How lawyers can use GenAI without exposing clients or getting sanctioned
AI is now baked into legal work. The upside is speed. The risk is simple: one bad prompt, one sloppy vendor, and you can blow privilege, leak PII, or file something you can't defend.
This guide gives you the playbook-what to watch, what to implement, and how to keep your license, your clients, and your reputation intact.
What GenAI changes in your risk profile
Traditional tools sit inside your stack. GenAI often sits outside it. Prompts, outputs, and embeddings can cross systems you don't control. That's why confidentiality, accuracy, and vendor risk all move to the front of the line.
Add in court-specific expectations and recent sanctions for fake citations, and you've got a real compliance problem if you wing it.
Your ethical duties still apply
Think through the same lens you already know: competence, confidentiality, supervision, candor, and fairness. Model Rules 1.1, 1.6, 5.3, and 3.3 aren't optional just because the tool is new.
If an AI system or contractor touches client data, they fall under your duty to supervise. If you submit AI-generated content, you own it-facts, law, and citations.
ABA Model Rules of Professional Conduct
Practical guardrails for daily use
- Set a firm-wide AI policy: What tools are allowed, what data is prohibited, who approves vendors, and how outputs are verified.
- Keep client data out of public models: No names, SSNs, financials, or facts that could identify a matter. Use approved, private instances only.
- Turn off training: Require written confirmation from vendors that your prompts and outputs aren't used to train their models.
- Rely on retrieval, not memory: Use private, read-only knowledge bases (RAG) so the model references trusted sources instead of guessing.
- Force human review: Every AI draft gets a human editor. Verify quotes, caselaw, and record cites before anything leaves your desk.
- Keep audit trails: Log prompts, versions, reviewers, and sources. If a question comes up, you can show your work.
- Use red-team prompts: Ask the tool to argue against its own output. Surface weak points before an opponent does.
Vendor diligence you can defend
- Security: SOC 2 Type II or ISO 27001, SSO/MFA, encryption in transit and at rest, role-based access, IP allowlists.
- Data handling: No training on your data, clear retention/deletion timelines, data residency if required.
- Contracts: Confidentiality, incident notice within 72 hours, indemnity for privacy/IP violations, right to audit, and subprocessor lists.
- Operational risk: Uptime SLAs, offline export, and a clean offboarding plan if you need to move fast.
Bankruptcy practice: filings, privacy, and courts
Bankruptcy matters carry heavy PII-SSNs, account numbers, addresses, payroll data. Treat every prompt like it's a public filing if you're not on a private, approved system.
Follow required redactions and privacy limits in your filings, and make sure AI tools never store unredacted data outside your controlled environment.
U.S. Courts: Records & Privacy Policy (Redaction requirements)
E-discovery and evidence
- Authenticity: If AI helped generate or summarize evidence, be prepared to authenticate the process and the source materials.
- Chain of custody: Keep original files, system logs, and a clear record of transformations.
- Deepfakes and altered media: Build a basic forensic review step for suspect audio, video, and images.
Court expectations and sanctions risk
Some judges now require certifications that filings were reviewed by a human and that citations are real. Courts have sanctioned lawyers for fake cases and misstatements pulled from public chatbots.
Best practice: disclose limited, non-privileged use when a standing order requires it, and always verify every cite and quote against primary sources.
Incident response for AI-related events
- Detect: Alerts on unusual data exports, admin changes, or mass downloads from AI systems.
- Contain: Disable tokens, rotate keys, lock affected workspaces, and suspend risky integrations.
- Assess: Identify what data was exposed, whose data it was, and what contractual or statutory duties apply.
- Notify and fix: Follow client, regulatory, and court notice obligations; patch the control that failed.
Prompts that reduce risk
- Context control: Provide only what's needed. Replace names with roles. Use matter codes, not client names.
- Citation requirement: "Cite only to primary sources and include pincites. If uncertain, say so."
- Fact flags: "List any assumptions. Separate facts from inferences. Mark anything that needs verification."
What to train your team on this quarter
- Firm AI policy, allowed tools, and no-go data categories.
- How to redline vendor terms and confirm no-training settings.
- Prompt hygiene and verification routines for legal analysis.
- Spotting AI failure modes: hallucinated cites, stale law, and subtle misquotes.
Bottom line
AI can speed research, drafting, and review. Your job is to keep it inside ethical and security boundaries. Set clear rules, pick the right vendors, and verify everything that hits the record.
If you need structured upskilling for your team, see these curated options: AI courses by job role.
© 2025. All rights reserved.
Your membership also unlocks:
