The Exam Is Closed-Book. The Profession Is Not.

The apparent contradiction is the point

Reuters reported on September 21 that at least a dozen U.S. law schools revised their artificial-intelligence policies over the summer. Some restricted laptops and phones. Some expanded or required AI instruction. Some did both. The University of Georgia, for example, adopted an analog-first classroom default while also requiring every student to complete at least one course involving artificial intelligence or related technology before graduation.

That combination is easy to caricature as institutional indecision. It is better understood as a claim about sequence. Students may need periods in which they cannot outsource recall, synthesis, or oral responsiveness. They also need supervised opportunities to use tools that will shape research, drafting, discovery, due diligence, and client service. A school that provides only the first experience risks training for a workplace that no longer exists. A school that provides only the second may never learn whether the student can recognize a bad answer without the machine that produced it.

The harder problem is not choosing between paper and software. It is identifying the human capability an exercise is supposed to develop or reveal, then setting the tool rule accordingly. A device ban without a stated learning objective becomes theater. Unlimited use without one becomes an authorship problem.

Law school is doing three jobs at once

The policy debate becomes clearer when legal education is divided into three functions. The first is formation: learning to read closely, extract a rule, hold competing arguments in mind, answer an unexpected question, and decide what matters. The second is assessment: giving a professor credible evidence that a particular student can perform the work being graded. The third is professional simulation: practicing how legal work is actually performed under duties of competence, confidentiality, candor, supervision, and accountability.

Those functions overlap, but they are not interchangeable. A closed examination may be an imperfect but legitimate test of unaided issue spotting and organization. A research assignment that prohibits every AI-assisted search may test something the student will rarely be asked to do in practice. A clinic exercise that invites a consumer chatbot to process client facts may simulate modern practice badly because the professional constraint is not merely accuracy. It is also whether confidential information may be disclosed to that system at all.

One policy cannot resolve all three jobs. Course-level variation is therefore not necessarily a defect. The defect appears when students must guess what a professor is testing, when substantially similar work receives inconsistent treatment without explanation, or when a school invokes academic integrity without distinguishing prohibited substitution from permitted assistance.

Columbia drew the line around intellectual responsibility

Columbia Law School's policy, effective August 1, offers one of the more developed attempts to name the boundary. By default, students may use AI to learn concepts, summarize their own materials, generate hypotheticals, identify potential authorities, test arguments, solicit criticism, explore alternative phrasing, and correct grammar. They may not use it to perform the legal analysis or judgment that an assignment is designed to assess. Examination answers must be produced without AI unless an instructor expressly provides otherwise.

The policy's most important phrase is not a list of approved tools. It is the requirement that submitted work remain student-composed and that the student be able to explain and defend every material proposition. Editing machine-generated prose does not automatically convert it into the student's own work. Disclosure is required when AI contributes to a submission, but disclosure alone does not cure intellectual substitution.

That is a sensible principle, but it is not self-executing. The line between testing an argument and originating one can be thin. Asking a model for five objections may expose the decisive weakness the student never saw. Requesting an alternative articulation may supply the structure that makes the analysis work. The better the tools become, the less useful it will be to classify conduct by verbs such as brainstorm, edit, or draft. The more durable inquiry is counterfactual: what capacity would this assignment still demonstrate if the tool supplied the most valuable move?

A policy can govern conduct without measuring learning

The public Law School AI Policy Archive now tracks policies and curricula across 180 schools. Its categories show how quickly the question has fragmented: studying and learning, submitted work, exams, disclosure, uploading course materials, mandatory instruction, courses on the law of AI, courses on the technology of law, and faculty support. Counting policies can show institutional movement. It cannot show whether students are becoming better lawyers.

A prohibition is relatively easy to announce and difficult to validate. AI detectors remain an unreliable foundation for serious academic judgments. Prompt logs can document one workflow while missing another. Oral defense can reveal whether a student understands a paper, but understanding after the fact is not identical to having produced the analysis. Permissive rules face the mirror-image problem: students can comply with disclosure requirements while allowing the tool to perform so much cognitive work that the grade stops measuring the student.

Schools will therefore need assessments that make the process observable without turning every course into a surveillance exercise. Sometimes the answer will be a supervised, unaided performance. Sometimes it will be an AI-assisted task paired with source verification and an explanation of what the model got wrong. The important design choice comes before enforcement: decide whether the school is measuring recall, reasoning, research judgment, tool supervision, written expression, or some combination. Only then can it define what assistance changes the meaning of the result.

The professional rule is competence, not technological enthusiasm

The ABA's Formal Opinion 512 does not tell lawyers to use generative AI. It says that lawyers who use it must understand its capabilities and limitations sufficiently to provide competent representation. The opinion also connects tool use to confidentiality, client communication, supervision, candor, meritorious claims, and reasonable fees. Fluency without judgment is not competence. Neither is principled refusal to understand a technology that colleagues, clients, opposing counsel, vendors, and courts may already use.

That makes the classroom disagreement more consequential than a dispute over note-taking. A graduate who has never used an AI research tool may not recognize when a model has obscured a weak source or fabricated an authority. A graduate who has used one for every synthesis exercise may have trouble building an argument when the system is unavailable, restricted, or confidently wrong. Both vulnerabilities will be inherited by employers.

Law firms should expect uneven preparation. A transcript will not reveal whether a student attended a school with mandatory AI instruction, a restrictive default, professor-by-professor rules, or a sophisticated combination. New-lawyer training has to establish the firm's own baseline: approved systems, prohibited data, verification standards, disclosure expectations, supervision, and the human work that may never be delegated. The degree certifies a legal education. It does not certify one shared AI curriculum.

The right question is what the work is meant to prove

The most defensible law-school policies will look inconsistent from a distance. They may require a student to answer a cold call without a screen, write an examination without AI, use an approved system in a research simulation, disclose the process, verify every authority, and defend the final judgment in person. Those rules do not express one attitude toward technology. They express different judgments about evidence of learning.

The same distinction belongs in practice. Lawyers work in open systems with research platforms, colleagues, templates, experts, and increasingly AI. Yet the professional obligation still attaches to the lawyer who chooses the argument, protects the information, signs the filing, advises the client, and charges the fee. Legal education should neither pretend that competent lawyers work alone nor graduate students who cannot work independently when independence is required.

The laptop ban is not an AI policy. The chatbot permission is not an AI curriculum. Each is only a tool rule. The educational question underneath is harder and more useful: after the assignment is complete, what does the result allow a professor, an employer, a client, or a court to know about the person who produced it?

This article analyzes current law-school policies and professional-responsibility guidance. It does not provide individualized legal advice, evaluate any student's compliance, or suggest that one classroom policy is appropriate for every course or institution.