When AI is the first stop for lease questions

Why a convincing answer can be wrong

Many chatbots are powered by a large language model. During training, it learns patterns from enormous amounts of material, including text. When answering, it generates a response piece by piece using those learned patterns and the context available to it. Further training shapes how it follows instructions and responds.

This supports useful work such as explaining difficult wording and organising information. It also means a fluent response is not proof that the facts have been verified. A model can produce plausible but false statements, often called hallucinations. OpenAI's research identifies training and evaluation practices that reward guessing over admitting uncertainty as one reason these errors persist.

Some tools can search the web or retrieve information from uploaded documents. Those features can give an answer a stronger foundation. They still leave things to check: whether the right document was used, whether an exception was missed and whether the cited passage supports the conclusion. A reference gives the reader somewhere to look; it does not complete that check.

A lease question can go wrong without an invented court case or an obviously absurd claim. The answer may overlook a definition, use another flat's obligations, or treat a new law as already in force. It may also accept an unproved assumption in the question, such as who caused a leak.

A leak question shows the difficulty

"The flat above me has had a flood. Can I claim damage off them directly?"

In a recorded example from the author's lease-analysis work, the case note records that the first answer missed a potentially important covenant, or legal promise. After the user challenged it, the revised response identified wording benefiting owners and lessees of other flats.

That opened a possible route for investigation. It did not establish that the neighbour owed compensation. The full provisions, the relevant leases, the cause of the leak and evidence of any breach still needed checking.

The example involved the author's own specialist lease-analysis tool. The need for scrutiny applies to specialist tools as well as general chatbots. Nor does a changed answer prove that the second attempt is right. The reason for the change has to stand up against the source.

Make the first answer useful

A useful first step is to identify the decision that needs making. In a leak enquiry, stopping the water, arranging repairs, claiming on insurance and recovering uninsured damage are different questions. The Leasehold Advisory Service's guidance explains that responsibility depends on the cause, the lease and insurance terms. Action to limit damage should continue while responsibility is investigated.

A reader can ask the tool to show the exact words and page it relies on, explain them in everyday language and identify anything it cannot establish. The check then means opening the original page and reading the surrounding wording, including definitions, exceptions and referenced schedules, alongside any supplied document that changes the lease. One flat's agreement may differ from another's.

Legal and factual questions need separate checks. For legislation, legislation.gov.uk is a starting point for reading the actual provision and checking where it applies and when it takes effect. If its relevance remains unclear, that becomes a focused question for the Leasehold Advisory Service or a qualified adviser. A lease cannot establish where water came from. That may require photographs, an inspection or a plumber's report. A confident answer cannot fill those gaps.

A useful follow-up question is: "What facts or missing documents could change this answer?" Asking what supports a different interpretation can expose an assumption that otherwise passes unnoticed. These questions help structure a review; asking the chatbot to check itself is not independent verification.

For a reader who cannot confidently assess the wording, the next step can be a focused question rather than a legal demand: "Please show me the provision that makes me responsible for this cost, explain how it applies and identify the evidence supporting the charge."

Free independent help is also available through the Leasehold Advisory Service. Where substantial money, disputed liability or legal deadlines are involved, advice from an appropriate professional may still be needed. AI can help organise the documents and questions for that conversation.

Helping residents before a dispute starts

For a managing agent, the work starts before a reply is written. Document-based AI can help identify who should arrange a repair, what permission may be needed and which provisions need checking before work is instructed. The agent still needs to establish the facts and validate the proposed action.

The next task is explaining that action. A response can show the relevant wording, explain why it applies and distinguish a confirmed responsibility from a point still being investigated. That gives the resident a basis for understanding or questioning the decision.

Managing agents may also choose to help residents obtain access to a suitable tool and ask questions about their own leases. A resident considering an alteration, or trying to understand a charge, could investigate the wording before an avoidable misunderstanding develops. The opportunity is earlier understanding, with the agent available to address what remains unclear.

The next letter affects the relationship

A leaseholder may be frightened by a bill and use AI to sound more confident. A property manager may be trying to answer several households while waiting for an inspection report. A forceful letter generated in seconds can leave the recipient with work to untangle, while making each side feel less heard.

Rejecting a complaint because it appears AI-generated does not resolve the underlying issue. Neither does sending back an equally confident AI-generated reply. Both sides need a way to identify what is established, what remains uncertain and what will happen next.

AI can help with that communication. A useful request is to review whether a draft explains the issue, distinguishes a request from an allegation and gives the recipient a specific question they can answer. The sender still needs to check that it has not added facts, deadlines or promises that nobody agreed.

For people running a building together, a short shared record can make the discussion easier to follow: the problem, the documents checked, the options, the missing information and the person taking the next step. AI can help prepare that record. Residents and their advisers still have to make the decisions, including difficult choices about money and priorities.

The value is in helping people understand the question and work towards an answer they can explain. Before pressing send, the useful test is whether the next person can see the evidence, understand the uncertainty and know what response is needed.

Adam Street is Founder of Building Trust and a resident director. 

Sources

OpenAI, Why language models hallucinate, 5 September 2025: https://openai.com/index/why-language-models-hallucinate/

Leasehold Advisory Service, Water leaks in leasehold flats: https://www.lease-advice.org/building-management/repairs/water-leaks/

Leasehold Advisory Service, Help and advice: https://www.lease-advice.org/

The National Archives, legislation.gov.uk, Understanding legislation: https://www.legislation.gov.uk/understanding-legislation