AI watermarking efforts must encourage independent public scrutiny

Your News story on Anthropic’s invisible watermarks highlights their potential as transparency signals for text generated with the help of artificial intelligence and their limits as evidence of misconduct (see Nature https://doi.org/rskt; 2026). But such watermarks can become targets for attack, too. In tests of previous efforts, attackers inferred patterns through model queries and then spoofed or removed the watermarks for less than US$50, with a success rate of roughly 80% (N. Jovanović et al. Proc. 41st Int. Conf. Mach. Learn. 235, 22570–22593; 2024).

Nature 658, 1356 (2026)

doi: https://doi.org/10.1038/d41586-026-03406-3

Related Articles

• Can Anthropic’s invisible watermarks curb ‘AI slop’? Researchers remain sceptical

• AI watermarking must be watertight to be effective

• Don’t let watermarks stigmatize AI-generated research content

• See more letters to the editor

Subjects

Latest on:

• Technology
• Machine learning
• Information technology

• 

  Nature’s job-satisfaction survey highlights the changing geography of science

  Editorial 20 OCT 26

• 

  Is AI eroding PhD students’ critical thinking and analysis skills?

  Career News 15 OCT 26

• 

  China’s AI research is booming: what it’s like to be on the front line

  News Feature 14 OCT 26

• AI-hallucinated citations should be treated as research misconduct

  Correspondence 27 OCT 26

• 

  Can maths survive AI? Fields medallist Terence Tao calls for a ‘Maths 2.0’ world

  World View 22 OCT 26

• 

  AI doctors’ orders come with a sky-high bill

  Research Highlight 21 OCT 26

• Decentralized AI needs more evidence on energy use

  Correspondence 13 OCT 26

• 

  Private, offline and yours: how to search the literature with local AI systems

  Technology Feature 12 OCT 26

• 

  Retrofitting language models to operate over bytes

  Article 07 OCT 26