AI Watermarks: What This Really Means for Businesses

Are companies required to label AI-generated text?

In most cases, no. The EU AI Act requires that AI-generated texts be labeled if they inform the public about matters of public interest and have not been reviewed by a human. This generally does not apply to offers, emails, or marketing texts. Errors in AI-generated texts are still your responsibility.

Key points at a glance

• AI watermarks only indicate that AI was involved. They don't say who or to what extent.
• There is no requirement to label most company content. However, any errors remain the company's responsibility.
• No one reads generic AI-generated text. But people do read your own experiences and opinions.

When Anthropic announced in August that Claude would be adding a watermark to its text going forward, my first thought as a student was: “Well, that’s it—back to ChatGPT.” But that didn’t happen, because OpenAI, the provider of ChatGPT, and Google, with Gemini, are both bound by the same EU AI Act. 

OpenAI has since followed suit and is now using watermarks in the EU as well. So switching platforms wouldn't have made much of a difference. Instead, I started looking into what this watermark actually means and to what extent it should concern me as a student.

As a student , I'm also interested in another perspective on this —namely, that of companies: Can a customer tell that an offer, an email, or a marketing text was created using AI? When does this have to be disclosed?

What exactly is an AI watermark?

First of all: It’s not what most people think. There are no hidden characters or markings in the text. The watermark works through word choice. Claude often has several words to choose from that all fit equally well. Usually, chance decides which one to use. With a watermark, a secret key makes the decision. Over the course of a longer text, this creates a pattern. Only someone who has the key can verify it.

Watermarks should not be confused with well-known AI detectors such as GPTZero or Turnitin. These tools do not use a key; instead, they use writing style to determine whether a text was generated by AI. Consequently, they are often wrong—even when it comes to texts written by humans.

For this reason, several universities no longer accept a detector score as sole proof. A watermark is more reliable when the text remains unchanged, since a specific pattern is checked. It proves involvement, but not authorship. 

How reliable is AI watermark detection? 

It only works one way. A watermark indicates that Claude was likely involved. However, the absence of a watermark does not automatically mean that the text was written by a human. It may have been generated by another model or edited so extensively that no trace of the pattern remains. 

A test conducted by OpenAI using its own watermark, textGrain, demonstrates just how fragile such patterns are. For a 400-token text , the detection rate dropped from about 92% to 66% when 10% of the words were replaced with synonyms. When 25% of the words were replaced, the detection rate fell to just 17%.

These are test results and do not represent a general detection rate. Editing, translation, short passages, and heavily restricted versions can weaken the signal. However, even a detected watermark reveals very little. It does not indicate who created the text, nor how much of it was written or edited by a human.

What the Law Requires: The EU AI Act and Switzerland

For businesses, another key question arises: Do I have to label AI-generated content? The labeling requirement applies to only a small portion of a company’s content. As of August 2, 2026 , the transparency requirements under Article 50 of the EU AI Act will take effect . 

The AI Act also applies to Swiss companies as soon as their content is used in the EU. The details depend on the specific case. I am a student, not a lawyer. This post does not constitute legal advice.

As a general rule, there is no requirement to label offers, emails, product descriptions, and marketing materials, as long as they are not intended to inform the public about matters of public interest.

A disclosure is required for texts that inform the public about topics of public interest, such as news reports. However, this requirement does not apply if a person has reviewed the text and a natural or legal person assumes editorial responsibility. This could also be a company. “Reviewed” means actually checking the content. Running a spell-check on the text is not sufficient.

What effect do I want my content to have?

There’s no need to debate that AI can be a huge help. It helps structure, organize, and draft texts in no time at all. However, the interchangeability of content remains a problem. Companies that create their content entirely with AI are likely using similar prompts and sources. While the article sounds accurate, it doesn’t offer me anything I haven’t already read elsewhere.

The question isn't whether anyone notices that AI was involved. The question is whether the post has any impact at all. If I want to know today how AI watermarks work, I'll ask my own AI and get an answer to my exact question in ten seconds. Before publishing, you should ask yourself critically what effect you want to achieve with the content. 

The Underestimated Risk: Errors and Copyright

In my opinion, the greater risk lies elsewhere. If AI makes a mistake, it ultimately comes down to your own fault. That’s exactly what happened to Deloitte in Australia in 2025. Deloitte is one of the world’s largest consulting and auditing firms. The Australian government had commissioned a report from Deloitte for 440,000 Australian dollars. A researcher at the University of Sydney found about 20 errors in the report that strongly resembled AI-generated mistakes. These included a fabricated quote from a court ruling and references to studies that did not exist. As a result, Deloitte had to refund part of its fee. In the corrected version, a note was added stating that AI had been used in the report’s preparation. For a firm whose clients ultimately pay for expertise, this is bad publicity.

Then there’s the issue of copyright. In Switzerland, only human creations are protected. So a text written by AI with little human input may not belong to anyone. Essentially, competitors could copy that text quite easily. The more original effort, perspective, and ideas a text contains, the more likely it is to be protected.

How can you prove that a text was written by a human being?

Technically, not at all. And many readers usually don't even notice it. But in the end, that's not the point. What makes an article worth reading is what the AI doesn't know on its own.

This is a personal opinion based on experiences from specific projects, mistakes that have been made, and what has been learned from them. Google, by the way, takes a similar view. Pages that are created in bulk without any original effort or added value for website visitors violate the guidelines. Whether AI was involved is not the deciding factor here.

Conclusion

AI is a useful tool. I translated the German version of this post using Claude. So the translation is probably full of watermarks.

The opinion expressed there is still my own, though. And that's the point: AI is fine as a tool. Your own experience and the process of verifying content cannot be delegated to a prompt.