EU makes AI content labels and watermarks compulsory August 2

On August 2 the EU AI Act's transparency rules stop being aspirational and start biting, and the enforcement architecture is unusual because it splits liability along two axes at once. According to the Guardian, providers of generative models must embed machine-readable watermarks into synthetic images, audio, video, and text, while deployers who publish that content have to disclose it in a way ordinary users can actually see. Chatbots have to identify themselves at the moment of contact, not tucked into the small print of a terms-of-service page.

The scope is broad in a way the industry has already flagged as a problem. Any AI-generated content designed to appear authentic is in, with carve-outs for personal use and for evidently artistic, satirical, and fictional works, plus an escape hatch when a human with genuine editorial responsibility has meaningfully reviewed AI-written text. Fines reach €15 million or 3% of global annual turnover, whichever is larger. Systems already on the market get until December 2, 2026 to embed the machine-readable marks, and a separate simplification package could shift the machine-marking deadline further, according to TNW's read of the current guidance.

The industry pushback centers on the collapse of a distinction the law was originally supposed to keep. CCIA Europe's AI policy lead argued the label was meant to flag deceptive content, and that with the deceptive-intent test gone, almost everything now gets labelled. That is a real risk for user attention: if a benign AI-assisted stock photo carries the same badge as a political deepfake, the badge stops meaning anything.

The honest caveat is that the technology under the policy is not settled. Watermarks can be stripped, metadata can be lost, and machine-written text is notoriously difficult to detect, so the rules will only be as strong as the checkers behind them. The reporting doesn't tell us how regulators plan to handle open-source models where the provider cannot easily bind downstream users, or how strict meaningful editorial review will be in practice.

The forward-looking part worth watching is who benefits. Platforms that already invested in the infrastructure, like Google, Meta, and TikTok, now have a regulatory moat over any generative startup that has to bolt watermarking on from scratch to sell into Europe. Google's SynthID reportedly covers over 100 billion images and 60,000 years of audio, and more than 180 organizations have already signed the EU's voluntary code of practice, which suggests the incumbents were positioning for this well before the deadline.