Anthropic plans to add an invisible watermark to texts generated by Claude. This is intended to comply with an EU regulation requiring AI-generated content to be identifiable from December onward. Claude’s random word choices would therefore be subtly altered to create a detectable pattern that remains invisible during normal reading. There are concerns that this intervention could affect the quality of the text and lead to less precise wording.
Another view is that users are unlikely to notice any difference, since randomness is already a fundamental part of how language models generate text. Watermarks could also help combat academic dishonesty and prevent “model collapse,” in which AI systems lose quality because they are trained on AI-generated data.
