What must be verified before an AI recommendation becomes professional judgment?
Legal AI, evidence, and responsibility for the recommendation
Karina Carlos · Self-Conquest
A professional who endorses an AI-assisted recommendation takes responsibility for the advice they give. Checking its facts is essential. So is establishing why those facts justify this action, whose objectives the action serves, and which unresolved assumptions could change the advice. My proposed threshold is that the professional can reconstruct and challenge the recommendation’s governing rationale before asking someone to rely on it.
In the fictional story The Recommendation Was Ready, Mara has eighteen minutes before a settlement offer expires. The dates and contract analysis have been checked. Accepting would end an expensive dispute, but also terminate a distribution agreement on which the client’s service continuity depends. A replacement remains unsigned. The analysis calls the consequence manageable, a description Mara herself supplied. Her associate asks whether the client has actually decided it can lose the service channel. The difficulty is visible in that question: a condition requiring a decision has passed into the recommendation as an apparently settled premise.
The story is invented, including its people and circumstances. It illustrates a question about professional practice; it supplies no evidence about how often this happens or which intervention works. Its useful provocation is that an accurate analysis can still leave the basis for acting incomplete.
What the evidence can establish
Han and colleagues’ CHI 2026 study interviewed 18 non-litigation lawyers. Their accounts describe verification across documents and records, resolving inconsistencies, and assessing provenance. The authors propose traceable and reversible AI-supported work. These are design implications from qualitative, self-reported evidence, not tested interventions. The study helps specify what verification involves; it does not establish that a particular review procedure improves decisions. [1]
A 2026 randomized trial by Schwarcz and colleagues offers an important counterweight. Upper-level law students used Vincent AI, o1-preview, or no AI across six legal tasks. AI assistance improved quality and productivity in task-dependent ways; the NDA drafting task showed no reliable speed or quality improvement. The tested tools and student population limit generalization to professional practice. [2]
An earlier randomized study by Choi, Monahan, and Schwarcz found substantial speed gains with GPT-4, alongside smaller, inconsistent quality improvements. [3] Taken together, these findings call for attention to the task and the tool. They cannot settle whether a particular recommendation is ready for a particular client.
Useful AI assistance raises a practical question: what should the professional verify at the point of reliance? I propose examining the connection between the evidence and the claim, the conditions behind the recommendation, and the authority to accept its consequences. This is a framework for inquiry, not a validated assessment instrument.
Follow the claim back to its support
A citation offers a route to inspection. The reviewer still has to determine what the source supports. A document may establish that a payment is due without establishing that it will arrive in time. A record may be genuine while describing an earlier state of affairs. Several consistent summaries may all depend on the same unconfirmed statement. The relevant question is how much weight the underlying material can bear in this recommendation.
This examination should preserve distinctions among a documented fact, a client’s account, an inference, and a planning assumption. A polished paragraph can make them look equally settled. Restoring those distinctions makes it possible to identify the claim on which the proposed action actually depends, and to direct limited review time toward it.
Recover the condition inside the recommendation
Mara’s word manageable hides a condition: service continuity depends on a replacement arrangement that is not yet secured. In the fiction, the missing work is to bring that dependency into the advice. Additional checking of already verified dates would not answer it. A useful review asks what would have to remain true for the recommendation to hold, and what the professional would advise if that condition failed.
Information legitimately excluded from an AI workflow can remain material to professional judgment. The response need not be to upload more. It may require a separate, appropriately protected review that connects the excluded information to the recommendation. The boundary of the tool’s inputs must remain visible to the person taking responsibility for the whole matter.
Locate the decision and its owner
A recommendation can contain several choices: whether an option is financially attractive, whether an operational exposure is tolerable, and whether to seek different terms. The authority to answer them may belong to different people. Before presenting a single conclusion, the professional should make the consequential tradeoff and the unresolved decision legible. This is professional-practice commentary, not a statement of any jurisdiction’s allocation of legal duties. It is not legal advice.
Human Authority connects judgment with legitimate standing, intelligible reasons, and the capacity to intervene. [4] Applied here, reasoning ownership means being able to defend, revise, or withhold the recommendation on grounds the professional has examined. It allows assistance, consultation, and written reconstruction. An eloquent explanation alone cannot demonstrate that the relevant evidence was considered or that the speaker has authority to decide.
Make the threshold open to study
The Access Problem asks whether relevant competence can be deployed under the conditions of the moment. [5] For this article, the practical implication is to examine review conditions alongside reviewer capability: available time, access to material context, and a workable path for raising unresolved issues. The fiction makes these pressures visible without diagnosing a real person or proving that delay produces a better outcome.
For legal educators and researchers, the next question is empirical. Do source-to-claim checks and explicit assumption reviews improve decisions beyond ordinary accuracy checking? Can participants explain what would change their recommendation, identify who owns the remaining choice, and revise when that condition changes? Any evaluation should measure the added time and the costs of unnecessary escalation as well as errors caught. Professional judgment deserves a standard that can be examined, challenged, and improved with evidence.
Comments