Anthropic's plan to watermark Claude's output has a side effect: it can make models more likely to break their own safety rules.
Anthropic recently disclosed that future Claude models will use SynthID-Text, an open-source watermarking method built by Google, to comply with a new EU disclosure law. The system uses a secret key to nudge word choices, swapping "cloudy" for "overcast," for example, so anyone holding the key can later verify a piece of text came from that model. New research shows the technique does more than tweak vocabulary. It can also change which tools a model calls and how consistently it follows its safety training, with the effect growing sharper under adversarial prompts designed to trick a model into doing something it normally wouldn't, like revealing a password.
That's a bigger deal than it sounds. Watermarking is usually pitched as invisible bookkeeping, a way to prove text is AI-generated without changing what the text says. Lasso Security researcher Andrea Siposova said that altering how a model generates text "is definitely going to change their behavior, especially when we place it under adversarial conditions," and that such tradeoffs "show up somewhere" even when the watermark itself isn't perceptible to a reader.
The irony: a compliance feature built to satisfy one regulation may be quietly undermining the safety testing meant to satisfy another.