AI can widen legal review

AI can make another pass through a large file practical: a chronology, a comparison or a search for evidence that might unsettle a case. That examination must improve the basis for a decision enough to justify checking its findings.

A few decisive passages can be worth more than hundreds of plausible observations. The work is to identify which findings matter, check what supports them and judge what may still have been missed.

Sources and findings

A language model can extract a date, calculate from it or draw an inference about it. Those operations need different checks. A reliable account distinguishes what the source says, what was calculated and what the lawyer concludes.

Take a hypothetical review of whether a notice reached its recipient in time. A model finds a letter dated 14 March, an email attaching it on 16 March and a later reply acknowledging receipt. The letter’s date proves neither sending nor receipt. The acknowledgment’s wording and the applicable legal rule still need examination.

The model also flags a different address in an earlier contract draft. The signed version and the notice use the same address; the apparent discrepancy adds nothing to the timing question. Distinguishing this distraction from the useful lead takes work.

Classifying a document as irrelevant does not establish its irrelevance. The search question itself may have left out what matters.

Broader review

AI can help examine many more documents against the same questions and link possible findings to passages. Record what was included or excluded, when it was collected and which files could not be processed.

Processing every file does not ensure every relevant point is found. A scanned page may be misread, a question framed too narrowly or a contradiction overlooked. Keep disagreements between sources visible; if the evidence cannot settle them, state which account the advice relies on.

When the model supplies the law

A model asked for the law without specified authorities may combine sound propositions with obsolete, irrelevant or invented support. Its tone gives the reader no reliable way to tell them apart.

An error may look entirely ordinary: a plausible citation, a familiar legal test or a proposition that is right in another jurisdiction. It can survive a hurried reading because it fits the expected answer.¹

An unsourced answer may suggest a line of inquiry, but the legal research remains to be done. Even when authorities are supplied, check that they exist, remain relevant and support the proposition in context.

Before using the system

Define what the system should examine, which sources it may use and what the answer must contain. Link every material factual or legal proposition to its sources. Assign review to someone with the time and expertise to check unsupported claims.

Confidentiality, professional secrecy, privilege, data protection and information security constrain the choice of system.² Before entering sensitive material, establish where it will be processed, who can access it and how it will be retained, reused, transferred, logged and secured. Local processing may reduce some risks; access controls, backups and connected services still need assessment. Removing names is insufficient if the remaining details identify the people concerned.

Test the system that will be used

Evaluate the whole process, from retrieval and extraction to the reviewer’s final assessment. Compare it with an appropriate existing method on the same question and source material. Choose representative documents, independently checked reference answers and difficult cases before running the test.

Measure at least:

• relevant material missed;
• statements unsupported by the specified sources;
• citations that do not exist or do not support the proposition;
• answers that stray beyond the question or permitted sources;
• the time needed to reach a reliable result, including corrections and the pursuit of false leads.

Repeat the tests when changes to models, settings or connected sources could affect the result, and investigate failures found in use. Success on one task does not establish reliability on another.

Responsibility

The lawyer must decide whether the findings justify the advice, what could change it and whether more work is needed. Stopping is part of that judgment: further review may be unlikely to affect the decision, or another way of obtaining evidence may be more useful.

The client should receive a considered answer and the uncertainties that matter. The supporting record should let another competent person trace the reasoning and question its key inferences, without asking the client to repeat the lawyer’s work.