tribunal Property Chamber and the Upper Tribunal saw a run of cases in which parties put forward case law that either did not exist or did not say what it was claimed to say. The housing law specialist Giles Peaker collected them, case by case, in a January 2026 review on Nearly Legal. It is worth reading in full, because the pattern is more instructive than any warning.

In one case, a landlord respondent cited ten authorities that were false or simply irrelevant, including one concerning Parole Board hearings. In another, leaseholders representing themselves cited three cases that did not exist, and the tribunal made a costs award against them. In a third, a judge tested an AI assistant during the proceedings and demonstrated that it produced confident, inconsistent and false answers to the same question. Elsewhere a party was debarred after relying on an authority that turned out not to exist, and in another matter counsel accepted that an AI had put words into a real judgment that were never in it.

Two things stand out. The consequences escalated over time, from submissions being disregarded, to costs awards, to debarment. And judges have started investigating AI use directly rather than waiting for someone to own up.

At a tribunal, ignorance is not an excuse. The document has your name on it.

The consistent theme across the reported cases

Note who was getting caught. Not only litigants in person. Represented parties and professionals too. Being sophisticated is not the protection here. Checking is.

Why UK property is unusually bad ground for a general AI

Three problems stack on top of each other, and each one on its own would be manageable.

The default is American

An enormous share of everything ever written online about property investing is American. So the statistical centre of gravity is American, and a general model drifts there unless you hold it in place. Ask about depreciation, closing costs, escrow or a 1031 exchange and you will get a fluent answer that has no bearing on a UK deal. The dangerous version is subtler: an answer that is structurally American but wears British vocabulary.

The rules change often, and they are not national

Property regulation in the UK moves several times a year, differs between England, Wales, Scotland and Northern Ireland, and in places comes down to your individual council. HMO licensing is the clearest case: whether you need a licence, and which of the three regimes applies, depends on the local authority and can change with a single cabinet decision. An answer that is correct nationally can be wrong in your street.

A model does not know where its knowledge ends

This is the one people underestimate. A model will not reliably flag "this may have changed". It cannot see the boundary of its own training. So the failure mode is not a refusal or a hedge, it is a confident, detailed, obsolete answer, which is the hardest kind to catch.

The UK rules most likely to come back wrong

These are the ones where a wrong answer changes the numbers or the legality of what you are about to do. Treat any AI answer touching them as a draft to verify, never as the position.

The three questions that catch it

You do not need to become a lawyer to use AI safely on this. You need a habit. Three questions, every time money or legality is involved.

1. "What is your source for that, and how current is it?"

Ask for the source of every specific claim: the statute, the HMRC page, the council register. Then open it. Two useful things happen. Sometimes the source does not exist, which tells you everything. More often the source exists but says something narrower than the summary you were given, and the gap between the two is exactly where the risk lives.

2. "What would make this answer wrong?"

This question is unreasonably effective. A model asked to argue against itself will often surface the condition it quietly assumed, that you are an individual rather than a company, that the property is in England, that the rule has not changed. Those assumptions are where a confident answer goes wrong, and they usually stay invisible until you ask.

3. "Am I asking it to remember, or to reason?"

This is the real dividing line. AI is genuinely good at reasoning over material you give it: read this lease and tell me what the ground rent review clause does, compare these two offers, draft this letter, summarise these forty pages and tell me what a buyer should worry about. It is unreliable when you ask it to recall a rule from memory. Same tool, two completely different levels of risk. Give it the current rule and it becomes useful. Ask it to remember the rule and you are gambling.

What good practice actually looks like

The investors who get real leverage out of AI on legal and tax matters all end up in roughly the same place.

• They supply the rules instead of trusting recall. The current HMRC page, the council's licensing position, the actual lease. Give the model the ground truth and it stops guessing.
• They use it to find the question, not to settle it. Its best output is often "here are the six things that could bite you in this deal", which is a brilliant agenda for a fifteen minute call with a solicitor and a terrible substitute for one.
• They keep a paper trail. If a decision was based on an AI answer, the source sits next to it. That is what turns an uncomfortable question later into a two minute answer.
• They never let it be the last word on anything with a penalty attached. Possession, licensing, tax filing, deposit handling. A £7,000 civil penalty is a great deal more expensive than an hour of professional time.

None of that makes AI less useful. It makes it useful on the things it is actually good at, which is most of the work, rather than on the one thing it is worst at.

Frequently asked questions