Anthropic、Claude の出力にウォーターマーク導入へ批判も

Anthropic's text watermarking for Claude is supposed to make AI-generated content detectable. But critics doubt that word choice stays unaffected, and lawyers are facing new transparency headaches.

Anthropic insists that text watermarking for Claude has no effect on content, creativity, or readability. The method is based on Google's SynthID-Text approach and tweaks the randomness source for word selection during text generation, creating a statistically detectable pattern. No visible marks are inserted. No hidden characters are added.

Blogger John Gruber disagrees. In a lengthy post on Daring Fireball, he argues that no two synonyms carry the exact same meaning. When Claude picks between "overcast" and "grey" based on a secret watermark key rather than semantic precision, text quality inevitably suffers. The system sometimes boosts the probability of a worse word and lowers it for the best one, Gruber writes. Anthropic's claim that the difference is "imperceptible" is simply wrong, he says.

Gruber has run the influential tech blog Daring Fireball since 2002. He also co-created Markdown, the formatting language now widely used by chatbots.

Even the SynthID study that Google Deepmind published in Nature, which Anthropic cites, doesn't convince Gruber. The thumbs-up/down rates measured there aren't a valid gauge of text quality, he argues. Nobody gives a chatbot a thumbs-down because it wrote "bananas" instead of "pineapples." Gruber suggests that Gemini's reputation as weaker than Claude and ChatGPT could partly stem from SynthID already being active.

Law firms need to prepare for transparency questions

Legal trade publication Artificial Lawyer examined what Anthropic's watermarks mean for law firms. The publication's verdict is mostly relaxed. The markings are harmless, Artificial Lawyer argues, and in most cases neither clients nor courts would object to AI use. Some clients are even actively requesting it now, the analysis notes.

Things get tricky in specific situations, though, according to Artificial Lawyer's assessment. If clients have explicitly banned AI use for their cases, or if a judge is skeptical of it, the AI contribution could be proven even when the submitted document is factually flawless. That could have consequences for how proceedings or a trial play out.

The watermarks also travel with the text, Artificial Lawyer points out. A complex contract might contain clauses that are AI-marked while a human wrote the rest. Future contracts built on older templates with AI-generated sections carry those markings forward. When multiple LLMs are used, different watermarks could even overlap in a single document.

Fee negotiations could shift, too. If clients demand discounts because AI made the work easier, the AI share becomes verifiable in principle, the publication notes.

Anthropic itself says that watermarking is sparser in fact-heavy passages because fewer word alternatives exist. For legal texts that depend on precision, that's a relevant caveat, even though there aren't any empirical studies on the topic yet.

Anyone who wants to strip the markings can do so with paraphrasing tools like Declaude. Its developer James Padolsey criticizes the underlying EU regulation as arbitrary. The system mostly hits ordinary users while doing little against deliberate circumvention, he says.

Anthropic is rolling out watermarking to comply with the EU AI Act but applies it globally because it can't limit the feature by region. All Claude models released after August 2 support the marking. Older models will be retrofitted in the coming months.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now