EU will require labels on realistic AI-generated content from August 2

The European Union is introducing rules that will oblige companies to label artificially generated content whenever it is created to look authentic. Starting August 2, digital watermarks will be required on synthetic images, audio, and text that are meant to seem real. The measure stems from the bloc’s previously passed AI Act.

The rules include notable exemptions. Personal content, such as a joke in a private group chat, will not need a label. Works that are clearly artistic, including satire and fiction, are also excluded. The August 2 deadline applies to new AI systems introduced on the EU market; systems that were already in place will have four additional months to adapt.

Sergey Lagodinsky, a European Parliament member who helped negotiate the AI Act, said the issue involves both consumer protection and the protection of democracy. He argued that such transparency is necessary to preserve democracy and the authenticity of facts online. Companies that fail to meet the requirements could face fines of up to three percent of their total gross revenue.

The EU has created its own black-and-white labels for anyone to use, while organizations may choose to design their own labeling systems. Boniface de Champris, an AI policy lead at the Computer and Communications Industry Association, told The Guardian that consumers might be surprised to see labels in certain places. He suggested that much AI-generated content on social media is already labeled, but people could be shocked to find it in areas where AI is used at scale without the public knowing, such as advertising, film, and publishing.

While large social media companies tend to have policies for labeling realistic AI imagery, enforcement is not foolproof. TikTok says creators need to label AI-generated posts, but various types of content still slip through. One current concern is short videos featuring AI-created doctors offering questionable health advice. Google says its SynthID tool has labeled more than 100 billion images and the equivalent of 60,000 years of audio. Meta has also adopted policies designed to help people identify AI-generated photo-realistic images, including several label options.