Summary

• The European Commission has published a final code to support AI Act transparency duties on marking and labelling generated or manipulated content.
• Disclosure will be an operational problem across publishing, customer communications, procurement, HR, public bodies, and software supply chains.
• Labels will be hardest where machine-generated drafts, human editing, vendor tools, metadata, and distribution platforms all overlap.

“Made in TSMC” is not a label anyone expects to see on a chatbot transcript, a council notice, or a corporate video, but the phrase points at the problem Europe is now trying to solve.

The physical economy has long understood provenance. Components have suppliers, food has origin labels, medicines have batch numbers, and industrial goods carry markings that tell buyers, regulators, and customers something about how they were made. Digital content has grown up with fewer visible clues. A report, advert, support message, video clip, social post, or public-service notice can now pass through machine-generation systems, human editing, translation tools, design software, and distribution platforms before anyone sees it.

Brussels wants clearer markings on that production line. The European Commission has published the final Code of Practice on marking and labelling AI-generated content, intended to help providers and deployers prepare for transparency obligations under the AI Act. Those duties begin applying on 2 August 2026 and cover areas including chatbot disclosure, deepfakes, and certain AI-generated or manipulated text published to inform the public on matters of public interest.

That can sound like a narrow regulatory exercise, but the practical burden will be much wider. The code is not just about whether a fake video should carry a label. It points towards a new operating discipline for organisations that create, approve, buy, publish, distribute, or archive digital content.

The label is only the visible part

The Commission’s code separates the problem into two broad areas. Providers of generative systems are expected to support marking and detection, while deployers are expected to label certain outputs, including deepfakes and specified public-interest material. The EU has also developed icons that organisations may use to signal AI-generated content.

The visible label is the simple part. The harder work sits behind it. Organisations need to decide when a disclosure is required, who makes that judgement, how it appears, whether it travels with the content, how records are kept, and what happens when the content is revised, republished, translated, embedded, or broken into smaller pieces for different channels.

Anyone treating this as a website footer problem will struggle. The same organisation might use one tool to draft a customer email, another to produce a product explainer, a third to generate training material, a fourth to edit video, and a fifth to handle chatbot interactions. An agency, supplier, employee, platform, and automated workflow may each touch the output before publication. By that point, asking whether the content is “AI-generated” can become less obvious than the phrase suggests.

Mixed authorship will cause the arguments

Most digital content will not arrive in a clean binary state. It will be drafted by a model and heavily rewritten by a person, or written by a person and then shortened, translated, illustrated, voiced, or formatted by software. A customer-service script may draw on approved human-written policy but be generated dynamically inside a support tool. A public consultation summary may begin as a machine-made draft before policy officials edit it line by line.

Those mixed workflows create practical questions that legal teams alone cannot answer. At what point does machine assistance become machine-made content? Does a human rewrite remove the need for a label, or simply change the form of disclosure? How should a disclosure appear in audio, video, chat, PDF, advertising, or internal knowledge systems? What evidence should be retained if a regulator, customer, journalist, union, or court later asks how the content was produced?

The public-interest text category is especially awkward. Public bodies, regulated businesses, trade associations, universities, consultancies, charities, and large employers all publish material that can shape decisions and expectations. Some of it will be clearly informational, some promotional, and some politically sensitive. Once generative tools are embedded in ordinary communications work, disclosure becomes part of governance rather than a final check by a web editor.

Software buyers will also have to become more precise. A vendor selling a content platform, chatbot, marketing suite, HR tool, learning system, customer-service assistant, or design product will increasingly need to explain how outputs are marked, logged, detected, exported, and labelled. A generic claim that a product is “AI Act ready” will not be enough when workflows differ by sector, audience, and risk.

That changes procurement questions. Buyers will need to know whether marking uses open standards, whether metadata survives export, whether labels can be configured by market, whether audit logs are accessible, whether human approval steps are built in, and whether the supplier’s own model outputs can be distinguished from user-edited material. The answers will shape contract terms, risk registers, staff training, and product design.

Publishing teams will face a different version of the same issue. The mechanics of disclosure have to fit real production cycles, where material is repurposed across websites, newsletters, social media, video, sales documents, press materials, and customer portals. If labels are added manually at the end of that chain, they will be missed. If they are added too broadly, organisations risk training audiences to ignore them.

Trust needs more than a sticker

Labels will not fix synthetic media on their own. A disclosure can be hidden, stripped, misused, ignored, or attached to material in ways that create more confusion than clarity. Bad actors are unlikely to become compliant because an icon exists. Even well-run organisations will struggle where content passes through multiple tools and jurisdictions.

Still, the absence of disclosure becomes harder to defend as machine-made content becomes ordinary. Customers, citizens, staff, investors, and regulators will not always object to automated production, but they will expect honesty when the form of production changes the nature of the message. A chatbot pretending to be human, a synthetic image presented as documentary evidence, or a machine-written public notice carrying institutional authority each creates a different kind of trust problem.

The EU’s approach turns that trust problem into paperwork, product design, and workflow control. It asks organisations to put receipts on parts of the content machine. The ones that adapt early will not treat labelling as a legal sticker applied moments before publication, but as a production discipline built into how digital work is commissioned, created, checked, and released.