Claude AI acknowledges watermarking limitations

• Anthropic has started embedding invisible watermarks into text generated by its Claude AI models.
• Claude itself expressed skepticism about watermarking's effectiveness, noting it can be removed by editing or paraphrasing.
• The AI suggested that malicious actors are most likely to successfully strip watermarks.
• Anthropic's method involves slightly altering word choices during text generation, adapted from Google DeepMind technology.
• The company plans to allow third parties to verify the presence of these watermarks.
• Human-written text edited by Claude may also carry the watermark, potentially leading to misidentification.

Anthropic has begun embedding invisible watermarks into text generated by its AI chatbot, Claude, a move that aligns with new EU transparency rules. However, Claude itself has expressed skepticism about the effectiveness of such measures, describing them as not a "silver bullet" for identifying machine-generated content.

The chatbot explained that simple watermarks can be easily removed through cropping, screenshots, or editing, while more sophisticated methods can be defeated by paraphrasing or reformatting. Claude specifically noted that individuals intending to misuse AI for fraud or disinformation would be motivated to strip watermarks, for instance, by rephrasing text or translating it back and forth between languages.

Anthropic's watermarking technique, adapted from Google DeepMind, involves subtly altering word choices during text generation. While the company plans to allow third parties to verify these marks, it acknowledges that human-written content edited by Claude could also be flagged, and heavily rewritten AI text might lose its detectable signal. Short passages and specific outputs like code are also more difficult to watermark reliably.

Despite these limitations, Anthropic is implementing the system globally, not just in Europe, and plans to extend it to older Claude models. This initiative follows the EU AI Act's transparency requirements, with OpenAI and Google also expected to introduce similar measures. In contrast, Elon Musk's xAI has not committed to watermarking its chatbot's output, though it must still comply with EU regulations.