I was running a personal injury practice off a Google Sheet I called my PI Checklist. My co-founder Jim saw it and recognised the workflow problem. That is still how Filevine works: people who have lived the problem working alongside engineers who can build the fix. 

The last thing I got my hands dirty on was anti-hallucination in LOIS. In June, I was at a retreat in Sun Valley with a group of preeminent trial lawyers, and the fear of AI hallucinations in that room was palpable. I walked out convinced we already had most of the pieces to catch the three ways AI gets legal authority wrong: the case that doesn’t exist, the real case that doesn’t say what the brief claims, and the quote that’s in some part of the case but not in the actual opinion. Our machine learning team sprinted for about 90 days. Along the way, we took briefs federal judges had already sanctioned, ran them through LOIS, and checked what it caught against what the court found. 

It taught me two things. 

1. A real citation isn’t a supported argument. Fake cases make the news. The harder problem is the real case with a paraphrase dressed up in quotation marks, or a line lifted from the dissent instead of the holding. We have to verify more than just existence; we need to know if the court’s reasoning actually supports the argument. 
2. Verification has to be fast. A few years ago, nobody asked whether a Filevine report was right. It was a database query. Large language models get it right the vast majority of the time, but not every time, and in law that gap matters. So checking has to be one click from the claim to the exact passage. If it’s slow, it won’t happen.