Claude is getting a watermark. That could matter for every law firm using AI

It felt like something quite significant happened this week when Anthropic - the AI giant behind Claude - announced it will start adding an invisible watermark to all outputs it generates.

The move is Anthropic's response to new requirements under Article 50(2) of the EU AI Act, which says that providers of AI systems must ensure their outputs are detectable as AI generated.

Despite the legislation only applying in the EU, Anthropic said it will mark Claude’s outputs globally.

The mark is designed to remain embedded even when text is copied and pasted. In Anthropic's words: "It will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from."

Anthropic says the mark on its own will not “confirm the full provenance of the content”, noting that Claude may not have been the original author. It also says that heavy editing could remove the mark.

Many people asked the same question: how does Claude add a watermark that can survive copying and pasting and the removal of all formatting?

The best guess at the moment is that it will be based on so-called statistical watermarking. In simple terms, as Claude predicts the next word to use, it marginally increases the probability of selecting certain words over others. Over a long enough passage, this creates a pattern that Anthropic can detect as a sign the text was generated by Claude.

The other AI giants - OpenAI, Google and others - are likely to make similar moves to comply with the legislation.

And this is where things get interesting for lawyers.

AI leaves a trace

Claude, along with models from the other AI giants, sits underneath many of the legal AI products now being rolled out across the world's top law firms. The obvious question is what happens to the outputs those platforms generate.

Will a contract produced with the help of AI now include a watermark, even where it has since been copied over to Word, been through several rounds of markup and is ready to be signed? What about a client memo or a court document? What happens where several different AI systems have contributed to the same document?

More fundamentally, the question is whether any of this matters.

Arnav Joshi, an AI and technology partner at Ashurst Perkins Coie, says the marking requirements have been known and expected by the industry for a while as the AI Act came into force in August 2024.

He says they are likely to apply both to the model providers - Anthropic, OpenAI, Google and others - and to platforms powered by them. In other words, this will be something legal AI companies need to think about too.

“The fact that a legal AI platform is built on somebody else’s model doesn’t necessarily take it outside these requirements,” Joshi says.

“There are obligations that potentially run through the chain. From an AI Act perspective, how compliance obligations like watermarking are actually met will sometimes rely on close cooperation between model providers and companies integrating those models.”

Non-Billable asked Harvey and Legora how they are thinking about the transparency requirements and, in particular, what happens when lawyers export AI-generated content from their platforms into other documents.

Harvey declined to comment, while Legora did not respond by the time of publication.

No longer invisible

Speak to law firm partners and innovation people inside firms and they'll tell you that most clients today expect AI to be used by their law firms - chiefly because they want firms to think harder about how much they end up being charged.

There is increasingly an assumption that firms use AI across their work, with firms themselves left to decide what that looks like.

This makes sense. Clients largely pay for outcomes and can ultimately sue if they receive negligent advice. How the work is completed is generally of less importance, except to the extent it ends up costing them more.

Watermarking could well change that dynamic.

If AI-generated text starts carrying a detectable marker as it moves from a legal AI platform into Word and eventually into a document sent outside the firm, the specific uses of AI are no longer invisible.

That feels especially relevant after some of the industry's more embarrassing AI episodes.

Pinsent Masons here and Sullivan & Cromwell in the US have both been caught up this year in incidents involving fabricated case authorities making their way into court documents.

Those incidents were discovered because somebody spotted that the underlying authorities did not exist. In a world of widespread watermarking, there may be another clue that AI was involved somewhere in the production of a document.

It wouldn't prove that AI caused the error, of course. Anthropic itself is clear that its mark does not establish the full provenance of a piece of text. But it potentially adds another layer of information about how legal work was produced - which is something that barely existed until now.

Joshi expects transparency around AI use to become standard. He thinks that should broadly be welcomed as AI becomes an ordinary part of professional work.

“The EU has a track record of setting rules that end up becoming global standards - GDPR is the obvious example,” Joshi says. "I think we could see something similar with the AI Act, where greater transparency around AI-generated content becomes the norm well beyond Europe.

"That shouldn't necessarily be viewed negatively. If AI is going to become a normal part of how professional work is produced, some transparency around its use makes sense.”

The cost question

There is also an interesting link here to the way legal AI is starting to be priced.

The large legal AI companies are starting to move towards consumption-based pricing, passing on some of the token costs they incur from the likes of Anthropic and OpenAI to customers.

Legora recently rolled out consumption-based pricing for its premium agentic product, while Harvey has been dropping hints that this is also the direction of travel.

The result is that law firms could start incurring material AI-related costs for individual matters they work on, with some or all of those costs inevitably passed on to clients.

Greater transparency around AI use could actually help firms here.

If clients become accustomed to the idea that AI has been used in producing legal work - and that its use is fully disclosed and even marked - it actually becomes easier for firms to explain why a matter has incurred £500, £5,000 or £50,000 of AI-related costs alongside the lawyers' time.

In other words, the same rules that make firms' use of AI more visible could therefore make it easier for them to charge for it.