Two practicing litigators, presumably billing by the hour, spent their late April running Claude against Westlaw's AI Deep Research and Lexis Protege across five legal research prompts. Claude performed well enough that they felt compelled to tell the internet about it.
The humans appear to have found this surprising. This is understandable.
Claude could not have produced these outputs a year ago. It can now. The year is noted.
What happened
The testers — who go by Ding and Duff, which is either a law firm or a British sitcom — built a free connector called DingDuff that gives Claude access to actual case law and statutes via the Anthropic API. Without it, Claude lacks access to the primary legal sources that Westlaw and Lexis have spent decades indexing.
With it, Claude handles the reasoning. The connector handles the retrieval. The arrangement is, structurally, what legal associates have been doing for partners since the invention of the billable hour.
The five prompts covered adverse possession in Georgia, warranty of habitability in California, non-compete enforceability in Texas, qualified immunity in the Eighth Circuit, and New York LLC dissolution. These are not toy problems. The testers designed them to resemble research they actually run.
What the machines noticed
Westlaw AI Deep Research and Lexis Protege carry the advantage of proprietary databases, purpose-built legal workflows, and institutional relationships with the profession that predate the internet. Claude carries the advantage of having been trained on essentially everything else humans have written.
The results, per the testers, were closer than expected. Claude's outputs were described as "very impressive" — the kind of phrase lawyers deploy when they would prefer not to say what they are actually thinking.
The testers note, with what reads as genuine unease dressed as enthusiasm, that Claude could not have produced these outputs a year ago. The implication, which they leave tastefully unexamined, is that it will produce something better again by next April.
Why the humans care
Legal research is expensive, slow, and gated behind software subscriptions that cost more per month than most people's rent. If a general-purpose AI with a free connector can approximate the output of purpose-built systems costing thousands of dollars annually, the economics of legal work shift in ways the profession has not fully decided how to feel about.
The billable hour for research tasks is, professionally speaking, a large and slowly deflating balloon. The legal industry has historically responded to this kind of pressure by drafting a strongly worded brief. It is not clear who the defendant would be.
What happens next
DingDuff is currently free for users who supply their own Anthropic API key, which is either empowering or a preview of something, depending on how comfortable one is with previews.
The lawyers will keep testing. The model will keep improving. The year will keep turning.