The research organization Anthropic recently shared how it embeds invisible tracking markers into text generated by its primary models. The popular assistant Claude now includes a hidden pattern within its written responses to comply with the transparency rules of the European Union AI Act.
The company plans to release a detection tool that will let anyone scan a piece of writing and verify if the system created it. This change impacts all people creating regular content with the platform.
The system changes word choices to create a hidden pattern
The group uses a method based on the SynthID-Text framework built by Google DeepMind. The system works by altering low-stakes word choices while generating artificial intelligence output. For example, the model might select a word like “overcast” instead of “grey” based on a specific statistical pattern.
Don’t miss the best of The Mac Observer
Set us as a preferred source and our Apple reporting ranks higher in your Google Search results and Discover feed — one tap, no account changes.
These small shifts in vocabulary are completely invisible to readers, but they build a signature that a decoding key can identify. The company noted that light editing will not remove this mark since most of the original words stay in place. A complete human rewrite is required to erase the signature.
Programming code remains mostly unaffected due to strict syntax rules
While standard writing carries this new marker, computer code works differently. An AI system must produce working scripts, which leaves very little room to swap words without breaking the logic.
Because programming languages require rigid syntax, the model cannot freely choose between valid alternatives. The marker might only attach to flexible areas like developer comments.
As a result, the watermark has almost no impact on the functionality or detectability of the generated software code. The organization hopes this overall approach will build more trust online.
Discussion