Critics say Claude's watermark trades quality for compliance

A week after Anthropic began embedding invisible watermarks in Claude's text output, critics are questioning whether the technique's tradeoffs are worth what it actually accomplishes, arguing the fix for one problem quietly introduces a smaller one of its own, The Decoder reported on August 17.

The watermark itself, which Anthropic began rolling out earlier this month, works by tweaking the randomness Claude uses when picking between roughly equivalent word choices, built on SynthID-Text, a technique Google DeepMind published in Nature in 2024. The system uses a secret key plus the preceding words to steer which synonym gets picked, embedding a statistically detectable pattern without inserting any visible marks or hidden characters into the text.

The core criticism, raised most pointedly by blogger John Gruber, is that synonyms aren't actually interchangeable: choosing between "overcast" and "grey" based on a hidden watermark key rather than semantic precision changes what the sentence communicates, however slightly. Gruber argues Anthropic's claim that the technique is imperceptible to readers doesn't hold up. Separately, tools like Declaude can already strip the watermark through paraphrasing, meaning the technique offers little resistance against anyone specifically trying to defeat it. Critics also point to a subtler legal wrinkle: watermarks persist when Claude-generated text gets reused inside later documents, and outputs from multiple watermarked AI systems could end up overlapping within a single piece of text. None of these are hypothetical edge cases; they follow directly from how the technique works rather than from any implementation flaw Anthropic could simply patch away.

“Anthropic's claim that the differences are imperceptible is simply wrong.”

— John Gruber, blogger, Daring Fireball

• Claude's watermark, based on Google DeepMind's SynthID-Text, biases word choice among near-equivalent synonyms
• Applies to all Claude models released on or after August 2, 2026; older models get it over the coming months
• Critics say biasing synonym choice for a hidden signal subtly changes meaning, contrary to Anthropic's 'imperceptible' framing
• Paraphrasing tools like Declaude can already strip the watermark
• Persistence and overlap across reused or multi-model text raise separate legal and provenance questions

Anthropic itself doesn't claim the watermark is airtight: the company has acknowledged it can't distinguish text Claude wrote from text Claude only lightly edited, that the signal gets sparse in short passages or factual writing with few word-choice options, and that a thorough rewrite can strip it entirely. That's a narrower claim than "detects AI content," and the gap between the two is exactly where this week's criticism is landing. The watermark was built to satisfy the EU AI Act's Article 50 transparency requirement, not to win an arms race against people actively trying to hide AI-generated text, and the tradeoff critics are describing is the cost of choosing regulatory compliance as the design target instead. That framing doesn't make the criticism wrong, but it does explain why Anthropic shipped a technique with acknowledged blind spots rather than waiting for something closer to bulletproof: the EU's Article 50 deadline didn't leave room for that. Whether readers ever notice the difference in practice, versus critics who are specifically looking for it, remains the open question underneath all of this.

Nothing here should be taken as financial advice — just information to consider.