Why Our Firm Still Prohibits Generative AI for Legal Research and Written Advocacy

Two years ago, at the request of Carlton Fields’ CEO, I published a short article titled “Why Our Law Firm Bans Generative AI for Research and Writing.” As always, a lot has happened since then.

More hype, more pressure. Can a hundred big law firms be wrong? Of course they can. Many institutional ideas—from DEI policies to climate alarmism, virtual offices, and debates over pronouns—are now being reconsidered. The enthusiasm for generative AI may prove to be another, especially for applications with little tolerance for errors.

A lot of claims about AI use in law firms are just not true. “Clients are demanding that firms use our product.” I review every set of outside counsel guidelines we get from the sophisticated clients that issue them. We have hundreds of them. The overwhelming majority require that if we use a generative AI product, we (a) identify the app, (b) get written permission, (c) justify the use case, (d) explain how we guarantee against the inherent problems of generative AI, (e) describe human supervision, (f) estimate any cost savings to the client, (g) bear responsibility if anything goes wrong, and various other things. Many clients appear to distrust it as much as I do.

“Your competitors are ‘all in’ on AI.” Many say they are, but to the extent we can find their actual use of generative AI, it consists of back-office uses, not legal research and advocacy.

There are daily reports in the legal press of fake cases, false quotations, and generated facts in legal briefs, expert reports, and judicial opinions involving generative AI tools.

Courts now regularly announce local court rules requiring a certificate that the signer has not used generative AI or a certificate that he or she has removed the fabricated portions. But most courts do not seem to get the point that removing the fake parts of a legal brief does not necessarily leave a competent brief—just a recitation assembled by the tool with no easily detectable falsehoods.

We do not advocate against artificial intelligence. Once collateral damage is accepted, (a) AI weapons may be more efficient, (b) self-driving cars may be better drivers than many of the worst human drivers, (c) robot vacuum cleaners save time if you get rid of the puppy and add an AirTag, and (d) phishing emails are much improved.

Disciples of generative AI tell us that we need to learn how to prompt. We need new skills to have a machine create a product that we can then use old skills to check for errors, omissions and fabrications.

I have reviewed the “conversations” between a skilled prompter and the generative AI tools. Yes, the next prompt, “Provide specific authority for proposition 3,” may get a programmed “You got me. I found no direct support for that proposition. I inferred it from a vast array of material that does not say that.” Every iteration in which the prompter attempts to “cross-examine” the tool still produces fabricated answers. A great cross-examination of a liar does not yield a trustworthy person. No more should a well prompted machine be deemed trustworthy.

The machine is still impressive, but skillful prompting does not save work or time, and confidence in it seems misplaced in work with no tolerance for error.

Generative AI has its uses. But judgment as to risk tolerance is necessary, and it depends on the application. For example: Assume a customer-facing chatbot is the application. If the typical 50% customer frustration level enables a cost savings of $X in employee cost and a 10% drop in sales, but results in a 5% increase in profit, a business may think generative AI is a godsend. But for legal research, reasoning, writing, and advocacy, an entirely different set of tolerances applies.

As important, it does not appear that the advocates for generative AI actually understand what lawyers are sworn to do. Lawyers deal with specific problems for a specific client in a specific set of disputed facts. It is not a matter of finding a list of statutes, rules, or cases. In litigation, the facts are of primary importance, and they are often nuanced and in dispute, sometimes by honest witnesses who saw, heard, remember, or understood things differently. And what a case holds depends on the facts of the case, rather than a rule. The entire common law is built on what a judge concludes is the right thing to do in each fact situation, and then whether a similar situation in a following case is similar enough or different enough to cause a judge to rule the same way or differently. Researching the law is largely a matter of understanding the context of each prior case.

Before generative AI, a lawyer dealing with an area of the law in which the lawyer did not already specialize would read an encyclopedia section or a treatise to get the history and context in which the cases were decided, plus maybe another area of law on which that area was built (e.g., Brokers, then Agency). There are arguable areas of treatises and encyclopedias, and there are errors in West Keynotes, but they were honest disagreements or honest mistakes. With the context from the basic research in mind, the facts (and all versions of the facts) could be examined to craft various persuasive reasons why one side should prevail in a dispute, or contract provisions that would advantage the interests of or minimize risks to a client.

Then, the lawyer could search for cases supporting those various persuasive reasons. Whether a case supported a position or not might be explained by peculiar facts, or might have been disagreed with by a subsequent court. The unhelpful cases would be analyzed and saved if they might need to be distinguished or discussed later, but the helpful ones would be tracked for subsequent history or for others factually closer to the case in hand. Legal research and advocacy are hard and imaginative work.

There is a difference between those drawbacks and output fabricated by tools that are agnostic to truth but instead made to assemble believable stories. By starting with generative AI, (a) the lawyer has avoided the important step of educating himself or herself in the subject matter. The lawyer is therefore less equipped to have a ready response to a judge’s questions about fundamentals; (b) the lawyer can be sure that an answer fabricated by AI has not “considered” all those steps and others necessary to develop a complete, cogent, supportive and forthright argument; and (c) to “verify” adequately the fabricated effort, the lawyer has to do the work again from scratch. Just checking for fake cases is not the same, and it is not enough.

If you do as the rules and legal opinions require and verify that your AI assistant is telling the truth, you must do traditional research, spend more time, not less, to provide accuracy. In fact, the time AI “saves” misleads, or has the potential to mislead, and may put you off track rather than giving you helpful background as a starting point.

If lawyers allow themselves to use a generative product to substitute for background research to gain context, how many will have the discipline to double-check the output? After all, they are “only using it for background.” They will not be citing fake cases from the background. But they risk the understanding and inspiration that they would have received from true research. A person tends to believe what they are first told. I once heard a five-year-old explain to his buddy that “dogs pee through their toes,” having learned that nugget by asking, “Hey, Dad! What is that dog doing on that fire hydrant?” I consider that an important parable. If it is not yet in the Bible, I expect generative AI will eventually put it there.

Businessmen do not hire lawyers because lawyers are smarter. Clients and their very smart engineers, financial folks, and marketing people have a problem that they have not been able to answer. They do not want “an answer,” and neither the judge nor the client wants “a brief.” Instead, they hope the lawyer will use his training, knowledge, investigation, research, and the expertise of others to find a cogent resolution of the problem or an authoritative, honest, candid and persuasive argument, if there is no clear answer. The deception of AI is that it can provide “an answer” or “a brief.” It cannot provide what is needed to help a client or a court reach a fully informed decision, and it cannot educate a lawyer to the extent immersive research can. The risk is that lawyers may be tempted by the hype, or will trust the plausible answer the machine creates, and will not do the hard and creative work they were trained for and have sworn to do. Clients can ask generative AI questions themselves. They need more from their lawyers.

AI tools may be useful as a final review of a completed or semi-completed brief. Ask the tool for a redline suggesting changes in grammar, highlighting passive voice, the word “blatant,” and all modifiers. Because it is a redline, the lawyer can remove all but the modifiers that make a difference and decide whether a passive phrase adds emphasis. But that is useful because the lawyer actually knows if any suggestion is good or bad, right or wrong. I really do not know whether that is generative AI. But whether it “generated” the suggestion or not is immaterial because the choices are completely up to the lawyer who knows what he or she needs to say. The tool cannot mislead the lawyer.

Some—many—wanting to compromise between the hype and the commonsense argument that one should not trust a fabricating tool as a source of truth, argue that maybe generative AI is better suited for transactional work—contract drafting, due diligence, coding. But that reasoning is flawed. Nothing makes it less likely that a tool whose nature is to fabricate is not fabricating in transactional tasks. It might be harder to find than a fake case; it might or might not be a more acceptable risk; but it is no more worthy of trust by a novice practitioner who does not immediately know if a suggestion is beneficial or not. I am not competent to talk about the scary things that might happen with hallucinated code. So, I am sticking with generative AI’s dilution of the quality of legal research and advocacy.

Moreover, case law in a common law system itself is fluid. There are no cut and dried answers to legal problems. An advocate might need to dig deeper into the circumstances of her own case to identify key facts buried there that seemed insignificant until the research is done. Good legal research is an iterative process. And as lawyers, we make the law with each new complex case. We help the law evolve. We do not simply take a case off the shelf that fits our case.

Lawyers gain their strength and effectiveness from hard work. We have not been able to persuade ourselves that the so-called shortcut of prompting a machine to produce a more believable answer is a good thing, if it robs the product, the client, the court, and the lawyer’s reputation of the benefit of immersive research into the facts, policies, goals, and history of the problem and the present state of the law.

Reprinted with permission from the March 23, 2026 edition of the Daily Business Review © 2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or [email protected].