Claude gets hidden watermarks. Invisible in prose, but the model steps back for source code

Proof of authorship without visual clutter

Anthropic is tackling the challenge of making LLM outputs auditable. This move responds to the requirements of the European AI Act, specifically the Transparency Code. The model will implement the SynthID Text system, originally developed by Google DeepMind. It operates by subtly adjusting the probability distribution when selecting words in low-stakes situations (such as choosing between "overcast" and "grey"). This creates a statistical fingerprint that is invisible to humans but detectable by a specific algorithm.

Code requires strict syntax, not creativity

The real issue emerges when Claude generates software. Unlike prose, where synonyms can be swapped without altering meaning, source code demands precise syntax. If the system attempts to force statistical noise into the structure, there is a risk that the code won't compile. To preserve the model's programming capabilities, Anthropic has to almost entirely exclude the watermark from the code. It is designed to function only in areas with arbitrary choices, typically within code comments. The generated functional code itself will inherently lack this protection.

Fragility against human edits and rewrites

The system faces clear functional limits once the output reaches human editors. Anthropic acknowledges that while light editing probably won't remove the watermark entirely, a complete rewrite replacing every word will destroy it. Furthermore, if a user decides to use Claude merely for proofreading human-written text, the system will have nothing to latch onto, as the vast majority of the words do not originate from the model.

Production environments will test its durability

The success of the entire watermarking initiative doesn't rely on laboratory demonstrations, but on whether the system can withstand the pressure of daily use in noisy environments. The deciding factor will be whether companies can reliably identify their outputs even after the text has been processed by other systems, and crucially, whether the fears of performance degradation in complex coding tasks, which is the main reason users pay for the model, turn out to be unfounded.

Lilith's verdict

The regulator got what it wanted on paper, but market reality wins. As long as the priority is to maintain perfect compilation and clean code logic, the transparent stamp has to step aside.