Verification in Legal AI Is a Design Problem
Lawyers remain responsible. Legal AI should make its work easy to check before they sign.
What Legal Professionals Need to Know About AI and How Not To Use It
A practical guide to task selection, client data, source verification, closed-matter testing and lawyer accountability.
Generative AI can assist with research, document review, factual extraction and drafting. The risk depends on the task, the material provided to the system and how the result is reviewed.
A useful firm policy should be more specific than “do not rely on AI”. It should tell lawyers which environments are approved, which tasks are suitable and what must be checked before the work is used.
Use AI for defined work
A product is easier to evaluate when the task has known inputs and a reviewable output. Examples include summarizing a specified document set, extracting dates or clauses, locating material relevant to an issue, preparing a first-pass chronology or editing a draft built from lawyer-approved facts.
“Analyze the matter” is too broad to create a reliable process or a fair test.
Protect client material before uploading it
Confirm the terms and controls of the exact product and account being used. The firm should know where data is processed, who can access it, how long it is retained, whether it is used for training and how permissions and deletion work.
An enterprise label does not answer those questions on its own. Apply the firm’s confidentiality, privacy, client and court obligations to the proposed use.
Check more than the existence of a source
A system can cite a real authority and misstate it. It can link to a real matter document and change the date, actor or event taken from it.
Review the proposition against the source, including the surrounding context. The paper discussed in When legal AI changes information you have already verified shows how an error can enter after correct information has already been supplied.
Keep accountability clear
The person signing, filing or advising remains responsible for the work.
The Ninth Circuit’s 2026 order in LNU v. Blanche applied that point to inaccurate filings produced with AI assistance. Existing duties attach when lawyers sign and file the work, regardless of where the error entered the drafting process.
The Federal Court of Australia’s generative AI practice note likewise expects users to understand the technology’s limitations, act consistently with existing obligations and be able to explain how AI was used when required.
Match the product to the task
Legal research, eDiscovery, factual review and general drafting products solve different problems. Assess the whole system rather than only the underlying model.
Look at coverage, source access, correction handling, security, integrations and evidence from a relevant benchmark or pilot.
Test on a known matter
A closed matter gives the firm a reference point. Give competing products the same documents and instructions. Measure review time, source accuracy, material omissions, support required and whether lawyer corrections persist.
The known-matter test sets out one way to compare products on work the firm can verify.
Uses to avoid
- Do not use an unapproved environment for confidential client material.
- Do not file, send or rely on generated work before the responsible lawyer has reviewed it.
- Do not treat a citation, confident tone or polished draft as evidence that the underlying work is correct.
- Do not use a product for a task simply because it can produce an answer. Confirm that the system was designed and tested for that work.
For a more operational starting point, see Getting Started (Safely): Six Tips for Attorneys.