The European Commission’s AI Office switched on Article 50 transparency enforcement and full fining powers over general-purpose AI providers on August 2, converting two years of drafting theatre into three operational duties: chatbots must tell users they’re chatbots, deepfakes must be labelled, and AI-generated content must carry a machine-readable mark. GPAI providers who fall short now face investigation, corrective orders, and penalties up to €15 million or 3% of worldwide annual turnover, whichever bites harder.

The staging is characteristically Brussels: obligations first, icons second, high-risk system rules phased into 2027–2028. DG CNECT reports that roughly 190 organisations had signed the Code of Practice on Transparency of AI-generated Content by end of July, a respectable pre-commit for a regime whose three disclosure icons (AI-involved, AI-generated, human-with-AI-modification) remain voluntary. Systems already on the market get until December 2 to add machine-readable marking, per Tech Policy Press, and deepfakes made before the switch date don’t need retroactive labelling.

The interesting fight isn’t the fine ceiling. It’s the carve-out for “obvious” AI interactions and “evidently artistic, creative, satirical, fictional” works. Desara Dushi, a researcher at Vrije Universiteit Brussel, flags this as the first interpretive battleground, noting the “obviousness” test is borrowed from EU consumer law’s “average consumer” standard. That doctrine has produced two decades of litigation over who exactly the average consumer is.

For SMBs, the surface area is wider than most realise. No-code platforms like Glean, Dust, and LemonLime, which many small teams use to run outbound and customer chat, put the deploying business directly in scope the moment an EU user is on the other end. The disclosure duty attaches to the deployer, not the toolmaker. Brussels has, quietly, made every marketing team a compliance surface.

Sources

Sources