Regression to the Mean

I've been experimenting with using LLMs for legal analysis. My problem isn't really with hallucinations but with "regression to the mean."

I read a brief, pulled the cases with WestLaw, uploaded them to a ChatGPT project, and then asked it some questions about the key arguments. It didn't hallucinate. It "read" the cases and would respond with citations to questions.The problem I found, however, was that its analysis was decidedly...well...pedestrian. It wasn't awful, but it just wasn't that insightful. It was able to summarize the arguments fairly well, but when asked questions like "Which argument is strongest?", "Do all the cases support the proposition they are cited for?", its responses were fairly mediocre. Not terrible; not implausible; just average.

I find this interesting for a few reasons.

First, it's not that surprising. LLMs are trained on a huge amount of data, and to oversimplify (enormously), they are predicting the most likely response based on that data. Accordingly, it's not a shock that what they give you is something like the average response.

Second, for just that reason, this may well be a fundamental limitation of LLMs. That's a hotly debated topic in many circles, but I'm curious what people will find in LegalTech. Are any of these legal-specific AIs actually better at analysis than ChatGPT/Gemini/Claude/etc.? That is, give them the same context (documents) and prompts, do they do better? My guess is no--but it's just a guess.

Third, average responses are perfectly fine for a lot of things--even some legal analysis--but it's hard for me to see how that will give you a competitive advantage.

I think of this last point in the context of the discussion about "will LLMs replace first year associates." To be sure, a lot of first year associates DO give average legal analysis, and there's nothing wrong with that. Moreover, sometimes first years (and even more senior!) give below average legal analysis, and LLMs could be an improvement on that. (And by the way, that's probably what makes using them tempting.)

But I can't tell you how many times it's made a difference in a case when I had a first year who gave ABOVE AVERAGE legal analysis. I wouldn't want to give that up. Nor would I want to be in a position where we are saying "well, the senior associates/partners will provide the real insight." That's not correct. Junior associates often do above average work because they are junior -- they are eager, they are hungry, they have fresh eyes, they haven't specialized yet. I wouldn't want to turn that over to an LLM that, by design, is going to give me something middle of the road.