I’ve been using AI in two very different parts of my life this year. At work, Copilot has become part of my daily routine: drafting, summarising, organising, and occasionally rescuing me from the administrative black hole that litigation can become. Outside work, I’ve been using AI in a completely different context - marathon training. Not instead of a coach, but alongside one. My coach sets the plan, but Copilot helps me understand pacing, heart rate zones, heat adjustments and all the other variables that make marathon training feel like a part‑time degree. If you’re not a runner, this probably sounds excessive. If you are a runner, you’ll know it’s entirely normal. We are, as a group, incapable of doing anything without data. I don’t even take my running watch off at bedtime. It’s also given me a new perspective on AI mistakes and legal reasoning, and why human judgement still matters.
Both experiences have been positive. Both have made me more efficient. And both have taught me something important about how AI behaves - something I think lawyers, and anyone relying on AI, need to understand. Because last week, Copilot confidently told me I’d finished far higher overall in the Berlin Marathon than I actually did. In fact, it briefly had me down as a world‑class, elite‑level athlete - the sort of placing where UK Athletics might start knocking on the door. They won’t. And the way that mistake unfolded says something much bigger about the risks of treating AI as an authority rather than a tool.
When I checked my race results, I saw three numbers: my overall place, my category place, and my age‑group place. I asked Copilot to interpret them so it could calculate my overall percentile - essentially, where I finished as a percentage of all runners. In doing that, it confidently promoted my category placing (female) into an overall placing, putting me far higher in the field than I actually was. It sounded authoritative. It sounded plausible. It sounded impressive. And for a second, I almost accepted it. In fact, I asked it several times to double‑check the interpretation - and it didn’t budge. It repeated the same conclusion with the same certainty, even though the numbers didn’t make sense together. A human would have spotted that immediately. The AI didn’t.
At that point, curiosity got the better of me. I wanted to know whether another Copilot agent - the one integrated into my work systems - would reach the same conclusion. So I ran the exact same numbers through it, word for word, just to see whether two agents would interpret the data consistently. The work Copilot immediately gave me the correct answer: my placing was nowhere near the lofty position the first agent had awarded me. Same data. Same labels. Completely different conclusion.
Using AI for marathon training has made me more aware of this than any legal example ever could. When you’re dealing with case law, you expect complexity. You expect nuance. You expect interpretation. But when you’re dealing with race results - simple numbers - you expect clarity. And yet even there, AI can misinterpret the data and stick to the wrong conclusion with absolute conviction. If I hadn’t questioned it, I could easily have posted my “elevated” placing on LinkedIn. And yes, I admit it: I did actually ask Copilot to draft a LinkedIn post about the race. Like most people on that platform, my “authentic voice” is often AI‑assisted; you can usually tell by the suspicious lack of typos.
This is the part that matters. AI is programmed to be authoritative. It is designed to give you a clear answer rather than a hesitant one. It is built to maintain a coherent line of reasoning rather than flip‑flop when faced with uncertainty. Those design choices make it feel helpful and decisive - but they also mean that when AI gets something wrong, it gets it wrong confidently. And unless you challenge it, it will continue to be wrong. This is exactly where AI mistakes and legal reasoning collide.
This is also why lawyers won’t be replaced. Not because AI can’t draft documents or summarise judgments -it can. Not because AI can’t analyse patterns or produce neat explanations - it can do that too. But because AI cannot reliably interrogate its own reasoning. It cannot apply scepticism. It cannot test assumptions. It cannot say, “This doesn’t look right.” It cannot do the thing lawyers are trained to do instinctively: challenge the answer. It’s the gap between AI mistakes and legal reasoning that keeps lawyers essential.
Most people won’t challenge an AI. They’ll assume it’s correct because it sounds correct. They’ll assume “it’s Microsoft, so it must be careful.” They’ll assume the confidence reflects accuracy. But confidence is not competence. And unless someone pushes back - repeatedly, if necessary - the AI will continue down the wrong path.
So if you’re using AI, whether in legal practice or marathon training, here’s what I’ve learned. Ask it to explain its reasoning. Ask it to show alternative interpretations. Ask it to sanity‑check the numbers. Treat it like a junior colleague: helpful, fast, and occasionally brilliant, but not infallible. And never forget that the skill lawyers bring - the ability to interrogate, challenge, and scrutinise - is precisely the skill AI cannot replicate.
AI is a powerful tool. It can help you train for a marathon. It can help you draft a witness statement. It can help you organise your day. But it still needs someone who knows how to ask the right questions. Someone who knows when something doesn’t look right. Someone who isn’t afraid to push back.
In other words: it still needs a lawyer.
Legal disclaimer The matters contained within this article are intended to be for general information purposes only. This blog does not constitute legal advice, nor is it a complete or authoritative statement of the law in England and Wales and should not be treated as such. Whilst every effort is made to ensure that the information is correct, no warranty, either express or implied, is given as to its’ accuracy, and no liability is accepted for any errors or omissions. Before acting on any of the information contained in this blog, expert advice should always be sought.
© Melissa Worth, October 2026