A new law in Europe will soon require AI companies to label content that is not written by a human. To keep up with these rules, the company Anthropic is adding invisible watermarks to text produced by its AI assistant, Claude, so that machines can later detect it as synthetic. WHAT'S HAPPENING: Anthropic is adopting a technology called SynthID-Text, which was originally developed by Google. This system functions like an invisible stamp embedded directly into the words Claude writes. For the human reader, the text looks and reads exactly the same as it did before. The goal is to create a digital fingerprint that only special software can find. This move is a direct response to the European Union's AI Act, which mandates that AI-generated content must be clearly identifiable. While this sounds like a massive overhaul, Anthropic claims that these invisible markers will not change the quality of the writing or cost users any extra money. ## The hidden signature of AI HOW IT WORKS: To understand how these marks work, you have to look at how an AI chooses its next word. When an AI generates a sentence, it is constantly calculating the probability of which word should come next. For example, in the phrase, The weather today was cold and, an AI knows that the word overcast is very likely, while the word sugary is extremely unlikely. Often, the AI has several perfectly reasonable options to choose from, like overcast or grey. In normal operation, the AI picks between these good options using a standard random number generator. The watermarking system changes the source of that randomness. Instead of a purely random choice, the system uses a secret key combined with the preceding words to decide which of the valid, high-probability words to use. Over the course of a long document, the AI makes thousands of these minor choices. By nudging the AI toward specific, subtle patterns, the system leaves a trail that a specialized detector can confirm later. Because the AI is still picking words that make sense, the reader never notices a change in flow or logic. WHY IT MATTERS: This is the first step toward a world where we need software to tell us who—or what—wrote the text in front of us. As these laws take hold in Europe, other companies like Google are already using similar methods, and others will soon follow suit to avoid legal trouble. The long-term challenge is that these marks are not permanent; if someone heavily edits or rephrases the AI-generated text, the pattern may be broken, rendering the watermark useless. For now, it represents a quiet shift in how we might verify the origin of the information we consume online.
Liked this one? The next lands at breakfast.
Every story in tomorrow's AI news, rebuilt in plain English — five minutes, sources linked, free forever.
By joining you agree to receive Article's daily newsletter — unsubscribe in one click. Privacy