Significant changes to the way internet users view and interact with images, videos and texts created by artificial intelligence could be brought about by the European Union’s new “watermarking” requirements, which came into force this month, with the aim of strengthening tools for detecting AI-generated content, whether produced automatically or not.

Since the “explosion” of generative artificial intelligence in late 2022, there have been fears that its dominance could spell the end of “photographic truth” — that is, the assumption that what we see in a photograph is inherently real.

The largest AI companies have said they want to address this risk by making it possible to verify the origin of content produced by their models. Until now, however, the relevant initiatives have been largely voluntary.

Now, European Union lawmakers are requiring internet users to be able to know whether they are interacting with an AI system and whether the content they are viewing was produced by a machine or by a human.

On Monday, U.S. company Anthropic announced that all its future models will support watermarking technologies in order to comply with European rules.

“I don’t think people realize the scale of the change,” said Ashley Casovan, CEO of the IAPP’s AI Governance Center, an organization specializing in privacy and artificial intelligence issues. As she explained, we are moving from a world in which users had to make an “informed assessment” of whether something was created by AI to a world in which such content can be accompanied by a verified symbol or relevant indication.

Brussels is setting the rules beyond Europe as well

Although Europe does not have any of the world’s leading artificial intelligence platforms (the U.S. and China have “been playing in a league of their own” so far), it has a significant influence on how the technology is developed and deployed around the world.

The European AI Act has already begun forcing the world’s largest platforms to make critical decisions on issues ranging from safety to transparency.

Brussels is thus once again assuming the role of a peculiar global regulator vis-à-vis Big Tech. The same happened in the past with the protection of personal data through the GDPR and with rules governing how digital platforms deal with harmful content — interventions that had an impact far beyond Europe’s borders.

As of August 2, providers of the most advanced generative AI models, such as Anthropic and OpenAI, are required to ensure that users know they are interacting with artificial intelligence and that generated content is labeled as an AI product.

The law’s transparency obligations are currently “the main subject of discussion,” according to Francesca Pohl, a lawyer at DLA Piper specializing in data, privacy and cybersecurity.

The “invisible” marks embedded in images and text

The industry’s largest companies have already stated that they will comply, both by signing a European code of best practices and by making technical changes to their products.

OpenAI said at the end of July that “people should have a better understanding of the content they see online, including whether it was created or modified by AI.” The company uses SynthID, a watermarking tool developed by Google’s DeepMind, to add an “invisible” watermark to AI-generated images. At the end of July, the technology was also expanded to audio.

Anthropic followed suit, announcing on Monday that all Claude models released from August 2 onward would add “embedded watermarks to text.”

According to the company, the watermark is embedded in the text itself. “You won’t see it, and it doesn’t change the meaning, quality or readability of Claude’s response,” Anthropic said. “Because the watermark is part of the text, it will remain with it when it is copied and pasted elsewhere.”

At the same time, the two companies have adopted the industry standard C2PA (Coalition for Content Provenance and Authenticity), which is designed to store and transmit metadata concerning the origin of content.

What we will see on social media

Adding the watermark is only the first step. The second — and equally critical — is detecting it and converting it into visible information for the user.

Companies such as Anthropic will need to work with the platforms on which AI-generated content circulates so that the relevant information becomes visible, explained Walter Pasquarelli, a researcher in synthetic content at the University of Cambridge and an OECD adviser.

He described how the system could work in practice: a user creates something in Claude, transfers it to a social media platform, and, when attempting to publish it, the platform detects the watermark and can indicate that the content was created by AI.

The consequences could be significant. According to Pasquarelli, there is already substantial evidence showing that when users know a post was created by artificial intelligence, their engagement with it decreases significantly.

In terms of today’s internet, he noted, the European rules could curb so-called “AI slop,” meaning mass-produced, low-quality artificial-intelligence content.

Platforms have already begun responding to its spread. LinkedIn, for example, recently added the ability for users to flag a post as content that “looks like AI slop.”

The discussion, therefore, appears to have moved beyond Brussels’ original objective, which was primarily to prevent deception through AI-generated content and, according to Pasquarelli, has shifted from misleading AI-generated content to the entirety of content created using artificial intelligence.