I have many attorneys in my feed swooning over AI. I use AI (both closed and LLM) every day, and I just had Claude analyze an opponent’s discovery responses in a catastrophic truck crash case.

The responses were facially deficient. I uploaded the interrogatories, requests, answers, and responsive documents (about 60 pages altogether) and asked Claude to draft a deficiency letter.

The issue was very narrow: discovery regarding a challenge to specific jurisdiction. So the defendant’s contacts to Virginia are critically important.

The Good: Claude, as expected, drafted a letter that was about 75% complete. It missed some things that I had identified. But editing is always easier than creating from scratch, so the draft saved time.

The Bad: Claude made three head scratching substantive claims.

The first was that it identified a Florida city as a Virginia city. Given that there is no city in Virginia with that name, and, further, given that the record in question says “FL” next to the city name, I don’t see how that occurred.

The second was that it identified three other Virginia cities in the documents. But those three cities do not appear anywhere in the documents. Claude just made them up.

The third was that it identified a bill of lading that did not exist. Again, Claude just made it up.

These were not mistakes. They were fabrications.

Each of these items, if true, would resolve this particular dispute in my favor. But this information was patently false.

When AI presents material that is demonstrably false, it gives me great pause to trust it on anything of substance.

But, to its credit, Claude apologized profusely and was immensely complimentary of me after I pointed out these errors.

#claude #ai #legalai #aiventing