TL;DR: yes, but less than the panic suggests. From 2 August 2026 the EU AI Act imposes transparency duties (Article 50) - yet for a company that publishes content made with third-party tools, they boil down to one thing: you visibly label deepfakes, meaning realistic images, video and audio that someone could mistake for authentic. Illustrations, abstractions and obvious fiction need no visible label. Machine-readable marking inside files is the tool vendors’ duty (from 2 December 2026) - your job is not to strip that metadata. Violations carry fines of up to EUR 15 million or 3% of turnover, so the topic is not to be ignored - but a sensible company handles it with one simple procedure, not by plastering labels on everything. (As of July 2026; not legal advice.)
What exactly does the law require from August 2026?
Transparency - in two different roles that must not be confused. Article 50 of the AI Act splits the duties between whoever builds the tool and whoever publishes the content:
| Duty | Who it applies to | What it means |
|---|---|---|
| Machine-readable marking (Art. 50(2)) | The generative tool’s provider (OpenAI, Google, Adobe…) | Invisible metadata or a watermark in the file - content must be detectable as generated |
| Visible deepfake disclosure (Art. 50(4)) | The company publishing content - that is you | A visible notice that the content is artificially generated or manipulated - deepfakes only |
| ”You are talking to AI” notice (Art. 50(1)) | The company running a chatbot | A clear message at first contact, unless it is obvious |
Deadlines after the recent amendment: the Digital Omnibus package, given its final green light in June 2026, postponed the high-risk provisions but left transparency almost untouched. The duties for publishers (deepfakes) and chatbots apply from 2 August 2026, while tool providers received a transition period for machine-readable marking until 2 December 2026. The sanctions are real: breaching the Article 50 transparency obligations carries a fine of up to EUR 15 million or 3% of total worldwide annual turnover - whichever is higher.
Good news for most companies: if you use Midjourney, Firefly or a video generator, you are only the publisher in this equation. We map the full set of AI Act roles and duties in Does the EU AI Act apply to my company.
What is a deepfake under the AI Act - and is your marketing one?
The definition is broader than the colloquial one and is not just about politicians’ faces. Under Article 3(60), a deepfake is AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events - and would falsely appear to a person to be authentic or truthful. The key words are “resembles” and “appear to be authentic”. Intent to deceive is not required - realism is enough.
For marketing this draws a simple line:
- a photorealistic product render, a generated “photo” from your production floor, the face of a non-existent “employee”, an avatar voiceover presenter - these are deepfakes within the meaning of the provision: label them visibly,
- an illustration, a stylized collage, an abstraction, obvious fantasy (a physically impossible character, a cartoon world) - it does not resemble a reality that could be taken as authentic: no visible label, metadata stays.
Mind the “artistic exemption” trap: the provision softens the duty for content that is evidently artistic, satirical or fictional - there, disclosure in a way that does not hamper the enjoyment of the work is enough. Advertising, however, is not an “artistic work” in that sense - realistic commercial material does not benefit from the exemption. Assume conservatively: if a customer could think “that’s a photo”, treat it as a deepfake.
Do I have to label this material visibly? A decision tree
Run every image, video and audio through four questions:
- Did AI generate or substantially alter this content? No - do nothing. Retouching and standard editing (background removal, colour correction of a real photo) is not a “substantial alteration”. Yes - continue, and do not strip the metadata.
- Is the content photorealistic and does it resemble real people, objects, places or events - could it pass as authentic? Yes - it is a deepfake: label it visibly.
- Is it evidently artistic, satirical or fictional, recognizable as such? A minimal, non-intrusive disclosure that does not spoil the work is enough.
- Is it pure fantasy or abstraction, physically impossible? No visible label - the file metadata stays.
Text is treated even more gently: the duty covers text published to inform the public on matters of public interest - and it lapses where the text has undergone human review and editorial responsibility. A product description or a blog post your team edited and stands behind needs no visible label.
How to label in practice - compliant and clean?
Visibly where required - and so the label is clear at first contact with the content:
- image - a permanent caption in the corner or beneath: “Image generated with AI”,
- video - a card or disclaimer at the start: “This video contains AI-generated content”,
- audio - a short announcement at the beginning of the track,
- a website footer is not enough - the disclosure belongs with the content, not in terms three clicks away.
The other half of the story happens in metadata - where your role is passive but important. The industry marks files with C2PA (Content Credentials) - an open provenance standard developed by Adobe, Google, Microsoft and OpenAI among others and watermarks such as Google DeepMind’s SynthID, imperceptible to humans but machine-detectable in images, audio, video and text. OpenAI embeds C2PA and SynthID in generated images, and Adobe Firefly automatically attaches Content Credentials to fully generated content. Your job: choose tools that do this, and do not strip metadata in post-production - aggressive exports, harsh compression or screenshots can cut a file’s provenance, and then the marking has to be restored manually.
Won’t an “AI” label hurt my brand?
A fair question - and the best argument for starting with how you work, not with a sticker. On one hand, nearly 90% of consumers want to know whether an image was created using AI; on the other, in experiments merely labelling content as AI-generated lowered its credibility and people’s willingness to share it. The law says label; audiences penalize labels. The way out?
The way we work ourselves: we do not create materials entirely with AI - we assemble. The base is a real product photo, the client’s actual render, footage from a shoot; AI removes backgrounds, repairs quality, adds elements, brings a frame to life. A designer composes the final piece and owns the result. Such material is usually not a deepfake at all - it is not a synthetic fabrication of reality but a processed real asset - so the visible-label duty does not cover it, and the tools’ metadata stays in the file. The mandatory label remains for fully synthetic content used deliberately. Why generating everything from scratch is a bad idea anyway, we cover in Will AI replace graphic designers.
There is one more reason for caution, outside the AI Act: if a generated “product” differs from the real one (different proportions, material, colour), you enter unfair commercial practice territory - regardless of whether you label the material.
What about chatbots? A footnote from the same article
The same Article 50 covers conversational bots: a person interacting with an AI system must be informed of it, unless it is obvious from the context. If your website runs an AI-powered chat - a “you are talking to an AI assistant” notice at first contact settles it. A bot posing as a human (“Kate from customer service”) is asking for trouble after 2 August 2026.
Want to be sure your materials and chatbot are ready for August 2026? We will go through your actual publications during a free consultation.
Frequently asked questions
Do ChatGPT-written texts on a company blog need labelling? No, as long as they pass human review and the company takes responsibility for them - the editorial exemption is written into the provision. The duty targets unsupervised “farms” of public-interest text, not edited content marketing.
Who is liable for missing metadata in a file - me or the tool vendor? Machine-readable marking is the tool provider’s duty (from 2 December 2026). You are responsible for visibly labelling the deepfakes you publish - and, in your own interest, for not stripping metadata from files.
Does retouching a photo in Photoshop make it “AI content”? No - the provision exempts assistive functions for standard editing that do not substantially alter the content. Removing a blemish or sharpening is editing. Generating a new background full of people who were never there is a substantial alteration.
Key takeaways
- You label deepfakes, not everything - the visible label applies to realistic content that could pass as authentic; illustrations and abstractions are out of scope.
- Two duties, two roles - in-file metadata is the tool vendor’s job (from 2 Dec 2026); the visible label on a published deepfake is yours (from 2 Aug 2026).
- Advertising gets no artistic exemption - a realistic product render or a synthetic face in a campaign requires disclosure.
- Do not strip metadata - C2PA and SynthID in files are your evidence of due diligence; exports and compression can cut them out.
- Label at the content, not in the footer - the disclosure must be clear at the audience’s first contact with the material.
- The best strategy is how you work - materials composed from real photos plus AI elements usually fall outside the duty, and they hold up on quality too.
- A chatbot must introduce itself - “you are talking to AI” at first contact; a bot posing as a human is a legal risk.