How to Design AI Rules for Assessed Legal Work

A rule about artificial intelligence in legal education should begin with a plainer question than whether AI is good or bad: what ability is this assignment meant to show? If a memo is meant to reveal whether a student can identify issues, weigh authority, and build a defensible argument, an assistant that supplies the structure or prose can hide the very work the course is evaluating. If an exercise is designed to teach careful supervision of AI output, a blanket prohibition can hide a different essential skill.

Berkeley Law's default policy, effective in summer 2026, makes that distinction unusually explicit. For work submitted for credit, it prohibits AI assistance with conceptualizing, outlining, drafting, revising, translating, and editing. It also prohibits AI in examinations, restricts AI research to identifying sources, and bars students from uploading course materials to generative systems. The same policy leaves instructors room to provide written, advance alternatives that fit a course and require disclosure.

The useful lesson is not that every legal assignment needs the same boundary. It is that the boundary needs to identify the competence being measured, the data that must remain protected, and the evidence a student should retain to show how a tool was used.

Define the assessed skill before naming a tool

A policy that simply says "no ChatGPT" ages badly and gives weak guidance for adjacent tools. Search engines, writing applications, legal research services, transcription products, and document-review platforms are increasingly adding generative features. The educational question is not the product label; it is whether the feature substitutes for a student decision the assignment is supposed to assess.

For an exam testing legal analysis, the protected work may include selecting relevant facts, identifying a rule, distinguishing precedent, and expressing a conclusion under time pressure. A rule can prohibit generative assistance across that workflow without claiming that all AI is inappropriate everywhere. For a research-skills assignment, the relevant boundary may be narrower: an AI system might help locate potential cases, but the student must read the authorities, verify citations, and decide whether they support the proposition.

This approach also makes assessment more defensible. Students can see what the course is asking them to practice, while instructors can explain why an otherwise convenient capability is limited in one setting and permitted in another.

Use a four-part permission model

A practical course policy can sort AI use into four categories. First, identify work that must be independently performed: for example, exam answers, a first legal analysis, or an initial outline. Second, identify bounded assistance that may be allowed, such as locating publicly available sources, generating practice questions, or critiquing a completed draft. Third, state when a student must obtain written permission before using a tool. Fourth, state what information must never be uploaded, including client-like hypotheticals, unpublished course materials, recordings, or documents containing personal data.

The categories should use observable verbs. "May use AI for research" is ambiguous. "May use AI to suggest search terms; may not rely on an AI-produced case summary or citation without reading and verifying the cited source" tells a student what to do. It also gives the instructor a clearer basis for feedback than a vague instruction to use AI responsibly.

Berkeley's distinction between source identification and submitted work illustrates the value of this precision. Finding a case is not the same educational act as deciding what the case means. A policy can permit the first in a defined context while still requiring the student to perform the second.

Make approved use reviewable, not invisible

Disclosure is most useful when it records enough context to evaluate a process without demanding a diary of every keystroke. A short declaration can name the tool, the permitted purpose, the material provided to it, and how the student verified any result used in the assignment. For higher-stakes projects, a prompt log or a version history may be appropriate.

That record changes the educational conversation. Instead of trying to infer authorship only from polished prose, an instructor can ask why a source was selected, why an AI suggestion was accepted or rejected, and how the student checked it. Oral follow-ups, staged drafts, and iterative assignments can test reasoning directly. Berkeley's own policy Q&A points to context-specific enforcement and assessment techniques such as oral presentations and iterative writing.

Disclosure should not turn into a presumption that every use is misconduct. It is a way to distinguish authorized assistance from undisclosed substitution and to give students a chance to demonstrate judgment.

Treat privacy and professional responsibility as part of the skill

Legal work often contains information that should not be sent to a public generative service. Even in a classroom, readings, assignments, recorded discussion, and hypothetical client facts can carry confidentiality, intellectual-property, or privacy concerns. A strong course policy names those limits separately from the academic-integrity rule so students understand that a tool can be inappropriate even when it would not write the final answer.

That separation mirrors legal practice. The American Bar Association has said that existing professional obligations continue to apply when lawyers use generative AI. Students preparing for that environment need practice with more than output quality: they need to ask what data a service receives, what a generated answer actually supports, and who remains responsible for an error.

Revisit the rule through assignments, not slogans

An effective AI policy is a teaching design that can be tested. After a term, instructors can review whether students understood the permitted uses, whether the rule worked for accessibility needs, where common misunderstandings arose, and whether the assignment still measured the intended skill. The answer may be a more specific example, a better disclosure form, or a different assessment—not simply a stricter ban.

The central principle is durable: legal education should protect independent reasoning where reasoning is the object of assessment, and it should teach supervised AI use where that supervision is the object of instruction. Clear boundaries, explicit disclosure, verified sources, and protected data make both goals easier to pursue.