EU's New AI Content Rules Take Effect in August—Here's What the February 2027 Deadline Means for Marketing Teams

Marketing teams got a real deadline this year, and the details matter more than usual. Article 50 of the EU AI Act, the transparency rule governing AI-generated content, took effect 2 August 2026. Two more dates matter: 2 December 2026 closes the grace period on machine-readable watermarking, and 2 August 2027 forces legacy AI vendors into full compliance. Ignore the "February 2027" date floating around some coverage; it doesn't match any official EU document.

What Article 50 actually requires: the four duties and who they bind

Four duties apply, regardless of risk classification, and each targets a different kind of content. Chatbots must say they're AI at the start of an interaction, while AI images, video, and audio need machine-readable provenance data, and deepfakes require disclosure outright. AI text on public-interest topics needs a label, unless a named human editor reviews it and takes responsibility.

Here's the part most marketing teams miss: the Act cares who hit publish, not who built the model. Download an AI product shot and post it, and you're a deployer, full liability included, and there's no exemption for brands outside the EU if the content reaches EU audiences.

Which marketing content types fall inside the rules, and which don't

Visuals face the broadest scope. Synthetic models, AI-rendered scenery, and product shots in fake environments all count, while authentic photos with color correction don't, and fines for non-compliant images can hit 3% of global turnover.

Text is narrower. Product pages and ad copy sit outside scope; editorial and thought-leadership content aimed at public interest doesn't, and human review is the main way through. Chatbots need disclosure immediately, before any terms page.

The color-correction line sounds tidy until you try to draw it on a real asset library. A background-removal tool and a generative-fill tool can touch the same image in the same session, and only one triggers disclosure.

The two-layer watermarking standard and what "machine-readable" means in practice

The Code of Practice, published 10 June 2026, calls for a multi-layered approach combining machine-readable provenance metadata and invisible watermarking techniques. Several major generative AI platforms have begun implementing such standards, but the gap sits in workflow. Metadata can be lost at various points in a typical publishing pipeline, in ways nobody on a marketing team tends to watch for. The tools exist, but the discipline to verify what survived the pipeline doesn't, yet.

The 2 December 2026 watermarking deadline and the 2 August 2027 GPAI legacy deadline

Disclosure duties are live now, with no grace period attached. Watermarking gets until 2 December 2026, while legacy GPAI vendors have until 2 August 2027.

The enforcement regime that makes these deadlines real

Fines reach €7.5 million or 1% of turnover for transparency failures, higher for GPAI breaches, and the AI Office favors "technical compliance dialogues" before formal action, though enforcement capacity varies sharply by member state.

How the UK divergence complicates compliance for cross-market brands

No equivalent UK rule exists yet, and that gap creates its own cost. Running two separate pipelines, one EU-compliant and one not, costs more than just applying the EU standard everywhere.

The compliance workflow marketing teams need to build before 2 December 2026

Map every AI asset type, then audit the image pipeline for stripped manifests, and fix chatbot disclosures today, not next quarter. Document editorial review as a real step, not an afterthought, since that's exactly what the safe harbor demands.