Banning AI Will Not Teach Legal Judgment
July 2026 · VdVLaw
Why verification, confidentiality, and judgment matter more than any classroom ban.
A shortcut wearing a blazer
Law schools are right to worry about AI. If a student uses AI to avoid learning how to read a case, test an argument, identify authority, or write clearly, that is not innovation. That is a shortcut wearing a blazer.
But treating AI like something that can be banned out of existence is not much better. That is the academic version of putting the router in timeout and hoping civilization improves.
The better question is not whether law students should ever use AI. The better question is whether they are being taught to use it without surrendering the one thing the legal profession still has to insist on: judgment.
Banning AI may protect an assignment. It will not teach judgment.
The issue is timely because Berkeley Law's Artificial Intelligence Policy is effective Summer 2026. The policy restricts AI use in exams and most credited coursework, but it also acknowledges the key reality:
future lawyers may need to use AI fluently, while also having the cognitive skills to assess AI work product and meet ethical obligations.
That tension is the whole debate.
Berkeley is right about one important thing. There are moments in legal education when students need to do the work themselves. They need to wrestle with the case, read the statute, write the bad first draft, realize it is bad, fix it, and become slightly less bad over time. That is how legal thinking gets built. It is also how humility enters the building, usually late and carrying coffee.
But denying that something exists does not create discipline. It creates a lack of understanding.
A ban may stop a student from using AI on a particular exam. It will not teach that student how to evaluate an AI-generated answer when a client, employer, discovery platform, legal research tool, opposing counsel, or co-counsel uses it later.
The real lesson is not being lazy
That is what I am teaching my 16-year-old son. Use AI responsibly. Make it give you sources. Then read the sources yourself. Check whether those sources actually say what the AI claims they say. Ask what is missing. Ask whether the conclusion follows from the authority. Do not assume that polished language equals truth.
That last part matters. AI is very good at sounding finished before the work is done. It can produce an answer that looks organized, confident, and adult. The danger is not only that it can be wrong. The danger is that it can be wrong with excellent posture.
David Hume had the cleaner version long before AI learned to hallucinate with confidence:
“A wise man… proportions his belief to the evidence.”
That is still the rule. The tool may produce the answer. The lawyer still has to measure it against the source.
The robot does not absorb the lawyer's responsibility
There is an old computer-age line, often attributed to Bill Vaughan: "To err is human; to really foul things up requires a computer." It is funny because it is still true, except now the computer also adds Bluebook citations and a tone of mild judicial disappointment.
Lawyers do not get to hide behind the posture.
ABA Formal Opinion 512 makes that plain in professional-responsibility language. Lawyers using generative AI still have to consider duties involving competence, confidentiality, communication, supervision, candor to the tribunal, and fees. In normal human language: the robot does not absorb the lawyer's responsibility.
That responsibility also includes knowing what kind of AI environment is being used. There is a real difference between using a vetted business, enterprise, API, or legal-specific account with reviewed privacy terms, admin controls, retention settings, and a default setting that does not use business data for training, and pasting confidential case facts into whatever free public tool happens to be open in another browser tab.
The first can be part of a responsible workflow. The second is how a shortcut puts on a suit and starts calling itself a malpractice exhibit.
Nothing is free
The State Bar of California's 2026 Practical Guidance makes the concern concrete. It explains that generative AI products may use prompts, uploaded documents, or other user inputs to train or refine the tool, may share such information with third parties, and may lack reasonable security even when the product does not use the information for training. California's guidance also says lawyers generally must not input confidential client information into a generative AI system that presents material confidentiality or security risks unless the client gives informed consent to the underlying risks.
That does not mean lawyers should avoid AI. It means lawyers should stop treating the login screen as the ethics analysis.
Before using AI on client material, attorneys need to know whether the tool trains on inputs or outputs, who can access the data, how long data is retained, whether prompts and uploaded files are protected, whether the account is governed by business terms, and whether firm policies and client instructions allow that use.
Florida Bar Ethics Opinion 24-1 is direct on that point: lawyers using generative AI must protect confidentiality and should understand the program's data retention, data sharing, and self-learning policies before using it with client information.
OpenAI, for example, currently states that ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, and the API platform do not use business inputs or outputs for model training by default. That kind of distinction matters. It is not a magic privilege cloak, but it is at least an adult in the room holding a clipboard.
For consumer or free-style AI use, the safer wording is simple: do not assume the tool has the protections a law office needs. OpenAI's current data-use materials say individual services may use content to train models unless the user opts out, while the Data Controls FAQ explains that users can choose whether conversations help improve models. That is exactly why lawyers should understand the settings and terms before putting anything sensitive into a system.
A careless prompt can become a confidentiality problem. A careless upload can become a privilege or work-product fight. A free tool can become the most expensive thing in the case, which is impressive, considering experts and court reporters already entered that competition with confidence.
A source that exists is not enough
The same problem shows up in the research. Stanford HAI summarized legal-AI benchmarking with a blunt headline: legal models still hallucinate in one out of six or more benchmarking queries. The underlying Stanford/RegLab research found that leading legal AI research tools still produced hallucinations at meaningful rates, between 17% and 33% in the tested tasks.
The data is moving, but the lesson is not. Vals AI's October 2025 Legal Research Report found stronger performance on 200 U.S. legal research questions: AI products scored in the 74%-78% range on average weighted scores, with reported accuracy around 78%-81%. That is progress. It is not permission to stop reading.
The exact percentages will keep changing because the tools keep changing. The duty does not. Better AI does not eliminate professional judgment. It makes judgment more important, because the output looks increasingly credible.
That is exactly why source checking is not a cute academic preference. It is a professional survival skill.
Supervise it like a very confident intern
Some law schools are responding by teaching AI rather than pretending it is not already part of practice. Albany Law School announced on June 30, 2026 that it is creating a mandatory legal technology and AI competency course for first-year students, and reporting the next day described the course as focused on both the benefits and risks of AI in the profession.
That approach makes more sense to me. Not because AI should replace legal training, but because legal training now has to include AI judgment.
Future lawyers do not need to be passive consumers of AI output. They need to be the adults in the room. Preferably caffeinated, skeptical, and in possession of the actual source material.
Source material with a drafting history
That is also where this becomes more than a law school issue.
The same discipline applies in discovery review, records analysis, legislative research, witness preparation, and case organization. A tool can help locate, sort, summarize, and surface information. But someone still has to understand what the material actually says.
Someone has to compare the report to the video, the medical record to the timeline, the witness statement to the physical evidence, the statute to the legislative history, and the shiny summary to the theory of the case.
And now there is a newer problem: sometimes the source itself may already be AI-assisted.
If AI helped create the police report, the source is no longer just the source. It is source material with a drafting history.
That is not theoretical in criminal cases. Axon's Draft One uses body-worn camera audio transcripts to generate police-report narratives. EFF has raised concerns about transparency, auditability, and the ability to tell what was written by a human versus software.
California has now enacted SB 524, which requires policies for AI-assisted official reports, disclosure when AI was used, preservation of first drafts and audit trails, and limits on vendor use of law-enforcement data. Utah's S.B. 180 also requires AI-assisted law-enforcement reports to include a disclaimer and requires the author to certify that the report was reviewed for accuracy.
If the source was AI-assisted, verification has to start before the summary.
When lawmakers start writing guardrails, the issue has usually left the conference panel and walked directly into the case file.
That connects directly to the concern I raised in my AI police reports blog: AI-written reports are showing up more. Are they being checked?
For defense work, that question matters because the police report is not just paperwork. It can shape probable cause, charging decisions, plea negotiations, suppression issues, witness preparation, and trial strategy.
Now the verification problem has moved upstream. It is not only, did AI summarize the discovery correctly? It is also, did AI help create the official narrative before the defense ever received it?
Wonderful. Criminal discovery has apparently decided it needed another layer of "hold on, who wrote this?"
There is no slow-motion scene where the PDF finally confesses
That work is not glamorous. No one plays dramatic courtroom music while someone catches a date inconsistency in a 600-page production. There is no slow-motion scene where the PDF finally confesses.
But that is often where the case starts to change.
Where my work lives
VdVLaw is built around that kind of careful legal support: getting into the discovery, records, statutes, timelines, police reports, body-worn camera footage, transcripts, and messy source material with the attorney's actual case theory in mind.
That includes watching for the new problem AI creates: the source material may already have been shaped by software before it ever lands in the defense production.
It also includes knowing when AI is useful and when it is the wrong container for the material. Some information belongs in a secure, reviewed workflow. Some information should be abstracted, anonymized, or checked without feeding confidential facts into a tool at all.
The skill is not just using AI. The skill is knowing what the source material is, what the risks are, what needs human review, and what should never be casually pasted into a box because the box gave a charming answer once.
Attorneys do not need one more confident paragraph. They need someone who can read the source material, understand what the attorney is looking for, flag what does not fit, connect the pieces, and bring back information that can actually be used.
That can mean discovery analysis, records review, legislative research, preparing questions for a licensed investigator, or finding the thing in the production that everyone else was too tired to notice.
A clean folder is nice. A verified fact that changes the case is better.
Probably fine is not a legal standard
That is why the AI conversation matters for legal support too. AI can make sloppy work faster. It can also make careful work more efficient, if the human using it understands the facts, checks the source, and refuses to confuse a polished answer with a supported one.
The future of legal work is not less verification. It is more verification, because there will be more generated material, more summaries, more search results, more transcripts, more video, more data, and more people insisting that whatever came out of the machine is probably fine.
Probably fine is not a legal standard.
It is what people say right before the printer jams.
The question is not whether they will use AI
Banning AI may be appropriate for some assignments. There are places where students need to prove that the work is theirs and that the legal reasoning came from their own brain, not a machine with a subscription plan.
But banning AI cannot be the whole answer. The legal profession is already encountering AI in research, drafting, discovery, client expectations, court filings, police reports, legal support workflows, and opposing counsel's work.
Students should learn what AI can do. They should also learn what it cannot do. They should learn how easily it can invent authority, flatten nuance, miss context, and sound certain when it should be apologizing quietly in the corner.
The question is not whether future lawyers will use AI.
They will.
The question is whether they will use it lazily, or whether they will be trained to use it with skepticism, verification, and judgment.
Because in law, the problem is rarely that nobody can generate words.
The problem is knowing which words are supported, which facts matter, which sources hold up, and which very confident paragraph needs to be escorted gently out of the building.
Sources
S1. Berkeley Law Artificial Intelligence Policy, effective Summer 2026. UC Berkeley School of Law. https://www.law.berkeley.edu/academics/registrar/academic-rules/artificial-intelligence-policy/
S2. ABA Formal Opinion 512 / ABA announcement on generative AI ethics duties. American Bar Association. https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/
S3. Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, 2026 update. State Bar of California COPRAC. https://www.calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf
S4. Florida Bar Ethics Opinion 24-1, Generative AI in legal practice. The Florida Bar. https://www.floridabar.org/etopinions/opinion-24-1/
S5. Business data privacy, security, and compliance. OpenAI. https://openai.com/business-data/
S6. How your data is used to improve model performance. OpenAI. https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/
S7. Data Controls FAQ. OpenAI Help Center. https://help.openai.com/en/articles/7730893-data-controls-faq
S8. AI on Trial: Legal Models Hallucinate in 1 out of 6 or More Benchmarking Queries. Stanford HAI. https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
S9. Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Magesh et al. / arXiv. https://arxiv.org/abs/2405.20362
S9A. VLAIR - Legal Research / Legal Research Report, last updated October 14, 2025. Vals AI. https://www.vals.ai/industry-reports/vlair-10-14-25
S10. Albany Law School Pioneers Mandatory Legal Tech Competency Course for First-Year Students. Albany Law School. https://www.albanylaw.edu/
S11. Albany Law School to require AI training. Times Union. https://www.timesunion.com/education/article/albany-law-school-requiring-ai-course-says-ai-22326584.php
S12. Axon Draft One Is Designed to Defy Transparency. Electronic Frontier Foundation. https://www.eff.org/deeplinks/2025/07/axons-draft-one-designed-defy-transparency
S13. Victory! California Requires Transparency for AI Police Reports. Electronic Frontier Foundation. https://www.eff.org/deeplinks/2025/10/victory-california-requires-transparency-ai-police-reports
S14. California SB 524, chaptered bill text. LegiScan mirror of California chaptered text. https://legiscan.com/CA/text/SB524/id/3272865/California-2025-SB524-Chaptered.html
S15. Utah S.B. 180, enrolled bill text. Utah Legislature. https://le.utah.gov/Session/2025/bills/enrolled/SB0180.pdf
S16. Rule 7.1 Communications Concerning a Lawyer’s Services. State Bar of California. https://www.calbar.ca.gov/sites/default/files/portals/0/documents/rules/Rules-of-Professional-Conduct-7.pdf
S17. Rule 1.6 Confidential Information of a Client. State Bar of California. https://www.calbar.ca.gov/Portals/0/documents/rules/Rule_1.6-Exec_Summary-Redline.pdf
S18. Quote verification for "To err is human...". Quote Investigator. https://quoteinvestigator.com/2010/12/07/foul-computer/
VdVLaw provides attorney-directed legal support services and is not a law firm. Robert van der Vijver is not an attorney. Nothing in this article is legal advice, and nothing here creates an attorney-client relationship. This article is general information written for attorneys and does not address the facts of any particular case. Criminal discovery obligations, public records access, evidentiary issues, retention issues, ethical duties, and law enforcement technology policies can vary by case, agency, county, court, and date. Statutes, rules, policies, AI product terms, and technology features change over time. Attorneys remain solely responsible for their own legal analysis, discovery decisions, ethical duties, litigation strategy, and verification of all statutes, rules, agency policies, discovery responses, and source materials.