Week by week: how this started
This section is the informal side of the project. The articles stick to a question and what the judgments show. Here I’ll write about what I’m working on, what’s going well, what isn’t, and where it’s heading.
Like many practitioners, I first ran into AI through clients. It arrived as questions that were unusually complex or technical, made sense on the surface, and fell apart in practice. Some of those questions show the limits of conventional legal research. The usual reference sites have little to say, neither do search engines or the tribunal register, and there is almost no appellate authority. Ultimately, the propositions are so unusual, and so plainly wrong, that nobody has ever needed to knock them down at a level anyone would report.
The project is also partly a response to a familiar LinkedIn theme: criticism of lawyers relying on AI. I’m a tech enthusiast by nature. Much of my experience has come from law centres and pro bono work, where finding tech solutions to routine tasks matters in a way that law firms can take for granted. I’ve built tools for smaller, everyday problems in the past, and in many ways this project is the natural next step.
It started with a question: could a more grounded approach fix the familiar problems with asking a chatbot for legal advice? I set that out in the first article in the series.
From there it moved to injury to feelings, a gap I hit constantly as a claimant lawyer. It’s no secret that settlement and mediation are often better routes than fighting a claim to the bitter end. But for someone with a discrimination claim, the hard part is usually knowing what the claim is worth and what a reasonable settlement looks like.
This is where I think one of AI’s more negative impacts is understated. Injury to feelings compensates the effect of the discrimination on the claimant, and that plays directly into one of AI’s core weaknesses: it rarely pushes back. A claimant describes how badly they’ve been affected, the model reflects that back, and the feedback loop can produce expectations no tribunal would meet. Anchoring the conversation in what tribunals actually award is an effective counterweight. The awards viewer came out of that work, along with a comparable cases tool that hasn’t yet made it onto the site, and I think it remains one of the most useful parts of the project.
After that I started looking at whether a tool like this could run on an ordinary local PC. I’ve spent much of my career working for charities and third sector organisations, and I’m still passionate about finding ways to make legal advice more accessible. This part of the project is at an early stage. The rough idea is a tool robust and reliable enough to be offered to advice charities, running on modest hardware they already have or could afford.
Lately I’ve gone down something of a rabbit hole with trends. One thing this approach does very well is search the tribunal judgments, pull out patterns, and then apply proper research and reasoning to them. So far the trend articles have been fairly broad, but there’s no reason they couldn’t be much more focused. For example:
- What are the most common claims brought against NHS trusts, and how do they fare?
- How often is the ACAS uplift awarded, and at what percentage?
- How do tribunals approach requests to work from home as a reasonable adjustment?
This week
The main things out of the door were the article on what tribunals see when parties use AI, and the extensions of time article and viewer. I’ve also started building a viewer to track AI use in the employment tribunal on an ongoing basis, and started the research for a section 15 article on autism and ADHD.
Writing about other people’s use of AI made me look harder at my own, and one thing stood out: a steady drift towards cloud AI. The method this site describes is local-first. A model on my own machine reads the judgments, rules check its answers against quotes from the text, and a cloud model is brought in only to settle the items that are genuinely uncertain. The local side is still doing the heavy lifting: finding the relevant judgments among tens of thousands and giving each one a first reading is work no cloud model could do at that scale. But in the newer trend studies, the cloud models had gradually moved from settling the leftovers to re-reading large shares of the cases and, in effect, having the final say. That covered the “something arising” in section 15 claims, the impairments relied on in disability claims, and the reasons an extension of time was granted or refused.
None of it came from a single decision. The newer questions were more interpretive than “how much was awarded”, and each study felt like a one-off, so handing the reading to a cloud model was the quickest route to an answer. Looking back, two things stand out. On the simpler questions, such as whether an extension was granted, the local model mostly reached the same answer as the cloud. On the harder, interpretive questions, the bigger models didn’t settle things either. They disagreed with each other, and sometimes with themselves on a second reading. The real problem was loosely defined questions, and that can be fixed without leaving the local model behind.
From now on, new cases going into the viewers will be read locally first. Only the disputed ones go to a cloud model, and a random sample of the rest gets checked too. I suspect this kind of drift is common, which is partly why I’ve written it up here.
Next up: ADHD, autism and section 15. That means what the “something arising” from the condition tends to be, how often tribunals accept the link, and what happens to the employer’s justification when they do.