Ask ChatGPT a legal question and you'll get a confident, well-structured answer. The problem is that the answer might be based on legislation that was amended two years ago, a case that doesn't exist, or a regulatory position that's since changed. The answer will sound just as authoritative either way. That's the danger.
This is the single thing every solicitor needs to understand before using any AI tool for research: there is a fundamental difference between an AI that retrieves the law and an AI that predicts what the law probably says. Both produce fluent, professional-sounding answers. Only one of them can be verified.
Why generation from training data isn't research
A general-purpose AI model, ChatGPT, Claude, Gemini used as a chatbot, doesn't look anything up when you ask it a legal question. It generates a response based on patterns in the text it was trained on. That training data has a cutoff date. It may not include the latest version of a statutory instrument. It may not include a recent Court of Appeal decision. It definitely doesn't reflect regulatory guidance published last month.
The citations it produces can look entirely real. The case name might be correct but the citation wrong. The statute might exist but the section it quotes might say something different. Or the case might simply not exist, generated because it sounds plausible, not because it was ever decided.
This has already caused real professional problems. Lawyers in multiple jurisdictions have submitted fabricated AI citations to courts, not through carelessness, but because the citations looked credible and nobody checked them. The SRA has been clear that verification is a professional obligation, but verification is only possible if there's something real to verify against.
What retrieval-based AI actually does
A retrieval-based tool works differently. Before generating any response, it queries live sources, legislation.gov.uk, BAILII, SRA guidance, court rules, retrieves the relevant documents, and then generates an answer grounded in those actual sources. The citations in the response point to real documents you can read.
This matters for a simple practical reason: you can check it. The professional obligation to verify research doesn't go away because an AI produced it. But with a retrieval-based tool, verification is a realistic step rather than an impossible one.
The test to apply before relying on anything
Before you use any AI output in your work, ask three questions:
Can you click through to the original source? If the tool doesn't provide a link to the actual legislation or judgment, there's nothing to verify.
Is that source an official, live publication? Legislation.gov.uk shows the current version of statute in force. BAILII has the full text of UK judgments. An AI summarising a secondary source, or citing a version of a statute from its training data, is not the same thing.
Does the source say what the AI says it says? This is the step that takes thirty seconds and catches most errors. Go to the page cited. Read the relevant section. Confirm it matches.
If any of those answers is "no" or "I can't tell", treat the output as a prompt for your own research, not as research itself.
The practical implication
For everyday use, this doesn't mean avoiding AI tools. It means using the right ones for different purposes.
If you're drafting, exploring a question informally, or preparing for a client meeting, a general-purpose AI is useful. The stakes for a hallucination are low, you're thinking things through, not advising.
If you're producing research you'll rely on for a client, you need a tool that retrieves from live UK sources and gives you citations you can check. The professional responsibility is yours. The tool just needs to give you something real to work with.
Writford's Research mode queries live UK sources, legislation.gov.uk, BAILII, SRA guidance, and attaches paragraph-level citations to every answer. That's the thing that makes verification possible.