AI Legal Research
Grounded in a real corpus, with every citation resolved before a human sees it.
Legal research is the single application where language models have caused the most professional harm, and it is also one where they work extremely well when built correctly. The difference is entirely architectural, and it is not subtle.
Why general models fail at this
A general-purpose model asked for authority supporting a proposition will produce it: correctly formatted, plausibly named, confidently cited, and frequently fictitious. It is not malfunctioning. It is generating text that resembles a citation, which is what it was built to do.
The failure is worse than an error because it is undetectable by reading. A fabricated citation looks exactly like a real one. The only way to catch it is to look the case up, which is the step under time pressure that people skip.
What a usable system requires
Three things, and all three are structural rather than behavioural. Prompting a model to avoid hallucination is not one of them and does not work.
- Retrieval from a real legal corpus, so the model only ever sees actual authority
- Automated citation resolution, so an unresolvable citation is removed and flagged before a human reads the draft
- Passage-level quoting, so the proposition a case is cited for is traceable to the text of the case rather than to the model's summary of it
The second-order failure
Even with real citations, a model can cite a real case for a proposition it does not support. This is subtler than fabrication and harder to catch, because the case exists and looks apposite.
Passage-level quoting is the mitigation: the system shows the specific text relied on, so a lawyer verifies the proposition against the source rather than against a summary. This is also why we do not build systems that summarise authority without quoting it.
Where firm-specific research adds most
Public authority is well served by the established research platforms. The gap those platforms cannot fill is your firm's own work: the memo written on this exact question four years ago, the brief that succeeded before this judge, the position the firm has taken before and should probably take consistently.
That corpus is private, it is not in any commercial database, and it is where a firm-specific system earns its cost.
Related
Common questions
Can we just use a general model carefully?
Care is not a control. The failure mode is invisible on the page and the people most likely to skip verification are the ones under the most time pressure, which is exactly when it matters. The resolution step has to be in the pipeline, not in the process discipline.
Do the commercial legal research tools hallucinate?
Grounded tools substantially reduce fabricated citations because answers are retrieved from a real corpus. They do not eliminate the second problem of a real case cited for a proposition it does not support. Verify against the passage regardless of the tool.
What is the value of a firm-specific system?
It answers questions no commercial database can: what this firm has argued, what worked, and where the internal precedent sits. That is institutional knowledge that currently walks out of the building when people leave.
How do you enforce confidentiality across matters?
Access control is applied when documents are retrieved, so a user cannot surface a matter they are walled off from. Enforcing it through prompt instructions instead is the most serious design error we see in firm systems.
Find out what this looks like for your organisation
A 30-minute call. We will tell you plainly whether AI is the right tool for the problem you have, including when it is not.
Book a free consultation