The EU Now Requires Labels on AI-Generated Images, Audio and Text

Key points

• EU rules requiring labels on AI-generated images, audio and text came into force on Sunday.
• The rules cover any artificially created content designed to look or sound real, including fake videos, voice recordings and written articles.
• Political deepfakes, meaning fake videos or audio clips generated by AI that look and sound like real politicians, were a key driver behind the legislation.
• Companies that fail to label AI-generated content face potential penalties under the broader EU AI Act framework.
• The rules apply to content distributed inside the European Union, regardless of where the company producing it is based.

For the past several years, European voters have watched fake videos spread online showing politicians apparently stealing elections, bullying staff or performing embarrassing stunts. None of it was real. All of it was convincing. Starting this Sunday, that kind of content, if made using AI and designed to look authentic, must carry a clear label saying so.

The Guardian AI first reported the rule's activation date. The requirement sits inside the EU's wider push to regulate artificial intelligence before its harms outpace its benefits.

What exactly has to be labelled?

Any AI-generated content meant to look or sound genuine. That means a realistic fake photograph of a politician, an audio clip mimicking a CEO's voice, or a written article that presents itself as human-authored journalism. The label must be visible and clear, not buried in small print.

Content that is obviously fictional or clearly a creative experiment is treated differently. The target is material designed to deceive.

Who does this affect?

Every company distributing content to EU audiences must comply, even if they are headquartered outside Europe. A social media platform based in California, a news aggregator run from Singapore: if their content reaches people in France or Germany, the rules apply.

For ordinary people scrolling social media or reading news online, the change means AI-made content should, in theory, start carrying a visible flag. Think of it like a nutritional label on food: it does not stop you consuming it, but it tells you what you are getting.

Content type	Must be labelled?	Key concern
AI-generated political images	Yes	Election manipulation
AI voice clips of real people	Yes	Fraud, defamation
AI-written articles posed as real	Yes	Misinformation
Clearly labelled AI art or fiction	No	Not designed to deceive
AI tools used to edit, not generate	Case by case	Depends on intent

Will this actually work?

Enforcement is the open question. Labelling rules only function if platforms and creators apply them honestly, and if regulators have the resources to check. The EU's AI Act, the broader law underpinning these rules, does give regulators teeth: fines can reach into the tens of millions of euros for serious breaches.

The bigger challenge is speed. New AI tools that generate convincing fake content appear faster than legislation can track them. Researchers who study disinformation say labelling alone will not stop determined bad actors, but it does shift legal responsibility onto the platforms hosting the content.

For now, if you see a striking video of a politician doing something outrageous, it is worth pausing before sharing it. A label should eventually appear. Until enforcement catches up, a moment of scepticism is still your best protection.

Common questions

Does this rule apply to content I see on social media?

Yes. If a social media platform serves users in the EU, it must label AI-generated content that is designed to look real, regardless of who created the post.

What happens if a company ignores the rule?

Companies that breach EU AI Act obligations face fines that can reach tens of millions of euros, or a percentage of global annual revenue, whichever is higher.

Does this cover AI tools I use privately, like image generators?

The rules primarily target content distributed publicly. Using an AI image generator for personal, non-public purposes is a different matter, though publishing that output without a label could trigger the requirement.