tech
Why Anthropic's AI watermark for Claude text goes further than rivals
Anthropic's AI watermark, which can appear in translations, summaries, and more extensively edited text, has prompted some backlash.
TL;DR
- Anthropic's new watermark for AI-generated text from Claude may go beyond what is legally required by the EU's AI Act.
- Article 50(2) of the EU law does not mandate marking AI outputs for standard editing or when meaning is not substantially changed.
- Anthropic's watermark can appear in translations, summaries, and more edited text, leading to concerns about misrepresenting AI involvement.
- Anthropic claims the global rollout is due to a lack of per-region application, and that lighter edits trigger the watermark less.
- Some Claude users have canceled subscriptions, and tools to remove the watermark have emerged.
- Experts suggest Anthropic's approach might be a broader interpretation of the law, aligning with its safety-first brand identity and potentially shifting regulatory burden.
- Other tech companies like Google and OpenAI are also implementing or planning AI content detection methods.