As of this month, content generated by most of Anthropic’s Claude models will include embedded watermarks to label it as AI generated, causing concern that is seeing some vendors and lawyers consider whether they should quit using Claude for drafting.
Anthropic, which has made the move in compliance with the EU AI Act, has explained that generated text will carry embedded watermarks and generated files will include digitally signed provenance metadata where supported.
Watermarks will apply to output from Claude models across the Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, not just in Europe, but worldwide, unless marks are not supported. At its simplest, the marks come from the sequence of words that Claude uses/chooses, which creates a statistical pattern.
Many are pointing out that the fact that this has been rolled out to comply with the EU AI Act also means that other major tech vendors will likely have to follow suit. We’ll come back to that point in a second, but it makes it no less problematic for users, who are considering the impact of new developments on both the way they use AI (and how that may be overstated); what it means for the credibility of the documents they produce; and what it means for the likes of ownership in patent law.
Created versus checked?
One cause of concern is the fact that a lawyer might create an original document that it is checked and slightly modified by AI, and that it will be labelled as AI generated.
Commenting on the watermark roll out on LinkedIn, Mauro Messias, a Stanford Law educated attorney and engineer, said: “The watermark is going to get lawyers in trouble.” He points out that copy-paste and light edits carry the mark, as does translating copy. And he observes that both Harvey and Legora use Claude under the hood – you can’t skip it – observing, “the mark is applied at the model layer, not in the legal app.”
Speaking to Legal IT Insider, Ryan Anderson, CEO and founder of $3bn legal operating system provider Filevine (which has circa 400 engineers that use Claude heavily) is one of those who is deeply concerned. He told us: “I write my team an email on a Sunday night. It’s very personal and I never miss it. Over the past three to four months, I had got Claude to do a grammar pass of it and that saves me time, but the idea that it is now marked as AI writing is kind of offensive and gives people the wrong idea.”
He adds: “We have 400 engineers and you couldn’t pry Claude out of their hands if you tried, I can’t say enough about what Claude Code has done of our industry, but not knowing how much of a work product is produced by AI is a problem.”
Filevine is considering swapping out Claude for drafting.
As to why this matters more specifically to lawyers, Anderson says. “Lawyers are good at coming up with creative arguments, Claude is not. At the end of the day, it’s a token predictor.
“What happens to the attorney who comes up with a novel argument that’s then checked with Claude and the whole thing is marked as AI? The courts are likely to look at it with less credibility. While that may change, I don’t like receiving AI slop and I don’t think opposing counsel like receiving it either.”
“AI is just like using Google?”
There are plenty of people who are quite fairly questioning why anyone in this day and age would and should care if a document is created or touched by AI. One lawyer on LinkedIn observed dismissively that using AI is now like using Google, which is interesting for more reasons than he intended, and takes us back to the point that other tech giants will be following Anthropic.
Posting on LinkedIn immediately after Anthropic’s watermark announcement, AI futurist Cat Casey said in a post well worth reading: “This is not new. Google has quietly had SynthID watermarking for years, OpenAI is already using SynthID on some image and audio output, and Google has opened up the text version for other model makers to use. Anthropic may be the one making noise about it now, but invisible text AI watermarking is already here.”
Lawyers will have to carefully consider the ramifications. If you disclose your invention to an AI service, have you lost your claim of novelty? If this is your area, patent attorney Roger Hahn’s post is interesting. It could also create new privilege and disclosure headaches for law firms.
Whatever your views on watermarking, it’s not nothing. The word you are going to hear a lot more of now in the context of AI is not just governance, but provenance.