GoKawiilNature correspondence warns AI text watermarks are vulnerable to cheap attacks
Responding to a Nature news story on Anthropic's invisible watermarking of AI-generated text, researchers point out that such watermarks can be undermined. They cite a 2024 study in which attackers inferred a watermark's pattern by querying the model, then spoofed or removed it for under $50, succeeding roughly 80% of the time.
GoKawiil's interpretation of the reporting above, not reported fact.
The correspondence suggests that watermarking schemes, while useful as transparency signals, cannot be treated as reliable proof of AI misuse if they can be cheaply defeated. The authors argue watermarking efforts need to be open to independent public testing so vulnerabilities are identified and addressed rather than assumed away.
- Anthropic's invisible watermarks were previously covered as a transparency tool for AI-generated text
- A 2024 study found attackers could spoof or remove earlier watermarks for under $50 with about 80% success
- The authors call for AI watermarking efforts to allow independent public scrutiny to expose such weaknesses
Source: nature.com — Hu, 2026-10-27
Published there as: “AI watermarking efforts must encourage independent public scrutiny”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.