In an article regarding Europe’s censors’ efforts to govern and limit American speech, Nathan Harden, RealClearEducation Editor, warned about AI packages embedding invisible and supposedly unremovable watermarks in writing (and eventually in imagery and recordings of speech, also) so that those censors can identify and block speech of which they personally disapprove. In the course of that, the author wrote this about Anthropic AI-generated watermarks in particular.
The marking technology subtly biases words the model picks, in a pattern Anthropic controls via a secret key. The watermark survives ordinary copy and paste, and there is no way to opt out. Anthropic says the mark carries no information that could identify a user and won’t affect the overall meaning of the text.
It isn’t possible to subtly bias words in a text without the overall meaning of that text being so influenced. Surely Anthropic’s managers understand this.
Harden then asked the critical question:
Americans should be asking a more fundamental question: Why is European law dictating how an American company produces words for Americans?
The answer is because [Anthropic’s] managers are all too willing to surrender to Europe’s censors in order preserve their precious presence there and not willing enough, if at all, to spend that energy preserving the freedoms of the nation that nurtured them and provided them the environment in which they grew and now prosper.
How, then, can [Anthropic] or its products be trusted in America or anywhere else?