Why law matters for AI safety
Legal questions rarely have a context-free answer. The right response depends on jurisdiction, dates, facts and procedure, and a model that ignores any of them can sound authoritative while being wrong.
People act on legal answers. They sign, file, or miss a deadline. Teaching a model where the edges are takes people who work inside them.
What legal experts do
Legal experts write reference answers, rank model responses and annotate the reasoning behind them. They review contract analysis, research summaries and drafted documents for accuracy and completeness.
They evaluate legal research and drafting agents, red-team models for harmful or misleading legal guidance, and produce alignment data on when a model should state its limits or recommend a lawyer.
- Expert-written answers and legal reasoning
- Ranking and grading model responses
- Review of contract and document analysis
- Evaluation of legal research agents
- Red-teaming and alignment data
Failure modes legal experts catch
Legal experts catch cited authorities that do not exist or do not say what the model claims. They catch rules applied in the wrong jurisdiction, and contested positions presented as settled law.
They also spot what is missing: a limitation period, a procedural step, a compliance obligation the user did not ask about but needs to know.
How we vet legal experts
Legal candidates are interviewed by AI voice agents on how they analyse a legal problem. They complete real-world work tests from legal practice, such as reviewing a clause or assessing a model's answer to a legal question.
Credentials and experience are reviewed as part of vetting. Paralegals and compliance specialists are tested on the work their roles involve, not on a lawyer's test.
Matching on jurisdiction and practice area
Law varies by country and by practice. We match experts to the jurisdiction and area a project covers, and add project-specific assessments when the work needs them.
A contracts lawyer and a compliance specialist will each catch errors the other misses. Matching on the task, not the title, is how a project gets the judgement it needs.