Anthropic's newly launched watermarking feature for text created by its Claude generative artificial intelligence platform is aimed at promoting transparency and complying with legal requirements.
Among other things, it is also aimed at reining in difficulties caused by the technology and the AI slop that has flooded the internet. But the move seems to have opened up more questions than answers.
Industry observers argue that it may have conflicting effects on how generative AI would be used and how its output would be treated, especially as the company has yet to offer technical-level details on how the detection would work.
“Some people see this as a first step towards future detection tools – which, going by track record, don't work reliably,” said Tina Austin, director of AI education and learning systems at UCLA.
“Because a mark only proves Claude touched the content [not that it wrote all of it]. And no mark doesn't prove human authorship either, since heavy editing or format conversion can strip it out.”
Compliance risk
The EU AI Act’s Transparency Code explicitly states that users must be informed when they interact directly with an AI system and that machine-readable marks must be added to enable the detection of AI-generated or manipulated content.
Tech companies have routinely scrambled to comply with regulatory requirements, lest they be unable to sell their products and services in specific markets.
ChatGPT creator OpenAI, Gemini maker Google, Facebook parent Meta Platforms and Microsoft, among other tech giants in the AI game, have also complied. Even LinkedIn and Snap have introduced a button to flag AI slop.
Anthropic says the machine-readable watermark will travel with the generated text or file when it is copied and pasted, and will be applied across all versions of Claude's stable services.
But by its own admission, the San Francisco-based company notes that the feature has its limitations, including a disclaimer that even if a watermark is detected, it may “not [be] fully conclusive” that Claude created the content.
Saif Kidwai, founder of Dubai-based Alnaas Advisory, called Anthropic's move an “irony” that adds more confusion about the direction of AI regulation.
“These models were built by copying everyone's work without asking. Our writing, our books, our posts. And now they get to put a watermark on writing that was ours to begin with and quietly decide what still counts as human,” he said.
“The thief is now issuing certificates of authenticity. Their reason? They signed the EU AI Act, then rolled it out to the whole world. I think this is the big problem, when you try to roll guardrails out retrospectively. What a mess.”
Academic issues
For academia, the issue is more profound, as AI has become a polarising technology for use in learning institutions.
There are students who have relied on AI – heavily, at times – to generate reports, shunning the need for actual research and thinking, while others are against it. The AI issue has even spilt over to commencement ceremonies, where a number of industry figures – including former Google chief executive Eric Schmidt – have been booed for touting AI.
Tim Mousel, an AI faculty fellow at Lone Star College in Texas, acknowledges that watermarking could be another form of evidence when reviewing students' work.
At the same time, he questions whether this is “the beginning of the end of students using AI to write their papers”.
“There are important limits. A watermark does not prove misconduct and the absence of one does not prove AI was not used. Students may also use AI appropriately for brainstorming, proofreading, tutoring, or other permitted purposes,” he said.
“The larger implication is clear. Faculty need assignment-level policies that specify what AI use is permitted, restricted, or required. That level of clarity is beneficial for our students. If AI provenance becomes widely adopted across major platforms, it could significantly affect how colleges approach academic integrity.”
Richard Whittle, professor of AI and public policy at England's University of Salford, said despite the “jubilation” around Anthropic's move, it “worries me more than ever”.
“Think for yourself and use AI to write. We can detect it and you will be punished. Use AI to think for you and regurgitate those thoughts down on a page yourself. We can't detect it and you will pass.
“I still think this is the wrong way around. Our obsession with detection is simply facilitating AI replacing our students thought processes. To be honest, I'd rather their writing was replaced and their thoughts their own.”
Simply complicated?
Anthropic has made it clear how its watermarking will work on the surface, but it has yet to share enough technical detail to say how the algorithm for the feature works.
That said, the fact remains that the watermark will be another test of how users will react to the latest attempt at tempering misleading AI-generated content.
Rachael Trickey, editorial director at industry publication Decoded Thinking, said: “On the one hand, it makes sense. As AI-generated content becomes harder to spot, having a way to identify it could help with transparency and trust.
“But I'm not sure it's quite that simple. If someone asks Claude to write an entire article, that's one thing. But what if they write it themselves and use AI to improve the structure, rewrite a paragraph or tidy up the language?”
Also, watermarking can be removed but it is notoriously difficult: a research paper from the University of Maryland points out that a watermark cannot be removed without modifying a significant fraction of the generated tokens. There is already software that can remove AI watermarks to varying degrees, so a fully fledged application to make the process easier is not far off.
“Think about where this could eventually go … the infrastructure [for detection] is being created,” said Nick Chasinov, founder of New Jersey-based advertising firm Teknicks.
“And I think that changes the AI content conversation. My guess is that simply using AI won't be the problem. Using AI to mass-produce thousands of pieces of low-value content probably will be.
“AI isn't going away. But neither is the value of original ideas, experience, expertise and a recognisable human point of view. In fact, those things may become more valuable than ever.”
