← All insights

Using AI Safely: Verification, Confidentiality and Firm Policy

Cover image for Using AI Safely: Verification, Confidentiality and Firm Policy
Article illustration

Photo: Pexels, free to use under the Pexels License.

Category: Blog | Audience: Partners, supervising lawyers, compliance and risk leads | Reading time: 6 minutes

The question is no longer whether lawyers will use AI. It is whether they will use it with controls. Courts have been clear that responsibility for what is filed stays with the human who signs it. Here is a practical framework built on three pillars: verification, confidentiality and governance.

Pillar 1: Verification

Courts across several jurisdictions have sanctioned lawyers and litigants for filing fabricated citations. In one federal case in the Northern District of New York, discussed in a Thomson Reuters analysis, the attorney acknowledged knowing that AI tools can invent citations, and the court still imposed a fine of 1,000 US dollars and required AI training. The lesson is that knowing the risk does not substitute for checking.

Use the same pass every time:

Article illustration

Check three things, not one. A citation can fail in three ways:

  • The case does not exist.
  • The case exists but the quotation is not in it.
  • The case exists and the quote is real, but it does not support your proposition.

Confirming existence alone catches only the first.

Check currency. Run every authority through a citator to confirm it has not been overturned or limited.

Keep a record. Note who verified what and when. If your work is ever questioned, a short verification log is powerful evidence of diligence.

Pillar 2: Confidentiality

Client information is the asset you are professionally bound to protect. Before anyone pastes anything into an AI tool, settle these points:

  • Data terms. Does the vendor retain your inputs? Are they used to train models? Can that be switched off in writing?
  • Tool tier. Consumer accounts and enterprise accounts often have very different data protections.
  • Privilege. Treat anything entered as potentially disclosable unless the vendor agreement and your jurisdiction's rules clearly say otherwise.
  • Minimise. Remove names and identifying details where the task does not need them.
  • Client consent. Some clients and engagement letters now address AI use explicitly. Check yours.

Pillar 3: Governance

A recurring finding in 2026 surveys is that adoption is outrunning policy. Paragon Legal's August 2026 research, based on more than 150 legal professionals, reported that 56 percent of organisations lacked formal AI governance even as use increased.

A one-page policy is far better than none. It should cover:

  1. Approved tools. A short list, and a process to request additions.
  2. Permitted and prohibited uses. For example, drafting internal summaries allowed; entering client-identifying data into unapproved tools prohibited.
  3. Verification standard. The pass described above, made mandatory for anything filed or sent to a client.
  4. Supervision. Junior lawyers and staff using AI must have a reviewer who understands the tool.
  5. Disclosure. When and how you tell courts or clients about AI assistance. Rules are evolving; the Chief Justice of India, for instance, called this week for model guidelines on disclosing AI assistance in legal research.
  6. Training. A short annual session. A survey of 380 law students in India found that 71 percent had received no formal training on ethical AI use, a gap worth closing in every jurisdiction.
  7. Incident response. What to do if an error reaches a court or client: who is told, how it is corrected.

A subtler risk to plan for

Researcher Damien Charlotin has argued that as fabricated cases become rarer, quieter problems may matter more. One is sycophancy, where a model tells you what you want to hear and fails to challenge a flawed premise in your question. Another is adversarial input, such as hidden instructions placed in a filing to manipulate AI readers. A Brazilian lawyer was reportedly fined by that country's Supreme Court for doing exactly that.

Practical responses: ask the AI to argue the other side, ask it what assumptions your question rests on, and read documents from opposing parties as a human before feeding them to any tool.

Tools that help you check

The verification market is growing. Georgia Tech researchers recently released an open-source citation checker, Thomson Reuters includes a feature it calls Verify in CoCounsel Legal, and Filevine launched a citator aimed at validating exact passages. See the accompanying news story for details. Tools can speed the pass up. They do not remove the lawyer from it.

Takeaway

Safe AI use is not complicated. Verify everything you rely on, protect client data by default, and put your rules in writing. Firms that do this can adopt new tools faster, because they have already built the habits that make speed safe.

This article is educational and is not legal advice. Check the rules of your jurisdiction and professional body.