Skip to content
Implementa.
OpinionGEO··10 min

Labeling AI-generated content at your company: when the law requires it and when it's theater

Labeling AI-generated content at your company doesn't mean stamping «made with AI» on everything. Article 50 of the EU AI Act asks for three specific things: disclose deepfakes, disclose public-interest text nobody reviewed, and mark output in a machine-readable way. The thesis: the decorative sticker satisfies none of them, and the mark that does matter is due for many providers on December 2, 2026.

Implementa Editorial

Editorial

Illustrative scenario. The marketing team at a 200-person company decides to «get compliant on AI» and adds the line «Content created with AI» to every LinkedIn post, every blog article and the caption of every campaign image. Very responsible. Meanwhile, the holiday video where the CEO thanks customers with a cloned voice —because he was traveling— goes out with no disclosure at all. And the images generated for the website run through the CMS, which resizes them and, along the way, strips the metadata that identified where they came from.

The thesis in one line: the law doesn't ask you to label everything an AI touched; it asks for three very specific things, and the decorative label satisfies none of them. Labeling just in case isn't caution: it's theater, and it buries the one label that actually had to be there. A note before we go on: this describes the EU AI Act as it applies today and isn't legal advice; for a specific case, talk to counsel. If you sell into the EU from the US, it applies to you too.

When does a company have to label AI-generated content?

The answer sits in Article 50 of Regulation (EU) 2024/1689, which has applied since August 2, 2026. It has four transparency duties. Two aren't about content —the chatbot that says it's a chatbot and emotion recognition that must be disclosed—, and we covered them in what the AI Act postponed and what it didn't in August 2026. The other two are the ones that hit marketing, comms and anyone who publishes:

  • The machine-readable mark (Article 50(2)). It's the provider's duty for any system that generates synthetic audio, images, video or text: the output has to be marked so another machine can detect it as artificial. It's not visible text; it's metadata or a watermark inside the file.
  • The visible disclosure (Article 50(4)). It's the deployer's duty —the company using the tool— and only in two cases: when it publishes a deepfake, and when it publishes AI-generated text meant to inform the public on matters of public interest, unless a person reviewed it and someone holds editorial responsibility.

That's everything Article 50 requires on content. It doesn't say «label every post a model helped write». It doesn't say «disclose that you used AI to fix a typo». And paragraph 5 adds the how: the disclosure must be clear and distinguishable, and arrive at the latest at first exposure. Not on the legal notice page, not in the third paragraph.

Content by content: what needs a disclosure and what doesn't

ContentVisible disclosure under Art. 50(4)?Why
Video with the cloned voice or synthetic face of a real person saying something they never saidYesTextbook deepfake: it resembles an existing person and can pass as authentic.
Event photo edited with AI to change what happened (bigger crowd, different venue)YesIt manipulates a real event in a way someone could believe.
Image of your real product in a setting that never existed, made to look like a photoProbably yesThe definition covers existing objects and places; if it looks like a real photo of your product, it's close.
Illustration, icon or abstract background nobody would take as realNoIt doesn't resemble anything existing or mislead anyone about its authenticity.
Obviously satirical or fictional ad with a synthetic characterYes, but lightDisclosing that generated content is present, without spoiling the work, is enough.
Blog article drafted with AI, reviewed by a person and published under the company's nameNoEven if it were public interest, there's human review and editorial responsibility: the literal exception.
Statement or notice generated and published without anyone reviewing itYes, if it informs on a matter of public interestThat's exactly the case the rule is written for.

Two rows deserve a warning. The product row is the one that surprises people, because almost everyone pictures deepfakes as politicians' faces; but the regulation's definition (Article 3(60)) covers content that resembles existing persons, objects, places, entities or events and could falsely appear authentic. And the statement row depends on what «public interest» means, which the regulation doesn't define for a company's text: our reading —inference, not doctrine— is that an ordinary marketing post isn't, while a notice about an incident, a product recall or a change that affects customers can be.

What counts as a deepfake, in one question

Forget whether the tool «generated» or «just edited». The operational question is: could a reasonable person believe this existed or happened as shown? If yes, and an AI produced or altered the image, audio or video, it gets a disclosure. If no —because it's obviously an illustration, a stylized render or a collage nobody would mistake for a photo— it doesn't, at least not under this rule.

The point of framing it this way is that you can delegate it. Legal doesn't need to review every asset: whoever publishes needs to ask that question, and the doubtful case needs somewhere to go. If that isn't written down anywhere, the problem isn't transparency; it's governance and control of AI automation.

Text: the exception almost nobody reads right

The text paragraph is the one producing the most unnecessary labels. The duty only covers text published «with the purpose of informing the public on matters of public interest», and it falls away if the content went through human review or editorial control and a natural or legal person holds editorial responsibility for the publication. Translated: the article a model drafted, your team edited and your company signs and stands behind doesn't need, under Article 50, any «written with AI».

But the exception comes with a condition good intentions don't meet: the review has to exist. If someone asks tomorrow who reviewed a piece and when, «someone on the team looked at it» isn't an answer. A minimal record —who approves, against what criteria, at what point in the flow— is what turns the exception into something you can defend. It's the same logic we lay out in the human in the loop of an AI automation: a review that leaves no trace doesn't count as a review.

The part that takes real work: the invisible mark due on December 2

The expensive duty isn't the visible label: it's the machine-readable mark in Article 50(2), and it belongs to the provider. If you use a commercial tool to generate images, the mark is the vendor's job. But if your company built its own system that generates content —a product-description generator you offer to distributors, an app that produces creative for clients— you may be the provider, and it's worth knowing before someone asks. We dig into it in provider or deployer under the AI Act.

The calendar has a nuance people keep misreading. Article 50(2) has applied since August 2, 2026, but the Digital Omnibus on AI gave generative systems already on the market before that date a transition period: they have until December 2, 2026. According to Bird & Bird's analysis of the May 2026 agreement, that window is shorter than the February 2027 date the Commission had proposed. And the Cloud Security Alliance note of September 21, 2026 stresses the part people get wrong: the extension covers only the technical mark; deepfake and chatbot disclosures were never postponed. Breaching Article 50 falls in the fine tier of up to €15 million or 3% of worldwide annual turnover (Article 99(4)).

There's a second problem, and this one is yours even if you're not the provider: the mark has to survive your pipeline. The same CSA note points out that resizing, format conversion, re-encoding on upload to a social network or a screenshot strips or invalidates provenance metadata. If your CMS cleans the metadata off every image it ingests, the mark the provider added never reaches anyone.

To make this concrete, in June 2026 the Commission published a code of practice on marking and labelling AI-generated content, voluntary, with common EU icons for the visible disclosure. If you're going to label, label with that: an icon readers recognize informs more than a sentence every team words differently.

Why labeling everything is worse than labeling nothing

This is opinion, and we stand by it. A label does one job: keep someone who could be fooled from being fooled. If «made with AI» shows up on the job-posting post, the quarterly infographic and the cloned-voice video, readers learn to ignore it within a week. And the day the label matters —the video that really could pass as real— it no longer informs anyone.

The universal sticker has a worse internal effect too: it feels like compliance. The team thinks it's done, and nobody checks whether images carry the technical mark or who reviews the text. It's the usual pattern: the visible action replaces the one that works. No theater here: label what the rule requires, mark what the technology requires, and everything else is a brand decision you make once and write down.

It isn't a visibility problem either. Google put it in writing in its February 2023 guidance on AI-generated content: it rewards helpful content however it's produced; what counts is who signs, what experience they show and whether someone stands behind what's published, which is also what an AI assistant weighs when deciding whom to cite, as we explain in the E-E-A-T signals that get an LLM to cite you. Signing and standing behind it is, conveniently, the same thing that triggers the Article 50 exception.

Four questions before you publish anything

  1. Is it an image, audio or video that resembles a real person, place, object or event and could pass as authentic? If yes, visible disclosure at first exposure, with the common icon if you can.
  2. Is it text that informs the public on a matter of public interest? If yes, either a named person reviews it and the company signs it, or it carries a disclosure.
  3. Is it generated by a system you built and others use? If yes, you're the provider: machine-readable mark, and if it was on the market before August, by December 2, 2026.
  4. Does the technical mark survive your publishing pipeline? Take a file generated last week, upload it the usual way and check whether the mark is still there on the other side.

If all four answers are no, Article 50 doesn't require a disclosure. You add «made with AI» if your brand wants it, not because you have to.

What to take away

  • Labeling AI-generated content at your company isn't a blanket duty: Article 50 asks you to disclose deepfakes and unreviewed public-interest text, and to mark output technically.
  • Text reviewed by a person and signed by the company needs no disclosure under Article 50; the review has to be provable.
  • The machine-readable mark is the provider's job and, for systems already on the market, is due December 2, 2026; the rest of Article 50 has applied since August 2.
  • Your CMS may be erasing the mark: test it with a real file.
  • The universal label doesn't protect you: it trains readers to ignore the one that matters.

If you can't say today which tools generate content at your company, who reviews what gets published and whether the technical mark reaches the website intact, don't start with the label: start with the inventory and put a name next to each flow. That's what we ship in AI Act compliance by operating AI, with data governance underneath. The sticker comes last, and it's the cheapest part.

Shall we get it shipping?

If this resonated, 30-minute conversation with no commitment. We tell you what fits, what doesn't and the approximate price.

See cases
Labeling AI-generated content at your company: when the law requires it and when it's theater · Implementa