The news
Anthropic has added watermarking at the model level to text generated by Claude. The change is presented as compliance with an EU regulation on AI outputs. The implementation marks every response rather than limiting the marks to cases the law explicitly requires.
Context
The regulation aims to make AI-generated text detectable. Earlier tools such as calculators and spellcheckers faced no similar demand that their results carry visible traces of use. The law’s text attempts to draw a line that preserves some distinction between routine assistance and full composition. Anthropic’s choice removes that distinction in practice by applying the watermark uniformly.
James Padolsey, writing on Daring Fireball, notes that the same logic applied to calculators would have required every sum to reveal its mechanical origin. The current rule instead places the requirement only once the tool can produce whole sentences. Padolsey calls this placement a moral premium on difficulty rather than a coherent boundary.
Details
The watermarking is described as a blanket, model-level feature. It does not appear limited to particular use cases or output lengths. Padolsey observes that this scope is wider than the regulation’s minimum requirement. The company therefore went beyond what the statute demanded while still framing the step as necessary compliance.
No technical details on the watermark method itself are supplied in the source. The criticism centers on the decision to implement at full model scope instead of narrower application. That decision, the argument runs, discards distinctions the law tried to keep.
The piece compares the policy to earlier moments when new tools changed what counted as acceptable assistance. Spellcheckers and calculators altered expectations without creating legal duties to label their outputs. The EU rule breaks that pattern by making detectability mandatory once the assistance reaches sentence level.
Why it matters
A regulation that treats capable language models as uniquely suspect creates an incentive for companies to over-mark rather than to calibrate. Anthropic’s wider implementation shows how that incentive works in practice: the safer corporate choice is to watermark everything and avoid any later claim of under-compliance. The result is a de facto rule that all Claude text carries a mark, even where the statute left room for narrower application.
This approach shifts the burden onto users and readers to treat marked text as lesser. It does so without evidence that the watermark improves accountability in the specific cases the law originally targeted. Over time the practice risks normalizing the view that any fluent output from a model must be flagged, regardless of how the model was actually used.
The precedent also travels. Other providers facing the same statute will weigh the same choice between minimal and expansive marking. If the broader option becomes standard, the regulation will have produced a uniform label on model output rather than a targeted signal for the uses it set out to address. That outcome follows directly from the decision to place the compliance trigger at the point where assistance becomes capable rather than at the point where it becomes deceptive.
---
Sources:
No comments yet