AI Is Forcing Law Schools to Defend the Mind

The Policy Is Not the Story

Berkeley Law has done something that many institutions will eventually have to do in their own way. It has drawn a line between using AI as a professional tool and using AI before the professional mind has been trained to supervise it.

Beginning in summer 2026, Berkeley Law’s artificial intelligence policy restricts students by default from using AI for a broad set of academic tasks connected to credited work. The policy is not limited to obvious drafting abuse. It reaches activities that many people now treat as harmless support, including brainstorming, outlining, translating, editing, grammar correction, legal-rule summaries, and exam preparation connected to work submitted for credit.

That is why the reaction has been sharp. Critics see the rule as too restrictive for a profession already adopting AI. They argue that law students should not be trained for a market that no longer exists. Legal employers are experimenting with AI tools, clients are asking about efficiency, and young lawyers will increasingly be expected to understand how these systems operate.

That criticism is reasonable, but incomplete.

The more interesting question is not whether future lawyers will use AI. They will. The real question is when AI use strengthens professional capability and when it quietly weakens the development of the judgment that professional work requires.

Berkeley’s policy turns that question into an institutional decision. It treats AI not only as a classroom integrity problem, but as a training problem. That is the part serious business leaders should pay attention to.

Professional Judgment Is Not a Software Feature

Legal education is not simply a process for producing memos, briefs, outlines, and exam answers. It is a training system for developing professional judgment under constraint.

A good lawyer has to identify the issue, understand the authority, distinguish relevant facts from noise, notice weak analogies, test the opposing argument, and recognize when a confident answer is legally unsound. That work happens before the polished sentence appears. The final document is only the visible artifact.

AI disrupts that sequence because it is very good at producing the artifact before the learner has completed the underlying work.

A student who uses AI to brainstorm a paper topic may be outsourcing the first act of judgment: deciding what is worth asking. A student who uses AI to summarize a legal rule may receive a clean formulation without learning why the rule is unstable, contested, narrow, or dependent on context. A student who uses AI for grammar correction may believe the tool is touching only style, while the system may also alter meaning, emphasis, tone, or legal precision.

Those are not small issues in legal training. They sit close to the core of the craft.

The policy debate should not be reduced to whether AI helps students work faster. Of course it can. The stronger concern is whether speed arrives before the student can evaluate what has been accelerated.

The Market Wants AI Fluency, But Fluency Is Not Readiness

The legal market is moving toward AI. That fact cannot be ignored. Law firms, in-house legal teams, legal research providers, courts, regulators, and clients are all confronting AI-enabled workflows. A law graduate who has no understanding of AI will be poorly prepared for the profession.

But AI familiarity and AI readiness are not the same thing.

A student can be taught how AI systems work, where they fail, how hallucinations appear, how confidentiality risks arise, and how professional duties apply without being allowed to use AI as the hidden engine behind credited legal work. A school can train AI literacy while still preserving unaided legal reasoning in formative assignments. A profession can demand tool competence without allowing the tool to replace the cognitive apprenticeship that creates competence.

That is the balance institutions now need to design.

The weakest version of AI adoption says that people should use the tools simply because the tools exist. The stronger version asks what the user must already understand before the tool becomes safe, useful, and professionally defensible.

A senior lawyer using AI to compare two drafts is not in the same position as a first-year law student asking AI to explain doctrine for the first time. A partner using AI to test a counterargument is not doing the same thing as a novice using AI to create the initial argument. The same tool can support expertise in one context and conceal the absence of expertise in another.

Legal AI Failures Have Already Shown the Cost

The legal profession has already seen what happens when fluent output is mistaken for reliable work. Courts have sanctioned lawyers for filings that included nonexistent cases, fabricated quotations, and false citations generated or supported by AI tools. The most prominent early example was Mata v. Avianca, where lawyers submitted fake authorities after relying on ChatGPT. More recent cases have continued to show that legal AI misuse is not a one-off embarrassment.

These incidents are often described as hallucination problems. That description is too narrow. The deeper failure is supervisory. The lawyer remains responsible for the work product, the verification process, and the professional consequences of submitting unreliable material. The model does not owe duties to the client. The software does not appear before the court. The interface does not carry the sanction. The professional does.

That responsibility requires more than a warning label. A lawyer has to know how to verify, what to verify, and when a plausible answer should trigger suspicion. Verification is not a final administrative step that can be bolted onto AI use. It is a habit built through training.

This is where Berkeley’s stricter default becomes strategically important. The policy suggests that professional institutions cannot wait until practice begins to teach skepticism. They have to preserve the conditions under which skepticism develops.

The Business Lesson Is Bigger Than Law School

The same pattern is appearing across business. Companies are moving AI into knowledge work at speed. Employees are using AI to draft emails, prepare presentations, summarize documents, analyze research, write code, review contracts, generate proposals, and produce internal strategy material. In many cases, the work looks better than before. It is cleaner, faster, and more fluent.

That improvement can be real. It can also hide a capability problem.

If junior employees use AI to avoid the difficult early stages of work, managers may see polished output while the organization’s training pipeline deteriorates. The first draft no longer reveals how the employee thinks. The summary no longer shows whether the employee understood the source material. The recommendation no longer exposes the assumptions behind the conclusion.

For leaders, that creates a governance problem that is easy to miss. AI can make weak reasoning less visible. It can help inexperienced employees sound more capable than they are. It can compress the messy process through which managers used to diagnose gaps, coach judgment, and develop talent.

A company may gain short-term efficiency while weakening the human capability it will need when the system is wrong, the context changes, the client challenges the recommendation, or the stakes rise.

That is not an argument against AI adoption. It is an argument against adoption without formation logic.

Governance Has to Include the Training Path

Most AI governance programs still focus on tool approval, data exposure, cybersecurity, disclosure, model risk, and output review. Those controls are necessary. They are not sufficient.

A mature AI governance model has to ask how AI changes learning, apprenticeship, supervision, and accountability. It has to distinguish between production use and training use. It has to define which tasks may be AI-assisted, which tasks require unaided work, and which tasks should be allowed only after the user has demonstrated baseline competence.

This is especially important in roles where junior employees are expected to become future experts. Law is the clearest example because the consequences of bad reasoning can be severe, but the pattern applies to consulting, finance, engineering, medicine, research, compliance, and management.

In each of those fields, professional judgment is not developed by reading finished answers. It is developed by working through uncertainty under supervision. The person has to struggle with incomplete information, make a provisional judgment, receive correction, revise the work, and learn what quality looks like from the inside.

AI can assist that process when it is used deliberately. It can generate practice questions, expose alternative arguments, test assumptions, and help trained users compare approaches. But when it replaces the early cognitive work, it can convert education into output production.

That is a bad trade for any institution that depends on future judgment.

The New Question Is When, Not Whether

The AI debate is often framed as a fight between prohibition and adoption. That framing is too crude for serious institutions.

The better question is when people should be allowed to use AI for particular forms of work. The answer should depend on the user’s level of competence, the consequence of error, the availability of supervision, the task’s role in training, and the organization’s ability to verify the output.

For a law student, unaided analysis may be essential in some assignments because the assignment is not merely testing the final answer. It is training the underlying reasoning. For a practicing lawyer, AI assistance may be acceptable or even valuable when the lawyer has the expertise to supervise the output and the process includes verification. For a business team, AI may be helpful in execution once the strategic judgment is owned by people who can defend it.

The point is not to preserve difficulty for its own sake. The point is to preserve the kinds of difficulty that build capability.

Institutions that understand this will avoid two mistakes. They will not pretend AI can be kept out of professional work. They will also not pretend that access to AI automatically makes people more competent. The first mistake is denial. The second is operational laziness disguised as innovation.

What Serious Organizations Should Watch

Berkeley Law’s policy signals where the next phase of AI governance is headed. The debate will move from whether AI may be used to whether the user is ready to use it responsibly.

For executives, that shift has practical consequences. AI policy cannot stop at acceptable-use language. Training programs need to preserve moments where employees must reason without machine support. Managers need to see how people think before AI polishes the result. Review processes need to assess not only whether the output is accurate, but whether the person responsible understands and can defend it.

This will become harder as AI is embedded into ordinary software. The boundary between “using AI” and “using a normal work tool” will blur. That makes the formation question more urgent, not less. If AI assistance becomes ambient, institutions will need clearer judgment about where unaided work remains necessary.

Berkeley’s policy may evolve. It may prove too strict in some contexts. It may need exceptions, revisions, and more granular implementation. But the underlying concern is not reactionary. It is one of the most important governance questions in professional work.

Before a person uses AI to accelerate judgment, what judgment must already exist?

That question belongs in law schools. It also belongs in boardrooms, consulting firms, banks, hospitals, engineering teams, and every organization building AI into skilled work.

AI will become part of professional practice. The institutions that handle it well will not be the ones that adopt the fastest or restrict the most. They will be the ones that understand sequence.

First, build the judgment. Then, accelerate the work.