Can Your Work Survive Disclosure?

Anthropic started watermarking Claude this week. Every model released after August 2 now embeds an invisible mark in the text it writes, and it travels when you paste the text elsewhere.

Which raises a question the bar has dodged for two years. Do you care if a judge, or opposing counsel, can tell that AI touched your filing?

I asked Eve CEO Jay Madheswaran. His answer was not the comforting one.

The watermark is not the problem. Your work is.

"If your firm and workflow collapses because AI use became visible, it just means you weren't doing a good job," Jay told me. "If your AI can't survive disclosure, that's a problem."

That puts the anxiety where it belongs. The worry this week was about being seen. Being seen only hurts you if the work does not hold up.

Two things to know before anybody panics.

The mark proves almost nothing. Anthropic's own documentation says a hit means the content "may have been processed by Claude," and that "Claude may not be the original author." Proofreading trips it. Translation trips it. A brief the machine drafted and a brief you wrote and asked it to tighten look exactly the same to a detector.

And nobody was thinking about your demand letter when they built it. The driver is the EU AI Act transparency code, and the target is copyrightable IP. An author lost a book deal this month on suspicion alone, an advance north of $2 million, gone.

Your complaint is not copyrightable. And the judge-runs-a-detector scenario everyone is picturing has not happened anywhere I can find. Bar guidance still says the same two things it always has. Check your facts. Review your work.

The lie was in the transcript. Finding it in time was the problem.

The watermark cannot tell you what the AI was actually doing, and that distinction is the whole ballgame.

Meredith Hinton spent six weeks this spring on a North Carolina wall collapse that killed a worker. The man her team alleged was running the job took the stand and told the jury he vividly remembered a phone call, that the workers' employer had signed off on backfilling the wall. Hinton was sure he had sworn the opposite somewhere in three days of deposition. Her team had until the end of lunch to prove it.

She asked her AI for every place he discussed that call. Volume, page, line. Seconds later, no such call. After lunch they walked him back through it, let him hold the line, then played the video of him swearing the opposite. The jury stopped taking notes on what he said and started taking notes on who he was.

$101 million, believed to be the largest personal injury verdict in state history.

Now run the watermark test on that. A detector catches nothing, because the AI never wrote a word that went in front of the jury. It pointed at a page number in sworn testimony, and the testimony did the rest. And when the AI does draft the document, the mark lands on something you signed, reviewed, and can defend line by line. Rule 11 never asked what software you opened. It asked whether you read what you filed.

Read Meredith's full account here.

One more thing

As a young journalist at the Miami Herald and Dallas Morning News, the em dash was a flex. You used one to hold two thoughts in a single sentence and land them both. We put them in on purpose, the way you wear a good tie.

Then the witch trials started. Torches up, evidence optional, the internet went hunting for em dashes door to door.

Readers decided the mark was a tell, that anyone reaching for one had let a machine do the reaching. The Ringer called it a "relentless campaign of online persecution," and pointed out that "any chucklehead who disagrees with an online post can, if that post happens to include an em dash, dismiss it with the claim that it was written by a chatbot." The humor site McSweeney's gave the mark a right of reply: "I'm the cigarette you keep saying you'll quit."

The cruel part is that the em dash is, in fact, "the most human punctuation mark there is." The tool writers use for "not sounding robotic in the first place." Use one now and you are the robot.

I resent that. Nobody had to prove anything, and a mark that used to say you were good at this became evidence you are a fraud. That is detection by vibes, and it is worth remembering the next time somebody squints at your work.

Hit reply and tell me: would you disclose AI use on a filing if nobody made you? I read every one.

~ Adam

PS: Want to be featured in The Tort Report? Hit reply and let’s chat.

PPS: I want the Tort Report to be worthy of your inbox. That means knowing what you actually want to read and where you go for news. Please take this 5-minute survey and tell me. I read every response.

Quote of the week:

“Try your lies. It is the oldest advice in this job. For the first time in my career, I can do it in time to matter.”

- Meredith Hinton, trial attorney at Ricci Law Firm in North Carolina, on using AI in the middle of a trial.