AI-Assisted Legal Research: Verification, Citation and Hallucination Risk

In December 2023, the First-tier Tribunal decided Harber v Commissioners for HMRC [2023] UKFTT 1007 (TC) — a UK tax case that became a widely cited cautionary tale after the appellant, a litigant in person, relied on case law summaries that turned out to have been generated by an AI tool and did not correspond to any real decisions. The tribunal noted the cases could not be found on any legal database because they did not exist. It is a useful reminder for legal professionals too: the risk is not that AI research tools are useless, it is that their errors are confident and easy to miss.

Why AI research tools hallucinate

Generative AI tools built on large language models are designed to produce plausible, fluent text based on patterns in their training data — not to retrieve and verify facts from a database of confirmed truth. When asked for a case supporting a particular proposition, a tool can generate something that reads exactly like a real citation — a sensible-sounding case name, a plausible court, a plausible year — without that citation corresponding to any actual decision. This is often called "hallucination," and it is a structural feature of how these tools work, not an occasional bug that improves versions will simply fix away. Some AI tools marketed specifically for legal research reduce this risk through retrieval from verified legal databases, but even those tools are not infallible, and the underlying discipline of verification still applies.

A practical verification workflow

The answer is not to avoid AI-assisted research — used properly it can genuinely speed up the early stages of a research task. The answer is to treat its output as a draft starting point that must be checked, never as a finished, citable source.

• Treat AI output as a first draft, not a conclusion. Use it to identify possible lines of argument or areas to research further, not as the final word on what the law says.
• Always pull and read the primary source. If a case, statute or regulation is cited, locate and read it yourself before relying on it. A summary of a case is not a substitute for the case.
• Cross-check every citation in an authoritative legal database (such as your firm's usual subscription research platform or an official source like BAILII or legislation.gov.uk) rather than accepting the AI tool's own description of what a case decided.
• Check that the case actually says what the AI claims it says. A citation can be real and still be misused — AI tools can accurately name a genuine case while misdescribing its holding or applying it to facts it does not actually support.
• Keep a record of what was AI-assisted. Noting which parts of a research task used AI tools, and what verification was done, gives supervisors something concrete to check and protects the fee earner if a question is later raised about the work.

Why this matters beyond the individual matter

An unverified AI-generated citation that reaches a court filing, a piece of client advice, or even an internal memo relied on by a colleague can cause real damage — to the client's position, to the firm's credibility, and to the individual solicitor's professional standing. The SRA's August 2026 warning notice on AI misuse specifically flagged accuracy risk of exactly this kind. Our related piece on safe generative AI use for legal professionals covers the wider set of obligations this sits within, including confidentiality and supervision.

Building verification into normal workflow

The firms that manage this risk well do not treat verification as a separate compliance step bolted onto the end of a task — they build it into how research is done from the start, so that pulling the primary source is simply part of the process, not an optional extra that gets skipped under deadline pressure. That habit costs very little time when applied consistently, and it is the single most effective safeguard against exactly the kind of error that embarrassed a litigant in Harber and has caused real professional consequences for lawyers in other jurisdictions who filed AI-generated citations without checking them first.

What exactly is an AI "hallucination" in a legal research context?

It is confidently generated but false information — most commonly a case citation, statutory reference or legal proposition that sounds plausible but does not correspond to any real, findable source.

Are AI tools built specifically for legal research safer than general-purpose chatbots?

Generally yes, because many retrieve from verified legal databases rather than generating purely from pattern prediction, which reduces but does not eliminate the risk. Independent verification remains necessary regardless of which tool is used.

What happened in Harber v HMRC?

A litigant in person relied on case summaries that turned out to be AI-generated and fictitious; the tribunal was unable to find the cited cases on any legal database because they did not exist. The case is widely referenced as an early, concrete illustration of AI citation risk.

How should verification be documented for supervision purposes?

A brief note of which research steps used AI assistance and how the output was checked — for example, which primary sources were pulled and confirmed — is usually enough to give a supervisor confidence and a clear audit trail if questions arise later.

Treating AI as a research assistant rather than a source of truth keeps its speed advantage without inheriting its risk. Learnsignal's CPD courses cover the practical research and verification skills that make that distinction workable day to day.