EU AI Act Enforces Mandatory Content Marking and Deepfake Rules Under Article 50

The change is about far more than adding another warning label to AI-generated content. Brussels is trying to establish common rules for identifying synthetic media before it becomes impossible to distinguish authentic material from AI-generated content online.

AI transparency is no longer optional

Until now, most AI companies decided for themselves how much information they wanted to provide about AI-generated content.

Some introduced watermarking technologies. Others experimented with provenance tools or small labels indicating that an image had been created by AI. There was no single approach, and participation largely depended on corporate policy rather than legal obligation.

That changes with Article 50.

The European Commission's Code of Practice on Transparency, published in June, remains voluntary. More than 180 organisations, including Google, have signed it because it offers a clear path to meeting the new legal requirements across the EU.

But the law itself is not optional. Companies that ignore the Code must still prove that their own systems meet the same legal standards. If they fail to do so, national regulators can impose fines of up to €15 million or 3% of global annual turnover.

For AI developers, transparency has become a compliance issue rather than a public relations exercise.

Different rules for developers and publishers

One important feature of Article 50 is that it does not place all responsibility on one side of the industry.

Developers of AI systems must ensure that synthetic content carries machine-readable information showing where it came from. The goal is to make AI-generated images, videos, audio and text easier for platforms and regulators to detect automatically.

Publishers have a different job. They must make sure users know when they are seeing AI-generated material on subjects that affect the public interest. The same applies to deepfakes and interactive AI systems, which must clearly inform people that they are communicating with software rather than another person.

The legislation also introduces notification requirements for emotion-recognition and biometric categorisation systems, giving people the right to know when such technologies are being used.

Instead of treating AI as a single product, the EU has divided responsibility across the entire chain - from the company building the technology to the organisation putting its output in front of the public.

Building a digital paper trail

The most important part of the new rules may be something most users will never see.

Instead of relying only on visible labels, the EU wants AI-generated content to include machine-readable metadata that software can detect automatically. In other words, Brussels is trying to create a digital record that follows AI-generated material wherever it goes.

The idea is straightforward. A label on an image can be cropped out or deleted. Metadata embedded in the file is much harder to remove without leaving traces. If this approach becomes widely adopted, online platforms, fact-checkers and regulators will be able to identify synthetic content much faster and on a much larger scale.

That is why technologies such as Google's SynthID and the C2PA provenance standard have become increasingly important. They are designed to make digital content traceable from the moment it is created rather than after it has already spread across the internet.

The ambition goes beyond helping users spot fake images. The EU is trying to build technical infrastructure that allows digital information itself to be verified.

Europe's rules are becoming global rules

Although the AI Act is a European law, its impact will not stop at the EU's borders.

Any company whose AI-generated content reaches users in the European Union must comply with Article 50, regardless of where that company is based. A developer in the United States or Asia faces the same obligations if its products are available in the European market.

This is another example of the so-called "Brussels Effect." Because the EU represents one of the world's largest digital markets, companies often decide it is easier to apply European standards everywhere instead of maintaining separate systems for different regions.

That makes the AI Act influential far beyond Europe. Even businesses with no headquarters or major operations inside the EU may have to redesign their products to keep access to European customers.

For many technology companies, complying with the AI Act is likely to become part of their global business strategy rather than simply a European legal requirement.

Compliance will not be equally easy

Not every company will face the same challenge.

Large technology firms have spent years developing watermarking systems, provenance tools and compliance teams. Many already had the technical resources needed to adapt their products before the rules took effect.

Smaller developers have a much harder task.

Open-source AI projects, in particular, face a practical problem. Once their models are released, they have very little control over how other people modify or distribute them. Keeping machine-readable markings attached to AI-generated content throughout that process is far more difficult than it is for companies operating closed commercial platforms.

As a result, the cost of compliance could fall much more heavily on startups and open-source developers than on established technology giants.

That may become one of the less visible consequences of the AI Act. Rules designed to improve transparency could also make it harder for smaller players to compete with companies that already have the money, legal expertise and technical infrastructure needed to meet the new requirements.

Transparency is only part of the challenge

Few people would argue against making AI-generated content easier to identify. The bigger debate is whether the growing number of labels, notices and technical standards will actually make life simpler for users.

Even companies that support the new rules have warned about this. Google, for example, signed the Commission's Code of Practice but has also argued that too many overlapping requirements could confuse users instead of helping them. Different watermarking systems, multiple warning labels and varying technical standards could create a patchwork that is difficult for both companies and the public to navigate.

Brussels has accepted that risk. Its view is that a complicated transparency system is still better than allowing AI-generated content to circulate without any reliable way to identify its origin.

That thinking explains why the EU has focused so heavily on provenance rather than just visible labels. The aim is to create a system that works not only for individual users but also for online platforms, regulators and fact-checkers trying to detect synthetic content at scale.

A new stage in AI regulation

Article 50 is another step in the gradual rollout of the AI Act rather than a standalone measure. Earlier phases introduced restrictions on high-risk AI practices and new obligations for developers of general-purpose AI models. Now the focus is shifting to the content those systems produce.

That may prove to be one of the most significant changes in the entire legislation.

For years, the discussion around AI centred on what these systems could create. Europe is now asking a different question: can every AI-generated image, video, audio clip or text be traced back to its source?