AI Labeling Requirement: Why You Should Mark July 22 on Your Calendar

If you've been following newsletters, LinkedIn feeds, or trade publications on AI regulation over the past few weeks, you've probably come across two dates mentioned in the same breath: July 22, 2026 and August 2, 2026. Both concern the labeling requirement for AI-generated content. Both are frequently confused. And this mix-up is currently causing unnecessary uncertainty for many Mittelstand businesses that really just want to know: do I need to do something now, and if so, what?

The short answer up front: July 22 is important, but not in the way it's often portrayed. The actually binding deadline, with the risk of fines attached, is August 2, 2026. Anyone who keeps the two dates cleanly separated will save themselves a fair amount of stress in the coming weeks.

Two dates, one misunderstanding

On June 10, 2026, the EU AI Office published the final version of the "Code of Practice on Transparency of AI-Generated Content." Companies wanting to register as so-called "initial signatories" can do so until July 22, 2026, 6:00 PM CEST. Anyone who misses this deadline is by no means excluded: signing is possible at any time afterward via an email form, just no longer as an "initial signatory."

One detail that gets lost in many articles is decisive here: the code is explicitly voluntary. The European Commission itself describes it as a "voluntary tool" that creates no obligations beyond the AI Act, no new requirements, and no additional administrative burden. Anyone who doesn't sign is not automatically in violation of anything.

The Code of Practice is a "voluntary tool" - it is meant to facilitate and demonstrate compliance with Article 50 of the AI Act, but it does not itself create any additional legal obligations.

The actually hard, statutory deadline is a different one: from August 2, 2026, the transparency obligations under Article 50 of the EU AI Act take effect bindingly - regardless of whether a company has signed the code or not. This obligation has also been explicitly left untouched by the so-called Digital Omnibus, which pushed some other AI Act deadlines to December 2027 and August 2028 respectively. August 2 stands.

For your business, this means: July 22 is a good occasion to engage with the topic - not because an obligation arises on that day, but because from that point on, practically only three weeks remain until the actual deadline. Anyone who uses the time until then to review their own labeling practice will be relaxed on August 2. Anyone who waits will be under time pressure.

This confusion between the voluntary code and the binding obligation is, incidentally, not a niche problem confined to a few blog posts - it is explicitly named as a widespread misunderstanding in several expert sources. That's understandable: both regulations concern the same subject area, both are carried by the EU AI Office or the European Commission respectively, and they lie only two weeks apart. Anyone hearing secondhand about "the deadline in July or August" can hardly tell whether that means a voluntary registration or a statutory obligation. That's exactly why a clean separation is worthwhile here: July 22 concerns a signature, August 2 concerns your actual conduct.

What the Code of Practice actually regulates

The code addresses three target groups: providers of generative AI systems that produce synthetic audio, image, video, or text, commercial deployers of such systems, and technology providers of labeling, detection, watermarking, or provenance solutions. For most Mittelstand businesses, the second group is the relevant one - you use generative AI tools in your own marketing or customer service, but don't develop them yourself.

In terms of content, the code distinguishes two levels of labeling. For the technical, machine-readable marking of AI outputs, C2PA (Content Credentials) is named as the only technology so far that meets the criteria defined in the code for digitally signed metadata. CAWG metadata assertions and IPTC provenance standards are also recommended in addition, along with invisible watermarks as a further machine-readable marker. This technical level primarily concerns the providers of the AI systems themselves - i.e. the makers of image generators or language models - and less so the companies that use these tools day to day.

For the second, visible level of labeling, the code envisages uniform, EU-standardized icons or symbols plus accompanying explanatory text, so people can recognize AI-generated or AI-manipulated content. Providers who sign the code are also expected to make publicly accessible, free detection systems available from February 2027, which can be used to verify AI-generated content.

Important for context: there is another, much older "General-Purpose AI Code of Practice" from 2025, which deals with safety and systemic risks in large AI models and which, among others, Anthropic, Google, Microsoft, OpenAI, and Amazon have signed - Meta, however, has not. This code is not identical to the transparency code for AI-generated content discussed here and should not be conflated in internal communications.

As for the concrete question of who has actually signed the new transparency code, there is currently no publicly accessible, verified list of signatories. For practical work in the Mittelstand, however, this is of secondary importance anyway: as described above, the actual obligation does not arise from a signature, but directly from Article 50 of the AI Act. Whether a large provider such as a particular image generator or language model maker has signed the code or not changes nothing about the fact that you, as a deployer of these tools, must meet your own labeling obligations.

What Article 50 of the AI Act specifically requires

So you can assess where in your business action is actually needed, it's worth looking at the four paragraphs of Article 50, which form the actual statutory core:

Paragraph 2 - chatbot labeling: Users must be made aware that they are interacting with an AI system and not a human. Applies in practice to almost every chatbot on your website or in your customer service.

Paragraph 4, first part - deepfake labeling: Realistic-looking, AI-generated or AI-altered images, videos, or audio content depicting real people, places, or events must be labeled as such.

Paragraph 4, second part - text on matters of public interest: AI-generated texts on matters of public interest generally require labeling.

Paragraph 5 - technical marking: AI outputs are meant to be marked in a machine-readable and recognizable way - this primarily concerns providers of the underlying systems.

One exception in paragraph 4 is particularly relevant for many companies in content marketing: the labeling requirement for text does not apply if the content has undergone human review or editorial control and a natural or legal person bears editorial responsibility for it. So if you use AI as a drafting aid and then revise, check, and personally sign off on the text yourself, it generally remains exempt from labeling.

What this concretely means for your Mittelstand business

Translated into everyday business practice, three areas are likely to be front and center for most Mittelstand companies: chatbots on your own website or in customer communication, AI-generated or AI-edited product images and social media graphics, and AI-assisted marketing texts.

With chatbots, the matter is comparatively clear: a brief note such as "Hello, I'm [company]'s AI assistant" generally meets the requirement under paragraph 2. With image and video content, the situation is stricter - here you generally need a visible notice, for example in a caption, when an image is recognizably AI-generated or AI-altered and realistically depicts real people, places, or events. The technical, machine-readable marking via watermarks or C2PA metadata is predominantly a task for the tool providers, such as the makers of image generators, not your task as a user of these tools.

For marketing texts, as described above, the editorial exception frequently applies. If a human in your content process reads, revises, and finally approves the AI draft, the labeling requirement no longer applies to that text. That covers a large share of typical use cases in content marketing - from blog articles to newsletters to product descriptions.

A practical first step that has proven useful in practice is a simple AI inventory: where exactly do you actually use generative AI? This concerns not only the obvious tools like ChatGPT or image generators, but also AI features that are long since embedded in CRM systems, email programs, or social media tools, often without the team being consciously aware of it. Only once you know where AI is actually being used can you assess where a labeling requirement might even apply.

The question of who, in the legal sense, is a "provider" and who is a "deployer" of an AI system also deserves a brief note here: this distinction is complex in the AI Act on a case-by-case basis and depends, among other things, on whether you merely use an existing model or substantially modify it under your own responsibility and offer it under your own name. The statement that most Mittelstand companies act only as deployers is therefore a simplified categorization for everyday practice and does not replace legal advice in an individual case - especially not if you yourself integrate AI features into your own products and pass them on to third parties.

Fines and the relief provided by the Digital Omnibus

Violations of the transparency obligations under Article 50 can, under the AI Act, be punished with fines of up to €15 million or 3 percent of global annual turnover - whichever is higher. That's the middle of three fine tiers under the AI Act; the top tier, which applies to prohibited practices, reaches up to €35 million or 7 percent of global annual turnover.

That sounds like a significant figure at first, and it is - but the Digital Omnibus provides reduced, proportionate requirements for small and medium-sized enterprises, and, newly, also for so-called "small mid-caps." A small mid-cap is defined as a company with up to 749 employees and either up to €150 million in annual turnover or up to €129 million in balance sheet total. For these companies, simplified documentation is envisaged, among other things, along with a tendency to apply the lower fine amount rather than the percentage. In addition, a roughly 30-day grace period with a warning is provided for before any actual sanction is imposed.

This does not change the fundamental obligation from August 2, 2026, but it noticeably eases the risk for the classic Mittelstand. It's also important to note: not signing the code is not an automatic violation. Anyone who doesn't sign the code merely has to be able to demonstrate compliance with Article 50 by other means, and in doing so tends to risk more follow-up questions and reviews from the relevant market surveillance authorities - in doubtful cases, also an individual gap analysis by the authority.

Another relief specifically concerns the technical watermarking obligation: AI systems that were already placed on the market or in service before August 2, 2026, get a transition period of roughly four months for the technical marking obligation, until December 2, 2026. New systems introduced only from August 2, 2026 onward, however, must label immediately. For most Mittelstand businesses as users of existing tools, this transition period should provide a bit of breathing room - but for visible user-facing labeling, such as on chatbots or image captions, August 2 applies unchanged.

What you should concretely do now

Let's summarize the practical steps that follow from the deadlines at hand. First: get an overview of where generative AI is actually being used in your company - from the obvious chatbot solution to inconspicuous AI features in existing software. Second: check for each of these use cases whether and in what form labeling under Article 50 is necessary - and whether the editorial exception applies for text content because a human reviews and approves the content anyway. Third: for visible image and video content, clarify what an uncomplicated but legally sound labeling approach could look like within your existing content production.

Signing the Code of Practice itself is not a necessary step for the vast majority of Mittelstand businesses, since you generally do not act as a "provider" of AI systems within the meaning of the AI Act, but as a deployer of existing tools. You achieve compliance with Article 50 through your own labeling practice, not through a signature under the voluntary code. July 22 is thus above all a good reminder in the calendar - an occasion to review your own practice in good time before August 2, rather than confusing it with a deadline of your own that doesn't actually exist for your business.

August 2, 2026 is getting closer, regardless of how much attention July 22 still gets in the coming weeks. Anyone who uses the time in between to soberly map their own AI usage and adjust their labeling practice accordingly has nothing to worry about on the deadline - regardless of whether a voluntary code was signed or not.