EU's mandatory AI watermarking rules kick in — hiding AI-generated content just got harder

Full text

BRUSSELS — It can sometimes seem like half the world is busy secretly using AI — and the other half is busy calling them out.

The AI detectors, human and otherwise, are about to get their biggest boost ever with new EU “watermarking” requirements that came into effect this month, which could upend how internet users experience and engage with AI-generated images, videos and text.

Ever since the rise of generative AI at the end of 2022, there have been fears that its dominance could mark the end of so-called “photographic truth”: what you see is by default true. While leading AI companies have said they wanted to address this risk by identifying the provenance of content generated by their models, their efforts have been largely voluntary.

Now, European Union legislators are demanding that internet users be informed whether they are engaging with AI and whether the content they see is the output of bots or the creation of a human.

On Monday, U.S. AI behemoth Anthropic announced that all its future models would support watermarking to comply with the EU rules.

“I don’t think that people understand the scale of change,” said Ashley Casovan, managing director of the AI governance center at the privacy- and AI-focused expert organization IAPP. She described a change from a world where people had to make an “educated decision” on whether something was AI-generated to one where such content has a verified symbol or overlay.

Even without having a top AI platform, Europe is shaping how this transformative technology is being rolled out globally. The EU’s flagship AI law, the Artificial Intelligence Act , has begun to force the world’s largest platforms to take crucial decisions about everything from safety to transparency.

It’s a reminder that Brussels often ends up as a kind of global regulator of last resort, ready to step in and set limits on the Big Tech industry that might otherwise barrel ahead at full speed. In the past, it was to safeguard privacy under GDPR and to police how platforms dealt with harmful content, which has had a knock-on effect around the world.

Now, since Aug. 2, providers of the most cutting-edge generative AI models, such as Anthropic and OpenAI, have had to ensure their users know they are interacting with AI and that the output is marked as AI-generated.

The law’s transparency obligations are “the main focus at the moment, in terms of what everyone is talking about,” said Francesca Pole, data, privacy and cybersecurity lawyer at DLA Piper.

Visualize the invisible

The most prominent companies have announced they would play ball: not only by signing an EU set of best practices for compliance, but also by announcing a series of technical tweaks.

OpenAI vowed at the end of July that “people should have better context about the content they see online, including whether it was created or edited by AI.”

The EU’s flagship AI law has begun to force the world’s largest platforms to take crucial decisions about everything from safety to transparency. | Kenzo Trbouillard/AFP via Getty Images

The company relies on SynthID, a watermarking tool developed by Google’s AI-focused branch DeepMind, to give AI-generated images an “invisible” overlay.

At the end of July, it extended that to audio.

Anthropic followed suit this Monday, announcing that all Claude models released since Aug. 2 would add “embedded watermarks in text.”

The watermark is woven into the text itself, the company said in a blog post.

“You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” the post read. “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere.”

Both companies also propagated an industry standard, developed by an organization called C2PA (the Coalition for Content Provenance and Authenticity), meant to store and transfer provenance metadata.

Watermarking content is only the first step though, detecting those marks and making them visible to internet users a crucial second one.

Companies like Anthropic will have to work with platforms where AI-generated content is being circulated to “visualize it,” said Walter Pasquarelli, a researcher in synthetic content at the University of Cambridge and advisor at the OECD.

Here’s how he described the process working in practice: “You generate something on Claude, you put it onto a social media platform, you’re about to make a post, and the interface where the post is put in may then have the detectors that can detect the watermark.”

The consequences could be far-reaching. “There’s plenty of evidence, … when it comes to media, that as soon people know that something is AI-generated their engagement drops quite significantly,” he said.

Or, in current digital parlance, Pasquarelli added, the EU’s move could reduce AI slop.

Platforms are already well aware of the rise of AI, with LinkedIn recently adding an option for users to flag a post that “seems like AI slop.”

That goes far beyond Brussels’ initial goal of preventing AI-generated content from being deceptive. “The conversation has shifted from AI-generated content which has a deceptive intent to AI-generated content as a whole,” Pasquarelli quipped.