Europe Starts Enforcing Its AI Act. The World’s Biggest Tech Companies Are Already Changing.

Europe's sweeping artificial intelligence law has entered a new and more consequential phase of enforcement, shifting from a regulatory blueprint to rules that companies are now expected to follow.

Under the new transparency requirements, companies must make it clearer when people are interacting with AI rather than another human. Providers must also use machine-readable markings to help identify content that has been generated or manipulated by AI. Deepfakes are subject to disclosure requirements as well.

The European Commission's AI Office can now begin enforcing provisions of the law alongside national authorities. The new powers represent a significant shift from preparation to oversight, although regulators have not yet announced enforcement action against companies under the latest requirements.

"This is a messy piece of legislation," Patrick Van Eecke, a Brussels-based partner at Cooley who advises companies on AI Act compliance, told Axios. For ordinary consumers, however, the most noticeable changes are more notices identifying chatbots as AI and more labels or technological markers indicating that images, videos, audio or text were generated or modified by artificial intelligence.

Those requirements are already influencing some of the world's largest technology companies. Anthropic announced this month that future Claude models will generate text containing a watermark designed to indicate the likelihood that Claude was involved in producing it. The company said the watermark will not be visible to readers and will not contain identifying information about individual users. Anthropic plans to make detection tools available to users.

The implications extend far beyond Europe. Anthropic said it plans to apply its watermarking globally at launch because it does not currently have a durable way to limit the technology geographically.

Meta has also signed the EU's Code of Practice on Transparency of AI-Generated Content and said it will work with the AI Office and industry partners to make transparency systems practical and interoperable.

OpenAI, meanwhile, has published summaries describing content used to train its general-purpose AI models and says its approach to provenance includes technologies such as Content Credentials, C2PA metadata and SynthID watermarking.

That growing adoption could ultimately prove more important than whether other governments reproduce Europe's legislation word for word. "We need a standard. This technological moment is happening really quickly. And right now, at least the EU AI Act gives us a standard," Amy Worley, managing director and data protection officer at Berkeley Research Group, told Axios.