AI Watermarking: What Industrial Executives Need to Know About the EU AI Act
A customer sends back your most recent application note together with a single comment: “Was this written by a person?” Since August 2, 2026, there has been both a regulatory and a commercial answer to that question.
AI watermarking refers to a machine-readable sign which AI companies currently place in the outputs of their tools, as this is mandated by Article 50 of the EU AI Act in Europe. The firm that developed the tool is the one that puts the watermark in. However, the company that publishes the content remains answerable for anything that is published. For an industrial business, the key issues are what information the watermark records, what the law demands, and where the actual risk lies.
Executive Takeaways
- The fact that the mark is there shows that a model came into contact with the content; it doesn’t indicate how much of the content the model handled, who was in charge of it, or whether the outcome is satisfactory.
- If your company publishes AI-generated content in the EU, the EU AI Act gives it two obligations: to disclose any AI-generated images, audio, and video that are realistic, and also to disclose AI-produced text which relates to issues of public interest, unless a specific individual has reviewed it and has editorial responsibility for it.
- Standard commercial material is mostly exempt from the text rule; the fine is the smaller exposure, while the greater one is if a customer or buyer finds nothing in your material that only your company could have written.
- Google and the AI answer services do not penalize a mark. Research carried out by Graphite indicates that human-written articles still account for 86 percent of Google rankings and 82 percent of citations from ChatGPT and Perplexity.
- The three questions will show you your position: what the AI produced, who reviewed it by name, and what parts of it came from your own plant.
What constitutes an AI watermark?
An AI watermark is a signal that is embedded in content as it is generated by an AI. The signal is invisible to a reader but can be detected by suitable software using the appropriate key. It shows that the content has been touched by a model, but it does not indicate the extent of the model’s involvement, who directed the model, or whether the final result is good.
Anthropic has explained the way it works in the case of text: if the model has several equally good options for the next word, the selection is guided by a hidden pattern rather than being based on chance. To a reader the prose appears completely normal; to a detector it shows a signature. Light editing is unlikely to eliminate this mark and only a full rewrite will do so. According to Anthropic’s own explanation, there are clear limits to what the method can achieve: it can only indicate that Claude was probably involved at some point, and it cannot tell the difference between “Claude wrote this” and “Claude heavily edited this”. As for Google’s SynthID, by May 2025 it had already watermarked more than 10 billion images, videos, audio clips and text passages.
The same term is used to describe three different technologies, and it is best for executives that they remain separate.
| What it is | Where it lives | What it proves | |
|---|---|---|---|
| Watermark | A signal woven into the content as the model generates it | Inside the file | A model touched this content |
| Content credentials (C2PA) | A signed record of which tool made or processed the file, and when | Attached to the file as metadata | How the file was made, if the metadata survives |
| AI detection | A third-party tool guessing from writing style | Outside the content | Nothing certain. It is a probability, with false positives and false negatives |
The first two are evidence while the third one is an opinion.
Why was it made necessary to have AI watermarking?
Since people can no longer tell whether a piece of content has been created by a human, the regulators have decided that it is the companies which build the models who must bear the burden of proof. Starting on August 2, 2026, Article 50 of the EU AI Act mandates that providers of generative AI shall include a machine-readable indication in synthetic text, images, audio, and video and must make it possible to detect such content.
On June 10, 2026 the European Commission issued its Code of Practice concerning the marking and labeling of AI-generated content. It requested that providers include at least two layers, usually consisting of signed metadata together with an imperceptible watermark, and that they provide free detection tools. Anthropic, Google, OpenAI, Microsoft, Meta and Mistral agreed to the code; on July 31 the Commission had listed around 190 initial signatories. Products that were already available on the market before August 2 enjoy a grace period until December 2, 2026. The Digital Omnibus, which had postponed the Act’s high-risk rules in July 2026, did not postpone Article 50.
In practice, as of October 2026, Google’s Gemini has included SynthID text watermarks since 2024. Anthropic is applying its text watermark to its new Claude models around the world since it does not have a reliable method of turning the watermark on only for Europe, and says it will add its older models over the coming months, ahead of the December 2 legal deadline for those models. Although OpenAI has promised to add watermarks, ChatGPT output still did not have one at the end of August. A regulation drawn up in Brussels now affects a document typed in an office located in Ohio, although the change is arriving tool by tool rather than all at once.
The United States is taking steps one state at a time. The AI Transparency Act passed in California came into effect on August 2, 2026. This act mandates that any generative AI system with over one million monthly users must provide embedded disclosures and make a free public detection tool available, and applies to image, video, and audio; it does not yet apply to text.
What requirements does the EU AI Act place on a company which uses AI tools?
The Act distinguishes between the providers, who are the ones that develop the models, and the deployers, who are the ones who use them. For example, an industrial company which uses a third-party AI tool to produce marketing material is a deployer. According to Article 50, deployers have two obligations.
In the first place, when you publish images, audio, or video that have been produced by AI or altered in some way and appear realistic, you are required to state that they have been artificially generated; but there is a lighter obligation in the case of work which is obviously artistic, satirical or fictional.
Second, when you publish text produced by AI with the aim of informing the public about issues of public interest, you are required to make this known, provided that the text has not been subjected to human review or editorial control and there is no natural or legal person who holds editorial responsibility. The Commission’s guidelines state that spell-checking and grammar review may not meet the threshold for review, and that giving only a superficial editorial approval does not count.
Routine commercial content is generally not subject to the text rule since that rule is activated by the aim of informing the public about matters of public interest and not merely by the subject matter. Fines for breaches of Article 50 can amount to 15 million euros or 3 percent of the company’s global annual turnover, with regard to proportionality for smaller companies. Ask your legal advisor to look at the details; you ought to be aware that they exist.
Will the use of an AI watermark affect our search rankings?
Google doesn’t penalize content for including a mark, any more than the AI answer services do. What loses out is content that provides nothing since a competitor could just as easily generate it from the same prompt.
The research carried out by Graphite backs this up with figures showing that its 2025 study of 65,000 web articles revealed that AI-produced articles had reached about half of new articles by May 2025. The accompanying study, which looked at the articles that actually rank, showed that humans wrote 86 percent of the articles appearing in Google’s results, and that people wrote 82 percent of the articles cited by ChatGPT and Perplexity. In the case of AI-generated articles that do manage to rank, they tend to rank lower down. About half of the new content has to compete for a small portion of people’s attention.
Let’s look at a real-world example. Your best engineer prepares a technical paper from scratch and then has it checked by an AI grammar checker before it is issued. Since almost all of the words are still the engineer’s own, there isn’t much for a watermark to latch onto, and Anthropic states that such a light editing process might leave no trace at all. In the case of a piece produced from a single prompt, the watermark appears everywhere. The detector can see the difference in degree. Your customer notices the difference in the actual content, and that is what wins the order. The tools have made production cheap, and it is cheap production that has left original thinking as the scarce resource. Our guide on content marketing for manufacturers tells what engineers actually read.
Where does an industrial company get real exposure?
The smaller exposure consists of the fine; the larger one involves a customer, a distributor, or an acquirer who reads your technical material and fails to find anything in it that could only have been written by your company.
Trust in a spec sheet, a datasheet, or an application note is the basis of industrial buying. The fastest publisher never received that trust; it was given to the supplier who clearly knew the product. Watermarking does not alter this rule; it only makes it easier and more tempting to break it.
The regulation also highlights an operational gap. With regard to the exemption allowing human review of text, it is assumed that someone at your company checks the content before it is published and takes editorial responsibility by name. However, many industrial companies cannot state who that person is. It is worth addressing this issue even if a regulator never raises it. The same ownership gap determines whether or not your company shows up when a buyer asks an AI assistant for suppliers; for more on AI search visibility for manufacturers, see our guide (AI search visibility).
What ought an executive to ask the marketing team?
Three questions will show you where your company is at, requiring three answers rather than the buying of new software.
For our published material, which items were drafted by AI and who reviewed them before they were published, including content from agencies and freelancers? Who, by name, has editorial responsibility for the material we publish? And with regard to this article, what portion came from our own resources, our own data, and our own engineers and could not have been produced by a competitor using the same prompt?
A rule worth taking up together with those others is this: if AI reduces a three-hour task to thirty minutes, then keep the three hours and use the time saved to increase depth. Marketing should be evaluated in terms of pipeline and performance, not on the number of pieces sent out.
The Duplia Perspective
There is the opinion that watermarking is solely a responsibility of the legal team and that marketing should therefore continue to produce at the speed that AI permits; on the other hand, there is the view that any use of AI constitutes a liability and that the safest course of action is to cease using these tools.
Duplia is positioned between the two of them. The mark is neutral and the tools are useful. The regulation brings to light a management issue that industrial companies have been able to avoid: namely, who is responsible for what the company publishes and what part of it is actually ours. A company that can answer both of these questions is in a good position with the regulator and, more importantly, with its customers; whereas one that cannot answer them has a gap which no detection tool can remedy.
Duplia Marketing is a fractional CMO service built exclusively for industrial B2B and manufacturing companies. We collaborate with leadership teams to establish marketing operations that are able to deal with those questions as and when they arise, incorporating AI into the workflow and having a named person in charge of it. The position of that named person within an agency or an in-house team is discussed in our comparison of a fractional CMO with an agency and an in-house team.
FAQ
Frequently Asked Questions About AI Watermarking
Would an AI watermark indicate that the AI wrote the content?
No. It shows only that a model was probably involved at some stage in the process, whether in producing the outline, the draft, the editing, or the translation. It has no way of telling whether a piece was entirely generated or was written by a person and then heavily edited by a model.
If there's no watermark, then is the content written by a human?
Not necessarily. The content could be from before watermarking was applied, it might have been extensively rewritten, it could be too short to check, or the metadata might have been removed. The fact that you do find a mark gives you a small but genuine piece of information; the fact that you don’t find one tells you nothing at all.
Does Article 50 apply to a manufacturer in the United States?
The rules concern content that is deployed in the EU; in reality, a number of the larger model companies currently add watermarks to their output around the world, and California’s AI Transparency Act also requires similar disclosure obligations with respect to the tools. Consider the EU regulations to be the developing benchmark and consult with legal advice regarding your particular level of exposure.
Must we put a label on each blog post that has been assisted by AI?
The obligation to disclose the text under Article 50 covers content which is published in order to inform the public about matters of public interest, but this obligation disappears when a person has carried out a substantial review of the content and when someone has editorial responsibility. Content that is merely routine and commercial is mostly outside the scope of that rule.
The paper was written by our engineer and the only part that was dealt with by AI was the grammar. Is there a watermark?
Probably not detectably. The watermark lives only in the words the model chose, and a grammar pass changes very few of them. The more the model writes, the stronger the mark.
Which AI tools currently watermark their output?
As of October 2026: Google’s SynthID covers Gemini text and its image, audio, and video models. Anthropic marks text from new Claude models worldwide and attaches C2PA content credentials to generated files, with older models to follow over the coming months. OpenAI has committed to marking text but had not shipped it for ChatGPT by late August. Check the help center of any tool your company uses for “watermark” or “content credentials.”
If you’d like to find out where your company’s content is positioned, you can have a 30-minute structured conversation using the Industrial Marketing Diagnostic.
For more on industrial marketing leadership and AI search, subscribe to The Duplia Perspective.
Patricia Gunter is the founder of Duplia Marketing and a fractional CMO with 25 years of marketing leadership inside industrial B2B companies and manufacturers. Connect with Patricia on LinkedIn.
The Duplia Perspective is published by Duplia, a fractional CMO and executive marketing leadership partner for industrial B2B organizations. Each edition presents two perspectives on a real industrial marketing challenge before arriving at a synthesis. Because growth happens when marketing and strategy work as one.