The EU launches a code to tackle AI-manipulated photos and videos
Fake doctors offering bad advice. Finance ministers suggesting investments. The faces and voices of celebrities asking for help on someone’s behalf. Fake videos are becoming increasingly common, from India to America, from Russia to Europe, partly because they are easy to create using generative artificial intelligence. The flood of photos showing public figures in indecent attire – caused by the lack of discretion in Elon Musk’s model monitoring system, Grok – flooded social media in early 2026.
It is time for Europe to take countermeasures. From 2 August, the provisions of Article 50 of the AI Act will come into force, requiring companies that produce artificial intelligence systems and those who use them to create and publish text, photos, videos and audio to take responsibility for their actions. And now, thanks to the work of a team of independent experts, the Commission is also setting out how to mitigate the associated risks.
In practice, the general principle is that developers of generative artificial intelligence and those who use it to produce content for publication must be transparent and ensure that users are aware they are interacting with content generated by artificial intelligence. System developers are obliged to mark the generated content with a sort of digital watermark or other solutions that can be read by specialised software and enable the artificial origin of, for example, images or videos to be recognised. For those who had systems available on the market prior to 2 August, the deadline is likely to be extended to 2 December. Those who produce and publish content created using artificial intelligence, in turn, must, from 2 August, inform the public – subject to exceptions provided for by law – by using special labels.
But it is not always easy, from a technical point of view, to comply with these obligations. Content creators must inform the public if they use artificial intelligence, but without compromising the enjoyment of the work. And system developers must label the outputs of their artificial intelligence, but in a way that is simple to use and recognise, ideally using standardised and interoperable tools. To make this task easier for those involved, the Commission has appointed a group of experts who have drawn up a code of practice – in other words, a sort of manual that allows those who follow it to demonstrate that they have done everything possible to avoid misleading interpretations of content generated by artificial intelligence.
The code is being previewed today. Dino Pedreschi, a lecturer at the University of Pisa, was one of the leaders of the expert group. “The code of practice is part of a comprehensive effort to move towards a less polluted information ecosystem. If things go to plan, a large proportion of synthetic material will be clearly identified as such. And, in the long run, it will also be easier to prove that content has been created by humans.” The aim is to develop the use of metadata that describes the information circulating online. “Our code of practice is adopted and applied voluntarily,” explains Pedreschi, “but those who adopt it can be certain they know what they need to do to comply with this part of the AI Act.” Consequently, amongst other things, this will make it easier to defend oneself in court should legal action arise over matters relating to this kind of content. Cases will be heard in the courts to which any affected parties have turned. In some cases, the Data Protection Authority also has jurisdiction. The code of practice will be of assistance to developers and publishers who adopt it voluntarily and find themselves having to defend their position. OpenAI has already signed up. Meta, for the time being, has refused.
