I’m a practicing attorney, and I want my clients to use AI.
That might sound counterintuitive coming from someone who earns her keep selling legal advice. Plus, many attorneys express frustration when clients show up with contracts full of unlawful provisions drafted by ChatGPT or implausible case strategies concocted by Gemini and then have to spend time explaining why those won’t work.
But after more than three decades of small law practice where I frequently harness my clients’ sweat equity to fight Big Energy, I see things differently. My clients who use AI to research their legal situation are better clients. They arrive well-organized. They understand the documents I’ve sent so we don’t waste precious billable hours on the basics. They’re fully engaged in the case without monopolizing my time. For clients on a budget especially, AI can be transformational.
A recent federal court opinion threatens to change all of that.
In a memorandum opinion dated February 17, 2026, Judge Jed Rakoff of the Southern District of New York ruled in United States v. Heppner that documents a defendant generated using Claude were not protected by attorney-client privilege or the work-product doctrine. The defendant, Bradley Heppner, charged with securities fraud, had used Claude to research his case and generate detailed legal analyses. When the FBI seized his devices, prosecutors claimed those documents were fair game.
The court agreed, finding that (1) Claude is not an attorney, (2) that the communications were not confidential given Anthropic’s privacy policy allowing disclosure to third parties, and (3) that even assuming the documents were prepared in anticipation of litigation, they were not protected by work-product privilege because they were not prepared by or at the behest of counsel.
The ruling may be defensible under existing doctrine. But it is a disaster for the 21st-century justice system.
To understand why, consider the crisis the American legal system already faces. According to the Legal Services Corporation, roughly 92 percent of low-income Americans receive inadequate or no legal help for civil legal problems. In most civil cases like evictions, debt collections, and custody disputes, at least one party is unrepresented. For these people, AI is a lifeline that Heppner turns into a liability.
The court treated Heppner’s AI conversations as no different than a Google search, which we all know from Court TV is not protected (think of all the convictions that have flowed from queries like “how do I clean blood stains from my carpet?”). But that equivalence isn’t quite right. When you Google “elements of securities fraud,” you generate links to public web pages and factual materials. No new information is created. AI is an entirely different animal. To get anything useful from an AI model, users must feed it specifics like a timeline and perceived vulnerabilities. What comes back is a custom analysis reflecting a user’s mental impressions and developing legal strategy. Produced in discovery, it hands your adversary your strategic calculations and your assessment of where a case is weakest.
By declaring AI research discoverable, Judge Rakoff allowed the government to rely on “wits borrowed from its adversary.” This is exactly what the Supreme Court sought to prevent in the foundational 1947 case Hickman v. Taylor, which established the work product doctrine. The Hickman court recognized that for the adversary system to work, a party must have a “zone of privacy” to prepare their case. Without that privacy, the court warned, “much of what is now put down in writing would remain unwritten,” and “the cause of justice would be poorly served.”
The Heppner decision also rests on a legal fiction about user expectations. The court found Heppner had no reasonable expectation of privacy because the terms of service for Claude stated that data may be disclosed. But the design of AI tools suggests otherwise. The conversational interface, the personalized one-on-one format, and the way the AI chatbots invite users to share their situation cultivates the sense of a confidential consultation. The Heppner ruling expects ordinary people, often at their darkest hour, to parse complex terms of service that most lawyers skip, while the product itself beckons with a siren’s song to “tell me everything.”
And the more these tools absorb your facts to sharpen their analysis, the more damaging the trail they leave behind. To follow the Heppner logic to its conclusion rewards ignorance and disempowerment. Do no research, and you have no trail. Try to be a diligent, informed participant in your own legal matter, and you hand your opponent a gift.
What is most aggravating about the Rakoff ruling is how tone-deaf it is to the high cost of legal services and to AI’s potential to reduce those costs. Heppner’s Quinn Emanuel lawyers bill upwards of $3000/hr, so why wouldn’t Heppner try using AI to save a few bucks? The opinion also acknowledges AI’s novelty – but instead of crafting an approach to encourage use of AI to level the playing field, it defaults to relying on an article penned by Ira Robbins, an ivory tower academic who arrives at the mean and utterly unimaginative conclusion that privilege for AI communications is “neither doctrinally justified nor normatively sound.”
Privilege isn’t some type of inherent protection. Some privileges like attorney-client privilege are created by legislatures while others like work product doctrine have been crafted by courts or even ethics regulators. For example, ABA Opinion 477 says that unencrypted email carries with it a sufficient expectation of privacy to confer confidentiality – even though we all know that’s a fiction. So why can’t we just say that an expectation of privacy applies to generative AI and speak a privilege into existence?
Under Rakoff’s opinion, work-product protection arguably survives if a client undertakes AI research at the direction of an attorney. But that only makes life more complicated, requiring lawyers like me to micromanage our clients’ work and remind them like a nagging parent to include the magic words “prepared under lawyer’s direction” every time they enter an AI chatbox. And that narrow exception doesn’t protect pro se litigants or clients who want to do their homework before ever stepping foot into an attorney’s office.
Judge Rakoff’s ruling mechanically applied old rules to a new world. Today, millions of Americans are turning to interactive AI to survive a legal system that has become too expensive and too complex for ordinary people to navigate. As a lawyer, I want my clients and my potential clients to keep using AI tools. The law should encourage them to do so, not punish them for it.
Carolyn Elefant is one of the country’s most recognized advocates for solo and small firm lawyers. She founded MyShingle.com in 2002, the longest-running blog for solo practitioners, where she has published thousands of articles, resources, and guides on starting, running, and growing independent law practices. She is the author of Solo by Choice, widely regarded as the definitive handbook for launching and sustaining a law practice, and has spoken at countless bar events and legal conferences on technology, innovation, and regulatory reform that impacts solos and smalls. Elefant also develops practical tools like the AI Teach-In to help small firms adopt AI and she consistently champions reforms to level the playing field for independent lawyers. Alongside this work, she runs the Law Offices of Carolyn Elefant, a national energy and regulatory practice that handles selective complex, high-stakes matters.
The post AI Research Can Be Used Against Clients In Court. It Shouldn’t Be. appeared first on Above the Law.