Until very recently this was a question about taste. Since 2 August 2026 it is also a question about paperwork, because that is the date Article 50 of the EU AI Act became applicable and disclosure of AI-generated content moved from good manners to a legal obligation in defined circumstances. Anyone writing brand copy for a European audience now has two separate problems to solve, and they have different answers.
Key takeaways
- EU AI Act Article 50 has applied since 2 August 2026, with machine-readable marking phased in by 2 December 2026.
- Meaningful human review with named editorial responsibility removes the text labelling duty.
- Google’s position is unchanged: AI use is not penalised, scaled low-value output is.
- Models regress towards the mean, so distinctiveness has to be supplied, not requested.
What changed on 2 August 2026
Article 50 of the EU AI Act sets transparency duties for both providers and deployers of certain AI systems, and the European Commission confirms it applies from 2 August 2026. Three obligations concern anyone producing marketing content.
First, people must be told when they are interacting with an AI system rather than a person, which covers chat and support interfaces. Second, deep fakes must be disclosed as artificially generated or manipulated. Third, and the one that has caused most confusion, deployers publishing AI-generated or manipulated text intended to inform the public on matters of public interest must label it. The Commission’s guidance frames public interest as topics such as politics, public administration, justice, fundamental rights, public security, health, environmental protection and consumer safety.
Most product marketing sits outside that third category. A campaign landing page for a handbag is not informing the public on a matter of public interest. A brand article about a health claim, an environmental credential or a consumer safety issue plausibly is, and that is where the line will be argued.
The exemption matters more than the obligation. Content that has undergone genuine human review or editorial control does not need labelling, but the standard is specific: a deliberate examination of the substance by a person with relevant knowledge and professional judgement, with one identifiable person carrying ultimate legal responsibility for publication. The Commission is explicit that superficial, purely formal or procedural checks, spell-checking being the example given, do not count. A separate grace period runs to 2 December 2026 for machine-readable marking of outputs from systems already on the market, and a Code of Practice on Transparency of AI-generated Content offers one route to demonstrating compliance.
Google’s position is a different question entirely
Search visibility and EU transparency law are frequently conflated, and they should not be. Google’s published guidance on generative AI content, last updated on 10 December 2025, states that using generative AI tools to generate many pages without adding value for users may violate its spam policy on scaled content abuse. The trigger is unoriginal output produced at volume, explicitly regardless of how it was created. Google also suggests, without requiring it, adding information about how content was produced where that makes sense for the audience.
The practical reading has not moved in two years. Nobody is penalised for using a model. Publishers are penalised for shipping volume that nobody needed, and a model simply makes that failure cheaper to commit at scale.
Why models flatten a voice
A brand voice is mostly a set of refusals. It is the vocabulary a company will not use, the joke it will not make, the claim it declines to state without a number behind it. A language model is trained to produce the most probable continuation, which means it reaches for the phrasing that appears most often across everything it has read. Those two objectives are in direct opposition.
What comes back is competent and unplaceable. Nothing is wrong with it, which is precisely the problem, because a voice that could belong to any of forty competitors is not doing the job a voice exists to do. Distinctiveness has to be supplied as input, through examples, constraints and an explicit list of forbidden formulations, rather than requested as an outcome.
A workable division of labour
| Task | Sensible to delegate? | Reason |
|---|---|---|
| Subject line and headline variants | Yes | Volume task with a human picking the winner |
| Reformatting for a second channel | Yes | Structure changes, the argument does not |
| Spotting repetition across a sequence | Yes | Pattern detection is a genuine strength |
| Positioning and category language | No | The point is to be unlike the training data |
| Claims, figures and comparisons | No | Fabrication risk plus advertising liability |
| Crisis and sensitive communications | No | Judgement, and named accountability, are the deliverable |
Two things follow from that table. The review step is not optional overhead, it is the mechanism that keeps the voice yours and, under Article 50, the thing that removes the labelling duty. And a documented editorial owner is now worth having written down somewhere, because the exemption depends on a person, not a process. Where this fits into a wider publishing operation is something we set out in our guide to building a content marketing strategy that holds up.
Questions this keeps raising
Does every AI-assisted marketing email now need a label in the EU?
On the current guidance, no. The text labelling duty in Article 50 attaches to content published to inform the public on matters of public interest, and meaningful human review removes it in any case. Chatbot disclosure and deep fake labelling are separate duties with their own scope. Obligations differ outside the EU and continue to evolve, so verify what applies to your markets before assuming.
Can a model be trained to hold a specific brand voice?
It can be steered a long way with strong examples, explicit constraints and a banned-phrase list. It still drifts back towards generic phrasing over long outputs, which is why the last edit tends to matter more than the first prompt.
Should a brand tell readers it uses AI?
Where no legal duty applies it is a positioning choice rather than a requirement. Google suggests explaining how content was created where that serves the audience. The failure mode worth avoiding is being discovered rather than being transparent.
The useful answer to the question in the title is unromantic. Give a model your brand voice and you get your brand voice, minus whatever made it recognisable, delivered faster and now with a compliance question attached. The teams handling this well have not banned the tools. They have simply kept a named human between the model and the reader, which turns out to be both the editorial answer and the legal one.
Working out which tools deserve a place in the stack?
Knowing what to delegate is only useful once you know what the current generation of tools can actually do.
Sources: Regulation (EU) 2024/1689 (AI Act), Article 50, and the European Commission’s guidance and FAQ on transparency obligations, applicable from 2 August 2026 with machine-readable marking phased to 2 December 2026; Code of Practice on Transparency of AI-generated Content; Google Search Central, “Google Search’s guidance about AI-generated content”, last updated 10 December 2025, and its spam policy on scaled content abuse. This article is general information on a fast-moving regulatory area and is not legal advice; confirm the obligations applicable to your markets before publishing. Updated August 2026.
