Article 50 of the EU AI Act has applied to freelancers, not just agencies, since 2 August 2026. Most commercial work, product pages, pitch decks, falls outside the duty entirely, and only AI-generated text published to inform the public on matters like politics or health needs a label. Genuine human review exempts you, meaning deliberate checking and fact checking, not a spell check or a skim. Fines reach 15 million euros or 3% of turnover, but the real risk for a solo business is a complaint, so document your review step and add a labelling clause to your contract now.
Most freelancers heard that the EU AI Act got postponed and stopped reading there. What got postponed was the high-risk chapter. The labelling rules that apply to people who produce content have been in force since 2 August 2026, and a one-person business is covered by them in exactly the same way an agency is.
The good news: for most freelance work, the duty either does not apply or is satisfied by something you should be doing anyway. The bad news is that "should be doing anyway" now has a legal definition, and spell-checking does not meet it.
What moved in July 2026, and what did not
The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It pushed the high-risk deadlines out by a long way: Annex III systems to 2 December 2027, Annex I safety components to 2 August 2028. White & Case summarised the limit of that in one sentence: the deferral "is limited to Chapter III and does not apply to other obligations relevant to high-risk AI systems (notably the transparency requirements)".
Article 50 sits outside Chapter III. It applied on 2 August 2026 and it applies now.
Three documents are worth bookmarking: the Commission's FAQ, the non-binding guidelines published in July 2026, and the Code of Practice on Transparency of AI-generated Content, published on 10 June 2026 with roughly 190 signatories by the end of July. The Code is voluntary. Orrick's alert makes the point that adherence "does not constitute conclusive evidence of compliance", so treat it as a template rather than a shield.
First question: are you a deployer or a provider?
The two roles carry different duties, and freelancers can end up in both in the same week.
A deployer is anyone using an AI system under their own authority in the course of professional activity. If you use a chat model to draft copy, an image model for visuals or a transcription tool on client calls, you are a deployer. Personal, non-professional use is out of scope, which is why "I only use it for my own newsletter" is a weaker defence than people think if the newsletter is part of your business.
A provider develops an AI system, or has one developed, and places it on the market or puts it into service under their own name or trademark. Building a client-facing chatbot and shipping it under your brand can put you in that category. Praxikon's checklist for small deployers is the right instinct here: map where your work touches Article 50 situations, decide which role you are in for each one, and if you resell or rebrand a third-party system, check whether your branding or modification changes that role.
The four duties, in the order they will hit you
1. Anything that talks to a person. Providers must design systems that interact directly with people so those people know they are dealing with an AI, unless it is obvious to a reasonably well-informed person. If you build the bot, that duty is yours, not your client's.
2. Machine-readable marking of synthetic output. Also a provider duty: audio, image, video and text output must be marked as artificially generated in a machine-readable format, as far as technically feasible. Assistive editing tools that do not substantially alter the input are carved out. For systems already on the market before 2 August 2026, providers have until 2 December 2026 to get the marking in place. You are not the one implementing this, but it is worth asking a vendor about it before you build a client workflow on top of them.
3. Emotion recognition and biometric categorisation. Deployers must inform the people exposed to these systems and handle the data under the GDPR. Niche, but not hypothetical if you work in UX research, retail analytics or event technology.
4. Deepfakes and public-interest text. This is the one that reaches content people. Deployers must disclose AI-generated or manipulated image, audio or video content that resembles real people, places or events. And AI-generated or manipulated text published to inform the public on matters of public interest must be labelled, unless it went through human review. The Commission's FAQ names politics and democratic processes and public health among the examples.
Note what is not in that list. A client's product page, an onboarding email, a pitch deck, a landing page: none of that is text published to inform the public on a matter of public interest. Most commercial copywriting sits outside the duty entirely. Press releases, health and finance explainers, and anything published under a client's name to inform a general audience are where you should stop and think.
The exemption most freelance work lives in
Article 50 lets published text off the hook when it "has undergone a process of human review or editorial control" and a person or entity holds editorial responsibility for it.
That sounds soft until you read what the regulators mean by it. The Commission describes a "deliberate examination of the substance of the content" by natural persons with relevant expertise, and says "superficial, solely formal, or procedural checks" do not qualify. Orrick's reading is blunter: spell checking, grammar correction and automated review processes fall short, and fact checking is a minimum, not a bonus.
So the exemption is not a formality you claim. It is work you do.
Julien Look, founder and CTO of Look Beyond, spent his Freelance Unlocked session on why 95% of corporate AI projects produce no measurable value. His diagnosis was not technical. Using Siemens as an example, he described what actually kills adoption:
"They saw that people are just not trusting the output. There will be hallucinations, there will be interpretations of data that might not fully align with what the people in the organization are used to, and that creates mistrust." (Julien Look)
His answer to it reads like a paraphrase of the AI Act:
"You need to make people aware of the limitations and also the possibilities of what your agent can do, and you need to make it transparent. You need to make the output traceable." (Julien Look)
Traceability, review, someone who can be named as responsible. The law and the practitioner arrived at the same shortlist from opposite directions.
Jan Schlie, VP of AI at Jimdo and a solopreneur for four years before that, made the same argument from the delivery side. His session was about giving AI enough context about your business that the output stops being generic, and he opened with what happens when you skip that step and paste a client enquiry straight into a chat window:
"It knows nothing about my company, it doesn't know my pricing. That answer is useless, because it's supposed to save me time and actually it makes more work for me." (Jan Schlie)
He hears the same complaint from the self-employed people Jimdo works with: "It's supposed to write me an Instagram post, but I have to rewrite it, because that's not how I talk at all."
If the raw output is not usable, then the rewrite you were going to do anyway is your editorial review. Write it down as a process step, name the person doing it, and the exemption is documented rather than assumed. Jan's other line is a useful warning against treating that as a one-off:
"You have to get into the mode where the setup is never finished. It keeps developing, it's alive." (Jan Schlie)
Where the label goes when you do need one
The law says disclosure must be clear, distinguishable and given "at the latest at the time of the first interaction or exposure". Orrick's summary of the guidelines gets more specific for published text: an icon or label above or at the top of the text, near the headline, or in the colophon. For short text, a contextual notice in the interface is acceptable. For anything non-digital, exhibitions, galleries, cinemas, the disclosure goes at the point of entry, in accompanying materials or on the physical carrier.
Creative work gets a lighter touch. Where a deepfake is part of an evidently artistic, creative, satirical or fictional work, the disclosure has to be made in a way that "does not hamper display or enjoyment" of it. A caption under the poster, not a watermark across it. The rights side of the same question is covered in our piece on AI and copyright for creative freelancers.
Who enforces this in Germany
Since 29 July 2026 the answer has a name. The German implementing act, the KI-MIG, took effect on that date and, according to the Bundesnetzagentur, makes the agency Germany's central market surveillance authority for the AI Act. It also runs a KI-Servicedesk aimed explicitly at SMEs and startups, with a risk classification tool and practical guidance, and it acts as the central complaints body: individuals and companies can report suspected violations there.
That last part matters more for a solo business than the fine ceiling does. Article 50 breaches carry fines of up to €15 million or 3% of worldwide annual turnover, and the FAQ notes that proportionality for SMEs is built in. Nobody is coming for a freelance copywriter with a €15 million penalty. A competitor or an annoyed former client filing a complaint is the realistic scenario, and the cost of that is your time and your reputation.
Put it in the contract before it comes up
Clients rarely raise this, which means the freelancer who raises it first looks like the professional in the room. Two or three sentences in the offer are enough:
The contractor may use generative AI systems in producing the deliverables. All deliverables undergo substantive review by the contractor before handover, including fact checking, and the contractor holds editorial responsibility for their content within the meaning of Article 50(4) of Regulation (EU) 2024/1689. The client informs the contractor if a deliverable is intended for publication to inform the public on a matter of public interest, or if it contains representations of real persons, places or events. In those cases the parties agree the labelling before delivery, and the client is responsible for the labelling remaining in place after publication.
The last clause is the one that saves you. You cannot control what a client does to your file after handover, so allocate that risk in writing instead of hoping.
What to do on Monday
- List every AI system in your delivery chain, including the ones inside tools you did not think of as AI: transcription, translation, image upscaling, voice cleanup. For each one, write down whether you are the deployer or the provider.
- Sort your last ten deliverables into two piles. Public-interest text and realistic depictions of real people or events go in the labelling pile. Everything else, which will be most of it, goes in the other one.
- Write your review step down as a process. Who reads it, what they check, that fact checking happened. One paragraph in your project template. That paragraph is what the exemption rests on.
- Add the labelling clause to your offer template, and send it to your two biggest clients before your next project starts rather than during one.
- Bookmark the Bundesnetzagentur's KI-Servicedesk if you work in Germany. It was built for businesses your size, and it is free.
The pattern under all five is the one both speakers landed on independently: a human who actually looked, and a record of it. Our piece on whether to tell clients you use AI covers the trust research behind the same decision, including what happens when a client finds out from someone else.
If you would rather work with companies who already know how you operate, create a free profile on 9am and put your terms where they can read them before the first call.
Freelance Unlocked is co-organized by 9am together with Uplink and freelancermap. This article draws on the sessions of Julien Look and Jan Schlie at Freelance Unlocked 2026. Watch the full talks above, and join us at the next edition: freelanceunlocked.com.
