What the New AI Disclosure Requirement Means for Your Website
The labeling requirement for AI-generated content has actually been in effect for a few weeks now—and it’s far less dramatic than the initial headlines suggested. This article explains who the requirement really affects, what exceptions exist, and what this means in practice for the text, images, and chatbots on your website.
Date
Category
Author
Labeling AI Content: What Will Actually Apply Starting in August 2026
“Do we now have to include a note with every blog image stating that it was created using AI?” This question has come up repeatedly in recent weeks, ever since the new labeling requirement for AI-generated content officially took effect on August 2, 2026. There was a lot of speculation about it beforehand, but far less concrete discussion about what this actually means for your company going forward.
Incidentally, this isn’t a standalone new law. The labeling requirement is set forth in Article 50 of the European AI Regulation—that is, the same AI Act that has been under discussion for some time and which we’ve already covered in more detail here on the blog. What’s new is that this requirement is now truly in effect, backed by specific implementation guidelines and a few exceptions that often get overlooked in the public debate.
This post explains what has actually been in effect since August 2, who is subject to the requirement and who is not, and what this means in concrete terms for the various types of content on your website.
A quick note to start with: This post is not a substitute for legal advice. For a definitive assessment of your specific case, we recommend consulting a lawyer who specializes in IT and data protection law.
What Article 50 of the AI Regulation Actually Requires
Article 50 essentially distinguishes between two different obligations that are often lumped together. The first concerns interactions with people: A chatbot or virtual assistant must be designed so that visitors can tell that they are speaking with an AI system and not with a human. Legally, this obligation falls on the system provider. However, anyone who places a chatbot on their website under their own name is often the provider themselves or must at least clarify the disclosure with the service provider. A brief, clearly visible notice in the chat window is generally sufficient for this purpose—preferably right in the first message. A sentence in the terms and conditions or a general statement such as “This website uses AI,” on the other hand, is not sufficient.
The second requirement concerns the generated content itself: Providers of AI systems must mark text, images, audio, and video content generated by their systems as “AI-generated” in a machine-readable format. This is the responsibility of the tool manufacturers, not your website. So-called “deepfakes”—that is, AI-generated or altered image, audio, or video content that resembles real people, objects, places, institutions, or events and would appear to be genuine—are subject to particularly strict regulations. The party publishing them—that is, the operator—must additionally label them in a way that is visible to humans, not just machine-readable in the background.
Both requirements have been in effect since August 2, 2026. However, for providers of AI systems that were already on the market prior to that date, there is a separate four-month transition period ending on December 2, 2026, specifically for the technical implementation of content labeling.
Who is subject to this requirement—and, more importantly, who is not
One point that often causes confusion in initial discussions: The labeling requirement primarily applies to providers and operators of AI systems, not automatically to every company that occasionally uses an AI tool for text or images. A company that uses AI tools such as image generators or writing assistants is generally considered an operator. However, an operator is only required to label two types of content: deepfakes and texts published to inform the public about matters of public interest. In addition, there is a disclosure requirement when using emotion recognition or biometric categorization, which rarely occurs on standard websites.
However, there is one important exception that is particularly relevant to texts: The labeling requirement for texts applies only to published texts intended to inform the public about matters of public interest, such as health, consumer protection, politics, or the economy. According to the Commission’s guidelines, advertising copy and product descriptions are generally not covered by this requirement. And even for these texts, the requirement does not apply if a person has actually reviewed the content and a natural or legal person bears editorial responsibility. This exception stems from the concept of editorial responsibility in the media sector but applies, in principle, to any person or organization that can demonstrate that it assumes this editorial responsibility.
For most corporate blogs, this means in practice: A blog post created with AI assistance but reviewed and approved by a human before publication does not need to be labeled as AI-generated. Simply skimming the content is not sufficient. The name or title of the person responsible for editorial content should be easy to find, for example, in the legal notice. This editorial exception does not apply to images and videos. In these cases, the only determining factor is whether the content is a deepfake—that is, a deceptively realistic depiction of real people, places, objects, or events. Only then is visible labeling required.
The Code of Practice: What the Labeling Should Look Like from a Technical Perspective
On June 10, 2026, a so-called Code of Practice was published that translates the rather abstract legal requirements of Article 50 into a concrete technical procedure. The Code of Practice stipulates at least two layers of marking for images, audio, and video: digitally signed metadata within the file itself and an invisible watermark that is intended to remain detectable even after editing. For plain text only, the watermark is sufficient. Fingerprinting or logging by the provider is provided for as an optional third layer, but is not sufficient on its own.
The reason for this multi-layered protection is simple: every single method has a weakness. Metadata can easily be lost when a file is saved or converted. An invisible watermark can be damaged by heavy image editing. Simply logging data with the provider is of little help if no one from outside can access it. Only the combination of several methods makes the marking reasonably robust.
For you as a company, this means one thing above all else: With many AI tools, part of this technical labeling already runs automatically in the background because the providers themselves are required to include machine-readable markings. The code of conduct is voluntary. Those who do not sign it must demonstrate in some other way that their own solution is just as effective. For operators, it also includes guidelines on how and where notices should appear, such as with the EU symbol for AI content. The actual task on your end is rarely to apply technical watermarks yourself, but rather to add visible, human-readable notices where it makes sense and is required by law.
What this means specifically for the content on your website
When applied to the day-to-day operations of a corporate website, the situation can be divided into three categories. First: AI-generated blog posts that were reviewed by a human before publication. In most cases, the editorial exception applies here; explicit labeling is generally not legally required, but may still be advisable for transparency reasons.
Second: AI-generated images, such as blog illustrations or background graphics. A visible disclosure is required by law only if the image appears to be a deepfake—that is, if it depicts real people, places, or events in a deceptively realistic manner. There is no labeling requirement for illustrations that are clearly artificial; however, a brief note in the caption is still good practice. Third: Chatbots or AI-powered customer service assistants on your own website. Here, the disclosure requirement is clear: a clearly visible notice at the start of the conversation is mandatory, regardless of how good or human-like the chatbot appears.
What is often overlooked in this list: Even minor linguistic revisions to a text written by a human using AI are generally not subject to the labeling requirement. AI translations are exempt in any case if a human has reviewed the translation. The line is drawn, therefore, not so much by whether an AI tool was involved in the process at all, but rather by who bears responsibility for the content of the result.
At first glance, this distinction may seem somewhat subtle, but in practice it is actually the key factor in determining the legal classification. A text that is published directly and entirely by AI without any human review differs fundamentally, from a legal standpoint, from a text that is created with AI support but for which a human ultimately takes responsibility. Those who clearly document this distinction in their own editorial process will be in a much stronger position in case of doubt.
A real-world example
A client in the education sector had already been using AI tools in several areas of their daily work: for initial drafts of blog posts, for illustrations on topic pages, and for a simple chatbot to schedule appointments. After the announcement of the labeling requirement, there was a brief period of genuine uncertainty within the team as to whether all existing content would now have to be revised.
After a structured review, a much more reassuring picture emerged. The blog posts had already been editorially reviewed and approved by the marketing management prior to publication, so the exception applied. For the illustrations, we added a small, unobtrusive note in the caption area. In the end, the only real need for action concerned the chatbot, which had previously given no indication of its AI nature. A single, clearly worded introductory sentence at the beginning of each conversation resolved the issue within a few minutes. What had initially been a major concern turned out to be a manageable afternoon’s work, involving three small, specific adjustments and a clear, documented chain of responsibility for future content.
Consequences of Violations
Violations of the transparency obligations under Article 50 are penalized under the general fine system of the AI Regulation. Depending on the severity and nature of the violation, fines of up to 15 million euros or three percent of global annual revenue may be imposed, whichever amount is higher. For small and medium-sized enterprises, startups, and—since the AI Omnibus (Regulation (EU) 2026/1744)—also for small mid-caps, however, the lower of the two amounts applies in each case. By comparison, for the most serious violations of the AI Regulation as a whole—such as prohibited AI practices—the penalty range is even higher.
For the vast majority of companies that are making a sincere effort to comply with the requirements, such maximum penalties are a theoretical scenario. We assume that, in practice, the initial response will involve warnings and deadlines for corrective action, rather than immediate maximum penalties for a single missed label. However, this does not relieve companies of their obligation to address the issue. In Germany, as of July 29, 2026, the Federal Network Agency is generally responsible as the market surveillance and complaints authority, and it is only just beginning to establish its inspection practices.
Why the Four-Month Deadline Is Often Misunderstood
One detail consistently causes confusion in discussions: the separate transition period lasting until December 2, 2026. Many interpret this to mean that the entire labeling requirement does not take effect until December, but that is not entirely correct. These four months apply exclusively to providers of AI systems that were already on the market before August 2, 2026, and specifically concern the technical implementation of automated content labeling at the system level.
For you as users of such systems, this deadline does not change your own obligations as operators—namely, to clearly label deepfakes and certain texts of public interest. So anyone who expects new obligations to suddenly arise in December is misunderstanding the actual logic behind the system. The deadline primarily applies to the providers of the tools themselves, not to the companies that use these tools in their day-to-day operations. So anyone who is already publishing AI-generated content should not rely on this deadline as a convenient excuse to delay fulfilling their own responsibilities.
The Difference Between Legal Obligation and Best Practice
One point worth clearly distinguishing: what is required by law, and what may still be advisable purely for the sake of trustworthiness. Even in cases where the editorial exception applies and disclosure is not legally required, transparent, open communication about the use of AI tools can build more trust in the long run than strictly exploiting every legal loophole.
We are now observing a clear trend in this direction across several client projects: Companies that openly communicate where AI is used to provide support—and where, ultimately, a human always makes the final decision—often come across as more credible to their target audience than companies that remain completely silent on this topic. This openness cannot be enforced by law, but it often pays off in practice—especially at a time when many people are already becoming increasingly sensitive to the issue of AI-generated content.
What you should specifically review over the next few weeks
Instead of resorting to knee-jerk reactions, it’s worth taking a structured, calm look at your own website. First, make a list of all content where AI tools played a role in the creation process—from blog posts and images to automated response systems. This inventory alone often provides surprising clarity about the actual scope of AI use within your company—which is usually much smaller than you might have previously assumed.
Next, for each of these content types, it is worth asking whether a human editorial review actually takes place and whether this responsibility is clearly assigned to a specific person or department within the company. This last point in particular is often overlooked: It is not enough that someone, at some point, glanced over a text. Responsibility should be clearly identifiable—and, if in doubt, documented—so that, in the event of a dispute, it is clear who actually performed this editorial role.
Conclusion: No reason to panic, but a good time to take stock
The requirement to label AI-generated content is real and has been in effect since August 2, 2026, but it affects most companies far less severely than the initial headlines might suggest. Those who editorially review AI-generated text, discreetly label AI-generated images, and clearly identify chatbots as such can meet the key requirements with a manageable amount of effort. A calm assessment of the current situation is generally more helpful than frantic, knee-jerk reactions—especially since the authorities’ enforcement practices will only begin to take shape in the coming months.
Related: AI Agents for Small and Medium-Sized Businesses in 2026: What Runs Autonomously, What Doesn’t, and What Costs Ten Slots a Month, and How to Show Up in ChatGPT and Perplexity Responses · Services: AI Consulting and Implementation
Not sure where your website stands when it comes to AI labeling?
We’ll take an honest look at your content and let you know where you actually need to take action and where you’re already on the safe side. Send us a quick message using the contact form. We’ll respond within 24 hours—personally and without any sales pitch.
Sources
Statutes, government and court publications, reviewed on November 4, 2026.
Article 113: Entry into Force and Effective Date – AI Act Service Desk
§ 2 KI-MIG: Market Surveillance Authority – gesetze-im-internet.de
European Commission: Code of Practice on Marking and Labeling AI-Generated Content
Stephenson Harwood: European Commission Adopts Guidelines on Article 50 Transparency Obligations
More Articles
© Marschfahrt Studio
Practical knowledge on web design, SEO, AI, and conversion optimization. Based on real projects, without any marketing spin.
