Most founders I speak to in Stockholm assume the new EU rules mean they’ll have to stamp “written with AI” on every LinkedIn post. That’s the wrong worry. The rule most likely to catch a founder’s content isn’t about text at all.
From 2 August 2026, Article 50 of the EU AI Act applies. It’s the transparency chapter, the part that says people should know when they’re dealing with AI. It has been sitting on the calendar since the Act entered into force in August 2024, and it landed more or less on schedule, which is not something you can say about the whole regulation.
So here’s what actually changes for a founder who uses AI to make content. And what doesn’t.
What is Article 50 of the EU AI Act?
Article 50 is the transparency section of the EU AI Act, and it applies from 2 August 2026. It covers four situations: systems that interact directly with people, generative systems that produce synthetic content, emotion recognition and biometric categorisation, and deepfakes plus certain AI-generated text. It bans nothing. It says: tell people.
The duties split between two roles. A provider builds an AI system and puts it on the market. A deployer uses one professionally. If you’re a founder generating content for your own business, you’re a deployer. Purely personal use sits outside the Act entirely. The moment it’s part of how you earn a living, you’re in scope.
That distinction matters more than it sounds, because the two roles carry completely different obligations, and a lot of the alarming coverage aimed at businesses is describing duties that belong to OpenAI and Synthesia rather than to you.
Do you have to label AI-written LinkedIn posts?
Almost certainly not. The text-labelling duty in Article 50(4) is narrow, and it only bites when three things are true at once: the text is published, it’s meant to inform the public, and it’s on a matter of public interest.
The Commission spells out what “public interest” covers, and the list reads like a regulator wrote it rather than a marketer: politics and democratic processes, public administration and services, justice and law enforcement, fundamental rights, public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments that might reasonably be the subject of public debate.
A post about your Series A. A carousel on your pricing change. A newsletter explaining what your product does. None of that is a matter of public interest in the legal sense, however important it feels on a Tuesday morning.
And even where text does fall inside the category, there’s an exemption that swallows most honest practice. If the content went through human review or editorial control, and a natural or legal person holds editorial responsibility for publishing it, no label is required. The Commission is specific about what counts. Human review means a deliberate examination of the substance by someone with relevant knowledge and professional judgement. Editorial control means someone with real authority to approve, change or reject the piece on substantive grounds, including fact-checking and checking that sources hold up. Running a spellcheck doesn’t count. Nor does skimming.
Which is a sensible place for the law to land. The thing being regulated is unreviewed machine output published as though it were reported fact, and if you’re writing under your own name about your own business, that isn’t what you’re doing. This is roughly the same line we draw when we talk about ghostwriting posts that still sound like you: the tool can shape the sentences, but the judgement has to be yours or the whole thing collapses.
What about AI avatars and synthetic video?
This is where founders are exposed, and most of them have no idea. The first paragraph of Article 50(4) gets far less attention than the text one, and it carries no public-interest qualifier at all. If you deploy an AI system that generates or manipulates image, audio or video content constituting a deepfake, you must disclose it. Marketing, sales, recruitment, internal. Doesn’t matter.
A deepfake, under the Act’s definition, needs three things to be true: the content closely resembles a real person, object, place, entity or event; that subject plausibly exists or could have existed; and the content would falsely appear authentic or truthful to someone encountering it.
Now think about the tools founders have been quietly adopting. The AI avatar that delivers your script in your face and your voice, so you don’t have to film. The cloned voiceover on a product walkthrough. The dubbed version of your webinar where your mouth moves in sync with Swedish you never spoke. Every one of those resembles a real, existing person. That person is you.
Whether a specific avatar “falsely appears authentic” depends on context, and the Commission guidance says to weigh the level of resemblance, what the content is substantively saying, the deployment context, and what the audience expects. An avatar openly framed as an avatar is a different case from one presented as an ordinary piece to camera. But if your audience believes you sat down and recorded that video, and you didn’t, you’re standing in the exact spot the rule was written for.
One detail catches people out. You cannot rely on the tool’s built-in watermark to satisfy this. Providers carry a separate duty under Article 50(2) to mark outputs in a machine-readable format so they’re detectable as AI-generated. That’s their obligation, not yours, and the Commission has been clear that it doesn’t discharge a deployer’s disclosure duty. Your disclosure has to be something a human being can actually see or hear, at the point they first encounter the content, without needing tools or extra steps.
Didn’t the EU just delay all of this?
Part of it, and that’s where most of the current confusion comes from. In late July the Digital Omnibus on AI became law, and the headline was a delay: the high-risk obligations under Annex III pushed back to December 2027, and the rules for AI embedded in regulated products to August 2028. A lot of coverage compressed that into “the AI Act has been postponed,” and the compressed version travelled further than the accurate one.
It hasn’t been postponed, not the part that touches content. The Article 50 transparency obligations stayed on their original date. The single exception is provider-side: AI systems already on the market before 2 August get until 2 December 2026 to comply with the machine-readable marking duty. A four-month extension, and it belongs to the tool makers.
Everything a founder needs to care about applies from 2 August. Worth remembering when someone tells you otherwise at a startup mixer this autumn.
What are the penalties, and will anyone actually enforce this?
Fines for breaching Article 50 can reach €15 million or 3% of total worldwide turnover for the preceding financial year, whichever is higher. Before that number ruins your afternoon: the Act allows proportionality to be taken into account for SMEs and small mid-cap companies, and enforcement sits mainly with national market surveillance authorities rather than with Brussels.
In Sweden, the proposal designates Post- och telestyrelsen as the main market surveillance authority, with a spread of other regulators covering specific domains. Here’s the honest caveat: at the start of August the Swedish supplementary legislation had not been formally adopted, even though it was proposed to enter into force alongside the Act’s provisions. The rule applies. The local machinery to enforce it is still being assembled.
Nobody is coming for a seed-stage founder’s avatar video this month. That isn’t the point. The point is that disclosure is about to become the visible default across the market, and being conspicuously late to it costs you reputationally long before it costs you legally.
Does anything published before 2 August need relabelling?
No. Content generated before 2 August 2026 doesn’t have to be labelled retroactively, though the Commission encourages it where that’s practical. There’s no archive project waiting for you.
Just a line in the sand, and a decision about which side of it you want to be standing on.
Why should a mission-driven founder treat this as an opening rather than a burden?
Because disclosure and trust are the same product, and trust is what you were already selling.
You Creatives is a Stockholm-based personal branding and content studio for founders. The pattern I notice most often with Nordic founders is a quiet discomfort with AI-assisted content that they never quite say out loud. They use the tools. They also feel slightly uneasy about it, and that unease leaks into the work: hedged posts, publishing delayed by a week and then another week, video that never gets made because the avatar version felt dishonest and the real version felt like too much exposure. It’s the same instinct I wrote about in Stockholm’s quiet founders, wearing a different coat.
A written rule helps with that. If the boundary exists on paper, you can stop renegotiating it with yourself every Monday. Use AI to draft, structure and move faster. Put your judgement into the substance. Say plainly when a face or a voice on screen was synthesised. That’s a workable position and a defensible one, which is more than can be said for the approach most people are running, which is to use the tools quietly and hope nobody asks.
There’s a trade-off here and I’d rather name it than pretend it away. Labelling a synthetic video will cost you some reach and some credibility with audiences who react badly to the letters A and I. Some people will scroll past. That’s a real cost, not a hypothetical one, and anyone telling you disclosure is free hasn’t tried it. My view is that it’s smaller than the cost of being found out later, particularly for a founder whose entire business case rests on being believed. But it is a cost.
The alternative worth considering, if synthetic video makes you uncomfortable anyway, is a format where you’re unambiguously present. Audio is good for this. So is anything where the awkwardness is part of the texture rather than something to be smoothed away, which is why podcasting suits founders who dislike being on camera more than they expect.
What should you actually do this week?
Start with an inventory, because most founders genuinely don’t know what they’ve got out there.
- List every place a synthetic voice or face appears in your content. Avatar videos, cloned voiceovers, dubbed or lip-synced versions, AI-generated images of real people or real places.
- Settle on one disclosure line and use it everywhere. Something plain, like “this video uses an AI-generated presenter.” Put it where a person sees it as they watch, not buried at the bottom of a description.
- Leave your written content largely alone. If you’re reviewing drafts properly and you stand behind what goes out, the text rule is unlikely to reach you. Keep the review real, though. That’s the whole basis of the exemption.
- Write down who holds editorial responsibility. For most founder-led businesses that’s one person, and it’s you. Getting it on paper takes ten minutes and it’s what makes the exemption defensible if anyone ever asks.
- Check what your tools claim. Providers carry the marking duty. It’s worth knowing whether the ones you pay for actually meet it, and whether they say so anywhere you could point to.
None of this is a content strategy on its own. It’s hygiene. But it sits next to a question that matters much more, which is whether the machines that increasingly answer questions about your category have any reason to mention you, and that’s a longer conversation about how AI search actually picks who it cites.
Frequently asked questions
These are the three questions that come up most often when founders realise the date has arrived. Short answers below, with the caveat that none of this is legal advice and the guidance around Article 50 is still settling as national authorities work out their own approach.
Does the EU AI Act apply to a solo founder?
Yes, if you’re using AI as part of a professional or business activity. The Act excludes purely personal, non-professional use, so a private joke image sits outside it. Content you publish to win customers does not. The Commission’s guidance is explicit that where a natural person gains economic benefit on a regular basis from an activity, or is otherwise engaged in a business, trade, occupational or freelance activity, that person counts as a deployer.
Do I have to label content made with ChatGPT or Claude?
Usually not, for written content. The text-labelling duty only applies to text published to inform the public on matters of public interest, and it falls away where the content had genuine human review and someone holds editorial responsibility for it. Image, audio and video are stricter: if the output is a deepfake, meaning it resembles a real person or place and could pass as authentic, you must disclose it regardless of the subject matter.
Does this apply if my company is based outside the EU?
It can. The Act reaches providers established outside the EU where the output of their system is used in the EU, and for deployers the practical question is whether you’re operating in the EU market and exposing people here to the content. If you sell into Sweden or the wider EU, work on the assumption that you’re in scope and take proper advice rather than trusting geography. I’m not a lawyer and this post isn’t legal advice.
Related reading
- Ghostwriting LinkedIn posts that sound like you
- GEO and AEO explained for founders
- Podcast strategy for founders who hate public speaking
- Stockholm’s quiet founders
If you want a content system where AI does the drafting and you stay clearly, honestly at the centre of it, that’s the work we do. Have a look at our services.
