AI-Generated Text Will Soon Carry Invisible Watermarks, Thanks to EU Regulations
The European Union’s Code of Practice for AI is taking a significant step forward with the implementation of invisible watermarks in AI-generated text. Anthropic, the company behind the popular AI model Claude, has announced that it will begin applying these watermarks globally to all text generated by its system.
For those who may not be aware, this means that any text created by Claude – whether it’s a short social media post or an in-depth article like this one – will carry a unique identifier. But here’s the key part: this watermark won’t be visible to the average reader. It’s more of a digital fingerprint, designed to help detect if a piece of content was generated by AI.
The watermarks are being implemented as part of Anthropic’s compliance with the EU’s Code of Practice for AI. This code requires companies serving the European market to clearly identify and label their AI-generated content. By applying these watermarks, Anthropic is not only complying with this regulation but also paving the way for more transparent use of AI in content creation.
So how exactly does this work? In a nutshell, Claude’s watermarking system uses a secret key to modify the randomness used when generating text. This modification introduces a subtle statistical pattern that can be detected by someone who has access to the key. It’s not like adding hidden characters or modifying the finished response – instead, it’s more like leaving a faint digital signature on the generated text.
According to Anthropic, its implementation of watermarking is based on Google DeepMind’s SynthID-Text approach. This means that Claude’s AI model will use a different source of randomness when making certain choices during generation. For example, if the model has multiple reasonable options for what to generate next, it might use the secret key and some preceding words as part of its decision-making process.
The good news is that this watermarking system won’t have any practical impact on the quality or content of Claude’s output. In other words, the AI-generated text will still be creative, readable, and free from bias – but now it’ll also carry a unique identifier.
Anthropic has confirmed that internal testing has found no significant differences in creativity or readability between watermarked and non-watermarked text. This suggests that the implementation of watermarking is both effective and subtle – after all, the goal is to make it invisible to human readers while still allowing for detection by those with access to the key.
As we move forward into a world where AI-generated content becomes increasingly prevalent, these watermarks will play an essential role in maintaining transparency and accountability. They’ll help content creators, researchers, and even law enforcement agencies identify whether a piece of text was generated by AI or not – which is crucial for tackling issues like deepfakes and misinformation.
For readers who want to stay ahead of the curve, here’s a key takeaway: be aware that AI-generated content will soon carry invisible watermarks. This doesn’t mean you’ll notice any difference in how it looks or feels – but it does mean that there are new ways to detect whether a piece of text was generated by AI or not.
Source: Bleeping Computer — 2026-08-14