If you’re training a model to do legal work, or buying data for a team that is, public case law gets you part of the way. It tells a model what the law says. It doesn’t show how a lawyer drafts a contract, takes a position in a negotiation or decides what to accept, and that record sits inside firms. Read on to find out the five kinds of legal data labs train on, what to check before you license any of it, and where to get it.
The five kinds of legal data
“Legal data” covers five different things, and the rights question changes with each one.
| Kind of data | What it teaches a model | The rights question |
|---|---|---|
| Statutes, regulations and court opinions | What the law says | The text is public, so every lab already has it |
| Publisher content (headnotes, annotations, practice guides, forms) | How experts summarize and organize the law | The publisher owns it, and a subscription isn’t a training license |
| A law firm’s own work (templates, precedent banks, playbooks, know-how memos, training guides) | How lawyers draft, negotiate and reason | The firm owns what it wrote for itself, and anything from a client matter stays out |
| A company’s own contracts and legal documents | What real agreements look like and how they change | The company holds them, and the other party’s terms may limit what it can share |
| Client files and matter documents | How a specific matter was handled | Covered by the lawyer’s duty of confidentiality, and not licensable |
The third row is the scarce one. A statute is public. A firm’s playbook for what to accept, what to push back on and why is not.
Why public case law isn’t enough
Opinions and statutes are the same for every lab, so they don’t set a model apart. They also show the end of a legal question and none of the work that got there. A model that has read every appellate decision still hasn’t seen a first draft, a markup or a fallback position.
Publisher content doesn’t fill the gap either, and it carries a risk of its own. In September 2026 a federal appeals court upheld a ruling against Ross Intelligence, an AI legal research company that trained on a publisher’s headnotes without a license. Our guide to AI training data lawsuits covers where the cases stand.
Why labs want legal work
Buyers tell us they pay a premium for full company datasets from industries built on high-value knowledge work, and legal is one of the industries they name. Drafting a contract, taking a position in a negotiation and reasoning through a question of law take years to learn.
OpenAI’s GDPval benchmark, published in September 2025, shows which jobs labs measure their models against. It tests models on real tasks from 44 occupations across the nine sectors that each contribute more than 5 percent of US GDP, and lawyers are one of the 44. OpenAI built each task from the work of a professional with 14 years of experience on average, and a task takes an expert about seven hours.
Legal AI companies like Harvey and enterprise AI companies like WRITER and Cohere are building models for this work. Agents raise the bar again, because an agent has to finish a legal task from start to end. Our guide to training data for AI agents covers the records agents learn from.
The record that trains best is how a firm’s forms changed
Buyers tell us that the more objectively a piece of work can be checked, the more readily they license it. Legal work is checked differently from a financial model, and the clearest record a firm holds is how its own forms evolved. A clause that was added after a dispute, a fallback position that got adopted and a template that was revised when the law changed each carry their reasoning with them.
The value is highest where the records connect. A template linked to the playbook that explains each clause, the training guide that teaches it and the revisions that followed a change in the law is a complete record of how a firm practices. Our guide to financial training data for AI covers the same idea for finance work.
What to check before you license legal data
Legal data comes with stricter limits than most, and a dataset that ignores them is a liability. Six checks sort it out.
- Is any client matter material in it? A lawyer’s duty of confidentiality covers more than privileged documents, so client files, matter documents and drafts written for a client stay out.
- Was any of it received under a protective order, a confidentiality agreement or a court seal? That stays out too.
- Is publisher content mixed in? Research content, practice guides and forms licensed from a publisher belong to the publisher.
- Have names and personal details been replaced with pseudonyms to a documented standard? That covers clients, opposing parties, witnesses and staff.
- Who approved the scope? A firm’s ethics and risk partners should have signed off, and the rules of its bar decide what’s possible.
- Does the license cover commercial model training in writing? Your own counsel will ask.
This isn’t legal advice, and your counsel should review any legal dataset before you train on it. Our guide to rights-cleared training data covers what clearing a dataset takes.
Where to get legal training data
You have four routes, and they get you different rows of the table above.
- Start with public law. Statutes, regulations and opinions are free to collect, and every other lab has them.
- License content from a legal publisher. The license has to name model training as a permitted use.
- Pay lawyers to write it. Practicing and former lawyers write tasks, drafts and model answers to your brief. You control the format, and every hour is billed. Our guide to AI data collection companies covers the vendors that work to order.
- License real work from the firms and companies that produced it. Law firms can license what they wrote for themselves, and companies can license their own contracts and documents. Our guide for law firms licensing their data covers what a firm can license and what has to stay out, and our guide to who buys company data for AI covers the five kinds of buyer.
Where Troveo fits
Troveo is a data licensing marketplace, and we’ve paid rights holders more than 20 million dollars. Law firms can license the work they wrote for themselves through us (templates and precedent banks, negotiation playbooks, know-how memos and training materials). Client matter material stays out entirely, the firm’s ethics and risk partners sign off on the scope, and we replace names and personal details with pseudonyms and document what’s included in a manifest. You sign one agreement with us and never negotiate with each firm. Tell us what legal work your model needs to learn and we’ll tell you what we hold.