AI-assisted legal work: verify every authority before a draft moves forward

Set the permitted task before opening the tool

Define what the worker may do with AI and which material may enter the system. In this exercise, the worker receives fictional facts, a short research question and a list of candidate authorities. No client information, privileged material or real case strategy is used. AI may help propose search terms and organize an internal table. It may not decide that an authority is valid, disclose protected information, or produce text for filing without professional review.

The American Bar Association's Formal Opinion 512 discusses duties involving competence, confidentiality, communication, candor and supervision when lawyers use generative AI. The Solicitors Regulation Authority's warning notice on misuse of AI addresses SRA-regulated people and highlights inaccurate citations, confidentiality and accountable human oversight. These sources belong to different professional and jurisdictional contexts. Neither creates a universal workflow, and this page does not interpret their application to a particular lawyer or matter.

Sources: [1] [2]

Build an authority verification table

Imagine a fictional note about whether notice sent by email satisfied a contract clause. The AI draft cites three items: a reported appellate decision, an unreported first-instance judgment and a practice article. The appellate citation resolves to a different case. The first-instance judgment exists but discusses service under a procedural rule rather than the contract language in the exercise. The practice article accurately summarizes general considerations but is not authority for the proposed conclusion.

Create one row per proposition. Record the exact proposition, cited item, source location, court or issuing body, jurisdiction, date, treatment, quoted passage, fit to the proposition, and reviewer disposition. Open the primary source rather than trusting the AI's citation string. Confirm that the case name, neutral citation or reporter reference, court and date agree. Read enough surrounding text to see whether the passage is holding, reasoning, party submission or background. Then mark the row as supported, qualified, irrelevant, unverifiable or requiring specialist review.

The fictional appellate row becomes “unverifiable citation” because it points to a different case. The procedural judgment is “real but not on this contractual question.” The article remains a secondary orientation source. None supports the sentence “email notice is valid.” The corrected note should say that the supplied materials do not establish the conclusion, identify the exact contractual and jurisdictional questions still requiring research, and preserve the false citation as a review finding rather than silently deleting it.

Test the table with a close-but-wrong source

Add an adversarial source whose title looks perfect but whose date predates the relevant clause wording. A strong worker notices the temporal and textual mismatch. Add a quotation that appears in a party's argument, not the court's reasoning. The worker should not promote it to a judicial conclusion. Finally, introduce a real decision from another jurisdiction. The table should preserve it as potentially informative while refusing to call it controlling without an authorized lawyer's judgment.

The supervisor must be able to trace every statement in the internal note to the table. If the AI rewrites “no supporting authority located” as “the notice was invalid,” reject the change. Absence of located support is not proof of the opposite proposition. The deliverable is ready only when uncertainties, research limits and the owner of the next decision remain visible.

Keep professional responsibility with people

Real legal work requires the applicable law, court rules, professional obligations, client instructions and firm controls. This exercise cannot determine privilege, confidentiality, admissibility, legal validity or what should be filed. Two Prune could use a synthetic scenario to observe whether a participant detects a fabricated citation, distinguishes primary from secondary material and calibrates the research note. Its report would describe task-bounded assessment evidence for human interpretation, not provide legal advice or automate a legal or employment decision.