Drycot

After AI tools have turned out to be sophisticated, an AI-generated text was hardly differentiable from real human writing. However, soon it is going to change. Major AI companies are now adding hidden signals, called watermarks, to the text generated by their models. This shift did not happen by choice alone as a new European law is requiring AI companies to make their content traceable. You might think you are safe if you do not reside in the EU, however, the change is not staying inside Europe. This article explains why this is happening, how text watermarking works, and what it means for people who use AI tools every day.

What Sparked the Decision: The EU AI Act

The European Union AI Act is the world’s first broad legal framework to oversee artificial intelligence. It became law in 2024, but many of its rules are only now taking effect. One part of the law, known as Article 50, requires AI companies to make their generated content detectable. This means images, audio, video, and text created by AI must carry some kind of marker that shows it was produced by a specific AI tool.

The transparency rules under Article 50 were enforced on August 2, 2026. That means any AI model that will be released after this date will have to watermark its generated contents. AI models that were released before this date have a period of transition to implement the watermarking. 

Not paying attention to these verdicts is severe. Companies that break Article 50 rules can face fines of up to 15 million euros, or 3% of their total worldwide yearly income, whichever amount is higher, which is a strong reason why large AI companies are taking the rule seriously instead of ignoring it.

Why the Rules Apply Globally, Not Just in the EU

You might expect this rule will have an effect only on the residents of the EU, but not really. The AI giant Anthropic, the owner of Claude product, has already announced their implementation of watermarking will be world-wide. There are rumors that other AI giants will follow the same rule. One obvious reason might be that companies like OpenAI, Google, and Anthropic build one main version of their AI models and offer it to users around the world. Building a separate, weaker version just for European users would cost more money and be harder to manage. Because of this, many companies are choosing to add watermarking features globally, not only for people inside the EU.

This pattern is not new. A similar thing happened with GDPR, the EU’s data privacy law from 2018. Many companies changed their privacy practices worldwide after GDPR, not just in Europe, because it was simpler to follow one strict standard everywhere. Some experts now expect the AI Act to have a similar global reach. This effect is sometimes called the “Brussels effect,” named after the city where EU laws are made.

What “Watermarking” Actually Means for Text

So what are the technical aspects regarding watermarking the text? Watermarks often suggest a visible logo stamped on a photo. Yet text watermarking is very different. It does not add any visible mark to the words you read, but instead, it chooses the words according to specific patterns, which are almost impossible for a human reader to notice but that a computer program can detect later.

However, this process is much harder to do with text than with images or audio. A photo has millions of pixels, and a sound file has a detailed audio wave. So there is a lot of room to hide a signal without changing how the image looks or the audio sounds. On the other hand, text consists of a much smaller number of words, and even small changes can affect meaning. This makes text watermarking one of the most difficult technical challenges in this entire area of AI regulation while Anthropic this will not affect the quality of generated texts. 

How the Leading Approaches Work

More detailedly, most current text watermarking methods rely on how AI models choose words in writing a text. When an AI model writes a sentence, it does not pick just one random next word. Internally, it evaluates a list of possible words, each with a certain probability of being chosen. Watermarking systems adjust these probabilities in a specific, hidden pattern while for the reader, the text looks all natural. But a detection tool provided by that AI company can analyze the word choices and identify the text as AI-generated with a high level of confidence. Actually, watermarking of texts is not something new. Google, through its Gemini, has been already marking its generated texts with a system called SynthID, which works in the way described above.

The Core Technical Challenge: Why Text Watermarks Can Be Fragile

Though watermarking is a powerful tool, it is not impossible to break. Antrophic, in their announcement following the EU’s decision, underlined that watermarking should not be fully trusted. Since the AI chooses words one by one according to a pattern, paraphrasing the text can remove the watermark. Other ways of doing so are translating the text into another language, or asking a different AI model, which is not obliged by the EU’s directive,  to paraphrase it, all of which can erase watermarks.

This is very different from watermarking in images or audio, where the hidden signal is often built to survive common changes like resizing a photo or compressing an audio file into a smaller format. Text does not have that same flexibility. Because sentences must still make grammatical sense and carry the same meaning, there are fewer places to hide a strong, unbreakable signal. This is one reason some researchers remain doubtful that text watermarking can ever become fully reliable, even as regulators require it.

What This Means for Everyday Users

For most people, these changes will be mostly invisible in daily life. If you are chatting with an AI assistant or reading an article, watermarking will not affect your experience as AI companies claim the quality of the output will not be altered. The change is way more important for people who rely on AI in producing texts to publish online. It will no longer be that easy to ‘act’ as if their post is written by them. Detecting it usually requires a separate tool built specifically to check for the hidden pattern, and these detection tools are going to be available for public use as in Anthropic’s announcement.  

It is also important to understand what watermarking does not do. It is not a guarantee that all AI-generated content will be caught. It is not a ban on using AI to write. And it does not automatically stop someone from removing the watermark through editing. For now, its biggest impact will likely be in professional and academic settings, such as schools checking student work, publishers reviewing submitted articles, or platforms trying to label AI-assisted content for their users.

Criticism and Open Debates

Not everyone agrees that text watermarking is the right solution. Some researchers argue that as long as a determined person can rewrite or translate text, no watermarking method will ever be fully reliable. Others point out a timing problem: regulators are requiring companies to implement watermarking before the technology has fully matured or before clear, shared testing standards exist across the industry. This means companies are being asked to comply with rules while the tools to meet those rules are still being built and tested.

There are also practical concerns about cost. Running a watermarking system across billions of pieces of generated text requires computing resources, and smaller AI companies may find this harder to manage than larger ones with more resources.

What’s Still Unresolved

Several questions remain open as this article is written. The exact legal status of the December 2 grace period is still being finalized in EU processes. There is no single, universal watermarking standard that all companies have agreed to use, which means detection tools built for one company’s system may not work on another’s. Enforcement actions and official guidance are still developing, so companies and users alike should expect more updates in the coming months.

Conclusion

The move toward watermarking AI-generated text did not start as a voluntary choice by tech companies. It began as a legal requirement from European regulators, but because major AI companies build products for a global audience, its effects are reaching users everywhere. The technology behind it is still young, and text remains one of the hardest formats to watermark reliably. For now, most users will not notice any visible difference in their day-to-day use of AI tools. But as detection systems improve and more companies publish details about their approach, watermarking is likely to become a bigger part of how we think about trust and authenticity online. This is a fast-moving topic, and readers who want the latest details should check official sources, since some elements described here are still developing.