Key Takeaways
From 2 August 2026, the EU AI Act’s transparency obligations apply. If you use AI anywhere in your content production and that content reaches people in the EU, you now have a legal responsibility to label it. The changes aren’t dramatic, but you do need to know what they mean for your content.
What’s changed, and what hasn’t
Most coverage over the past year has been about high-risk AI: recruitment screening, credit scoring, biometrics. The deadline for those obligations has been pushed back. The European Parliament approved the Digital Omnibus amendments on 16 June, the Council signed off on 29 June, and standalone high-risk rules now apply from 2 December 2027, with AI embedded in regulated products following on 2 August 2028. The reason was practical rather than political: the harmonised technical standards that compliance depends on weren’t finished in time.
Article 50, however, comes into force this Sunday, overseen by national market surveillance authorities. Penalties reach €15 million or 3% of worldwide turnover.
This also affects creators based in the UK. No matter where the company is registered, the Act applies if the content is made for an audience in the EU. A Nottingham agency producing content for EU audiences needs to be aware of it, whether that’s for a UK jeweller shipping to Ireland or running Meta ads across Europe.
The four obligations
Chatbots have to identify themselves. If an AI assistant handles enquiries on your site, people need to know they aren’t talking to a person. This is usually straightforward, and most tools already do this.
Generative AI output has to carry machine-readable marking. This one sits with the companies building the tools rather than with you, and systems already on the market before Sunday have until 2 December 2026 to comply. It’s worth asking your suppliers what they’re doing about it, since your evidence of compliance partly depends on theirs.
Deepfakes need a visible label, and this obligation falls on the deployer – you, or whoever produces your content.
AI-generated text published to inform the public on matters of public interest needs labelling too, unless a human has reviewed it and holds editorial responsibility for the publication.
“Deepfake” covers more than you’d think
Most people read “deepfake” and assume the rule is about fabricated politicians and cloned voices. The Act’s definition is much wider than everyday usage. Article 3(60) covers AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic to someone looking at it. It doesn’t matter whether you intended to deceive anyone – what matters is whether the image looks real.
The definition doesn’t stop at people. It also covers objects, places and events.
The Commission’s final guidelines, adopted this month, work through advertising examples directly. An AI-generated image that shows a real product exactly as it is stays outside the deepfake definition. An AI-generated image that makes the product look different from, or better than, it really is falls inside the deepfake definition and needs labelling. Content that’s obviously fantastical falls outside the definition.
A lot of photorealistic jewellery content is now in a grey zone:
- A synthetic model wearing a ring that was never photographed on a hand
- A generated lifestyle scene of a shop interior or a proposal that never took place
- Retouching where the retouching quietly changed the stone, the setting or the metal colour
- AI-generated stock imagery standing in for a real workshop or a real customer
Evidently artistic, creative, satirical or fictional work gets a carve-out, but the guidelines interpret it narrowly and exclude anything primarily commercial. Advertising will rarely qualify. Where content mixes creative and informative purposes, the informative purpose wins.
Your captions and product descriptions almost certainly don’t fall under the text obligation. Marketing copy isn’t published to inform the public on matters of public interest. Sponsored content written to look like editorial coverage is different, though: it’s more likely to count as informing the public, particularly on health or finance topics.
What this means for jewellery and retail brands
Jewellery and retail depend heavily on product photography and lifestyle imagery, so a large share of their marketing content falls inside these rules. A product catalogue can run into hundreds of images, and each one needs checking individually.
Whose job is it – yours or your agency’s?
The person responsible for labelling isn’t necessarily whoever owns the campaign, it’s whoever decided to use AI and controlled how it was used. If an agency generates the images, the agency is usually responsible. If a client specifically asks for AI content and signs off on it, responsibility can shift to them. In practice this needs to be spelled out in the contract. We’ve started doing exactly that in ours.
What we’d do this week
Here’s what we’d prioritise in the next seven days:
- Ask your tool providers what marking and provenance data they apply, and get the answer in writing.
- Work out which published assets were AI-generated, and when. Anything generated before 2 August doesn’t need labelling retroactively. For images the date that matters is the date of generation; for text on public-interest topics it’s the date of publication, so something drafted in July and published in August still needs a label.
- Settle on a labelling convention before you need one. The Code of Practice on Transparency of AI-Generated Content proposes a standard visual “AI” marker that can be translated into different languages. Signing up to it isn’t mandatory, but regulators are likely to treat it as the industry standard, so it’s worth adopting.
- Make sure any chatbot on your website tells visitors it’s an AI, not a person.
Our take
None of this is a reason to stop using AI in content production. It’s a reason to know where you’ve used it. When we ask brands which assets in their library were AI-generated, most don’t have an answer – and that gap is the real problem. Labelling is easy once you have a record. Reconstructing eighteen months of campaign history from memory is not.
If you want a hand auditing what’s in your asset library, or working out which of your content actually needs a label, contact us online, email us at [email protected], or call us on +44 0115 678 7377.
Sources
