Every few weeks there’s a new AI model to talk about. Faster, cheaper, apparently smarter than whatever came before it. It’s the kind of news that’s easy to get swept up in, and if you work anywhere near legal tech, it can start to feel like keeping up with model releases is the job.
It isn’t, though. Not really. Not if you’re the person whose name ends up on a report that a judge or a jury will eventually read.
There’s a quieter shift happening in UK legal technology at the moment, and it matters more than any single model launch. Legal AI tools are starting to be built differently. Not as clever text generators with a legal skin bolted on, but as systems designed from the ground up to show their working: who used them, what they were asked to do, what a human checked and changed before it went anywhere near a client.
Why now
Part of the reason is that the UK hasn’t gone down the route of writing one big AI law. Instead, the existing regulators are being told to deal with it themselves. The Solicitors Regulation Authority put out updated guidance earlier this year making the point, again, that a solicitor is personally responsible for AI-assisted work regardless of which tool produced it. The Information Commissioner’s Office has been working through updated guidance on automated decision-making. And from August 2026, parts of the EU AI Act start applying to UK firms that offer AI-powered services to clients in the EU, which brings its own paperwork.
None of these things add up to a single clear rulebook. What they add up to is a situation where, sooner or later, someone is going to ask “can you show me exactly what happened here,” and firms that can’t answer that properly are going to have a bad day.
What this actually looks like in a tool
A handful of things tend to separate the legal AI products built for this environment from a generic chatbot with a legal wrapper on it.
There’s usually a proper audit trail. Not just a note saying AI was involved somewhere, but a record of which tool was used, what was asked of it, what sources it drew on, and what a person changed before it was signed off.
There’s a human checkpoint that’s actually built into the workflow, rather than something that depends on someone remembering to do it. The system won’t let output move forward until a qualified person has reviewed it.
The output tends to be tied to something real. A verified legal database, a known source, rather than open-ended generation that sounds plausible but can’t easily be traced back to anything.
And increasingly, UK data handling is designed in from day one rather than adapted later from a product that was originally built somewhere else with different rules in mind.
Where forensic psychiatry fits into this
Most forensic psychiatrists won’t be the ones choosing an AI tool. But they’re increasingly likely to be asked, by an instructing solicitor or in cross-examination, what was used in putting a report together, and whether that can actually be accounted for. As the rules around expert evidence keep shifting in the UK, “the conclusion was right” is looking less and less like it will be enough on its own. Being able to say clearly how something was produced, and who checked it, is where this is heading.
That’s not really a psychiatry-specific problem. It’s the same pressure the whole legal profession is under right now. But for a field that depends so heavily on the credibility of an expert’s own judgement, there’s a real difference between a tool that can show its working and one that just hands you a confident-sounding answer with nothing behind it. One of those survives being picked apart in court. The other becomes the story itself.
The actual takeaway
Whatever model gets announced next month will get its moment and then fade into the background, same as the last one did. The more useful question to ask about any AI tool, whether your firm is buying it or you’re being asked about one in your own work, isn’t which one is newest. It’s whether, if someone asked you to walk them through exactly how a piece of output was produced, you could actually do it.
If you can, the tool was probably built with this in mind from the start. If you can’t, that’s worth sorting out now, before someone in a courtroom asks the question for you.