The Lawyers Strike Back

The predictions that AI will kill the legal profession this year may have underestimated the power of their pens.

Shortly before Christmas, an article was published titled “AI Will Kill All The Lawyers.” A few days ago, Matt Shumer’s viral article on X repeated the claim that the job-apocalypse hitting software developers is coming for the lawyers next.

Later that same day, the humans struck back using tools that AI was supposed to make obsolete: paperwork and bureaucracy. The tools in question were a lawsuit and a ruling from the bench by Judge Jed Rakoff. If this ruling holds, the bureaucracy that is the legal profession will protect lawyers against the AI threat to their jobs, discouraging clients from using AI to replace workers without disrupting our ability as lawyers to use those same tools.

Before we get to the ruling, first the facts of the case: Bradley Heppner was arrested and charged with securities and wire fraud late last year. His mansion was searched and his devices were seized. On those devices, federal agents found approximately thirty-one documents generated by Claude. It turns out that, after Heppner found out he was going to get busted and hired a lawyer, he used Claude to prepare research reports evaluating his legal defenses.

The prosecutors moved to admit those thirty-one documents into evidence. Heppner’s lawyers objecting, claiming attorney-client privilege over the documents. They argued that the documents Claude generated were protected because they included information Heppner got from his lawyers and that he created them specifically for organizing his thoughts to communicate with counsel. They also invoked the attorney work product doctrine, arguing that a report prepared by a client in anticipation of litigation should be protected even if not done at counsel’s direction.

The government’s motion attacked both claims. On attorney-client privilege, the prosecutors argued, first, that Claude is not an attorney, that Anthropic’s terms of service expressly disclaim any attorney-client relationship, and that the documents were not confidential because Anthropic’s privacy policy permits use of prompts for model training and disclosure to governmental authorities. On work product, the government pointed out that Heppner’s own counsel conceded they were not the ones prompting Claude. Heppner, their client, did it on his own initiative and then shared the outputs with his laywers after the fact. So Claude’s generated outputs cannot be the attorneys’ work product.

Judge Rakoff agreed with the prosecutors and held that the generated documents are not protected. On attorney-client privilege, he held that Heppner had disclosed the information to a third party, Claude, which had an express provision that submissions were not confidential. On work product, he held that Claude’s reports did not reflect the legal strategy of Heppner’s counsel. Since neither Heppner nor Claude are licensed attorneys, and Heppner was not working at his counsel’s direction, the materials were unprotected.

(Rakoff has not yet issued a written opinion, but the government’s motion, which he granted, can be found here.)

Now step back and consider what this means for all the doom-saying about professional work. Claude may have been able to pass the bar exam back in 2023, but as of today, Claude cannot be a registered attorney. It doesn’t have a bar number, it isn’t an officer of the court, it isn’t bound by duties of confidentiality.

Anything you say to Claude can and will be used against you in a court of law.

There’s an irony in the fact that the very paperwork and bureaucracy that Claude was supposed to render obsolete is what makes the attorney-client relationship legally special in the first place. The credential, the license, the oath, the regulatory apparatus, is what creates the privilege.

And this isn’t unique to law. Every licensed profession has the same structural defenses baked in. Your accountant’s work papers are protected under Section 7525. Your therapist’s notes are shielded by psychotherapist-patient privilege. Your doctor’s communications with you are covered by HIPAA and state evidentiary rules. Try replacing any of them with an AI chatbot and then asserting those same protections in court. See how that goes. The protections attach not to the information being generated but to the relationship within which it is generated, that between a client and a licensed human professional. AI can generate the information, but not the relationship, and so, not the privilege.

This doesn’t mean, of course, that Claude won’t transform these professions. It will. Lawyers who use AI will outperform lawyers who don’t, and the same is true for doctors, accountants, and every other knowledge worker. But there’s a difference between a professional using a power tool and a client trying to do the job themselves with that same tool. When the professional uses it, the licensing regime wraps the output in legal protections like privilege, confidentiality, malpractice insurance, and fiduciary duty. When the client uses it alone, none of that applies. Heppner just learned this the hard way.

So AI kill all the lawyers? Maybe, but not today. Turns out the lawyers have a secret weapon in this fight that software engineers don’t: a centuries-old regulatory moat, enforced by the courts themselves, that makes the human professional legally necessary for the protections that clients actually need. AI can write the memo, but only a lawyer can make it privileged. And if Judge Rakoff’s ruling is any indication, the lawyers are in no rush to change that.