To label or not? AI-generated content and the EU AI Act

Martin is CPTO at Kontent.ai and a past Boye conference speaker

AI is already part of many content workflows. The harder question is whether, where and how its involvement should be disclosed.

New transparency requirements under the EU AI Act have prompted strong reactions, new tools and plenty of uncertainty. The rules are narrower than some headlines suggest, while important practical questions remain unresolved: Should the label apply to a paragraph, a page or an entire website? What about translated or AI-assisted content? And when is human editorial oversight enough?

Article 50’s disclosure requirements have applied since 2 August 2026. For content teams, this makes disclosure a current publishing question: which content falls within scope, how is it reviewed and who takes responsibility for publication?

In a recent members’ call, Martin Michalik, CPTO at Kontent.ai, helped bring some perspective to the debate. He distinguished what the rules already establish from the practical questions still taking shape, and outlined steps content teams can take without waiting for every detail to be settled. A useful starting point was separating what technology needs to record from what readers need to see.

Start with what you are publishing

For a content team, the first question is what the content is intended to do. An AI-generated summary of a town council decision and an AI-assisted product description may share the same technology, but their purpose and implications differ.

Article 50 includes disclosure requirements for AI-generated or manipulated text published to inform the public on matters of public interest. Martin’s examples made this more concrete: information about public decisions, health or investor matters can fall within that scope. Ordinary product descriptions generally sit outside it, although claims about health, safety or sustainability can change the assessment.

The rules also address deepfakes: AI-generated or manipulated images, audio or video that resemble existing people, objects, places or events and would falsely appear authentic or truthful. These need a separate assessment from text.

This gives teams a more manageable starting point than treating every use of AI as the same disclosure problem. Look at the content you publish, its purpose and the consequences of someone relying on it. The assessment may differ across a single website.

Translated and AI-assisted content need the same careful assessment of purpose and editorial process. Those descriptions alone do not answer whether disclosure is required.

Human review needs editorial responsibility

For public-interest text, there is an important exception where content has undergone human review or editorial control and a person or organisation holds editorial responsibility for its publication. That exception does not extend to deepfakes.

The practical question is how that responsibility works inside the organisation. Who reviews an AI-generated health article? Who approves changes to an investor report? What happens when a colleague uses AI to revise content that has already been approved?

Martin recommended documenting the editorial process and giving each step a clear owner. This makes review easier to understand and sustain as tools and workflows change. An approval step has more value when the reviewer knows what they are checking and has the authority to challenge the output.

These are familiar content governance questions, given fresh urgency by AI. Organisations may already have suitable review practices. Others may discover that automation has moved faster than their arrangements for oversight.

What systems detect and what readers see

Martin distinguished two mechanisms that are easily confused. Marking makes AI-generated or manipulated content detectable by machines, using techniques such as watermarks or metadata. Labelling communicates its artificial origin to people encountering the content.

AI system providers have obligations concerning machine-readable marking and detection, subject to exceptions. Organisations using those systems have disclosure obligations for content within the relevant scope.

Martin also highlighted 2 February 2027 as the milestone under the voluntary transparency Code of Practice for participating providers to make detection work across systems. The aim is to help others identify which detection method to use for a piece of content. This does not postpone publishers’ existing disclosure obligations.

For a publishing team, a provider’s watermark, an internal CMS field and a visible label serve different purposes. The watermark supports detection, the field can retain information about the workflow, and the label informs the reader. Understanding those differences helps teams ask more precise questions about what their technology provides.

Record the detail without overwhelming the reader

Our member Laura Delnevo from Magnolia CMS raised a question that becomes increasingly relevant as AI is integrated into content platforms: should metadata about AI involvement support internal auditing, inform the end user, or both? If automated workflows become commonplace, how useful will an “edited by AI” flag be?

Martin suggested a simple internal field to record AI involvement, set automatically when AI modifies content. Recording this at component level, such as an image or a body of text, preserves information that can support later decisions.

A page might combine human-written text, an AI-generated image and a revised summary. Retaining information about those components gives the publishing team more flexibility than a single website-wide flag.

The discussion left open how that detail should translate into visible disclosure. A label needs to communicate something useful without overwhelming the reader. Internal records can be more detailed because they serve a different purpose: helping the organisation understand what happened and decide what to disclose.

Ask more of your technology vendors

Useful records depend partly on what platforms make available. Martin encouraged members to ask vendors for detailed audit logs, information about AI changes and support for passing metadata through the publishing process.

Can the platform show what was modified? Does that information survive export or reuse? Can reviewers see it before approving publication? These questions become more important when content travels between systems.

Simon Jones from Studio 24 in Cambridge shared Claude’s explanation of how it marks AI-generated content, providing a concrete example of a provider’s approach. Detection also has limitations: a positive or negative result should not be treated as a complete account of how content was produced.

Providers have their own obligations, while publishing organisations still need to own their editorial decisions. Technology can support that responsibility by making the process more visible.

Teams do not need to settle every question before improving their publishing processes. Knowing where AI is used, how its output is reviewed and who approves publication provides a useful foundation while guidance and expectations continue to develop. Whether to label becomes easier to assess when the organisation understands the content behind the decision.

The conversation continues

If reading and being a part of an online call is not enough, you’re very welcome to join us more actively. Our community is built around learning together, comparing notes, and exploring how theory meets practice across roles, industries and regions.

You can also:

However you choose to engage, we’re glad you’re here and part of the journey.