Assistance, not authority
Use AI to surface candidate passages, compare translations or outline issues for human counsel. Final opinions remain with those competent to give them. Write that boundary into office policy. A fluent wrong citation is more dangerous than a slow correct one.
Workflow ideas
- Retrieve from approved corpora with citations
- Cluster related documents for a matter
- Draft research memos clearly labelled as drafts
- Track open questions for the human reviewer
Confidentiality
Case materials may be highly confidential. Use hardened environments only. Default public models are inappropriate for active matters. Align with secure AI practices and role-based access.
Quality checks
Verify every citation against the source. Models invent references. A wrong canon citation is a serious professional risk—not a “hallucination” to laugh off in a staff meeting.
Working with limited specialist time
Canonists and trained officers are scarce. AI should reduce time spent hunting files, not invent conclusions for them to rubber-stamp. Design queues so the specialist sees candidates and citations, not a paragraph that pretends to be a rescript. Measure specialist time saved on retrieval, not “opinions generated.”
When the system is unsure, it should say so. Confident silence is worse than an explicit unknown.