Fund's Legal AI Investment Files Patent

Counsel Stack filed a provisional patent last week in the area of citation accuracy verification.
Founder and CEO Von Wooding spoke about the patterns that led his Pittsburgh startup, and Brown and White Ventures portfolio company, to filing for the patent.
โI think they matter for anyone building in legal tech, and for the profession more broadly,โ said Wooding .
Counsel Stack believes the legal ๐ฐ๐ถ๐๐ฎ๐๐ถ๐ผ๐ป ๐ฝ๐ฟ๐ผ๐ฏ๐น๐ฒ๐บ ๐ถ๐ ๐ฒ๐ป๐๐ฒ๐ฟ๐ถ๐ป๐ด ๐ฎ ๐ต๐ฎ๐ฟ๐ฑ๐ฒ๐ฟ ๐ฝ๐ต๐ฎ๐๐ฒ, beyond what most know as standard LLMs hallucinate citations. These failures are often ones that a database lookup away from being caught.
โThe industry is approaching the point where most AI generated citations will be technically correct. As in the Bluebook details will be mostly real. The case exists. The reporter volume is correct. The year checks out,โ Wooding said.
But it is then when legal tech enters an entirely more challenging dimension of the problem, when the question will no longer be โdoes this case exist?โ but rather โdid the model correctly represent what this case stands for?โ
Some project this response is generated by the language model simply to appease the author's request.
Wooding said these gauge ๐บ๐ผ๐ฑ๐ฒ๐น๐ operate as general reasoning machines and ๐ฑ๐ผ ๐ป๐ผ๐ ๐๐ต๐ถ๐ป๐ธ ๐น๐ถ๐ธ๐ฒ ๐น๐ฎ๐๐๐ฒ๐ฟ๐, but instead, process patterns.
โA language model is a general reasoning machine. It processes patterns in legal text the same way it processes patterns in cooking recipes and Reddit threads,โ Wooding said.
โIt does not distinguish a holding from dicta the way a trained attorney does. It does not understand that a case with three interconnected issues may have distinct rulings on each, and that characterizing โthe holdingโ requires judgment about which ruling the citing context is invoking. It does not grasp that a concurrence carries different weight than the majority opinion., Wooding said.
Leaders in the field like Wooding call these models โnot deterministic. If a lawyer asks the same model what the holding of a complex case is, they will get a different answer each time. Not a wrong answer, necessarily, but a different emphasis, scope, or framing of which sub-holding is primary.
Wooding believes itโs a problem of stewardship running against the legal systemโs nature of
When we delegate that characterization to general-purpose reasoning models that don't think in legal categories, aren't trained to make the judgments lawyers make, and can't produce the same answer twice, we're risking a rapid degradation of how legal knowledge is represented and transmitted.
The systems that solve this should NOT be more general reasoning models stacked on top, according to Wooding.
โYou can't fix non-determinism with more non-determinism. That approach turns into a game of telephone that ironically enough increases hallucination risk. You need something that is auditable, verifiable, reproducible, and deeply grounded in sources.โ




Comments