John Moltz, in his weekly members-only column for Six Colors (no gift links, but Moltz’s column is worth the subscription in and of itself): This week
[Anthropic announced]that in order to identify AI from human-generated content it would be watermarking all text it generates by subtly altering the randomness at which its models picked the next word. While this might seem a perfectly cromulent means of accomplishing a worthy goal, it really rubbed John Gruber the wrong way, as if someone suggested putting autocomplete Swiss on his Philly cheesesteak AI.
Cooper Sharp cheese is the way to go. The best new cheesesteak joint in town — Uncle Gus’s — only offers Cooper Sharp, in fact.
Surely a fair number of the hundreds and hundreds of bees that have nested in Gruber’s extremely fancy bonnet are a result of the fact that this all ostensibly in response to a mandate from the EU. But, as James Thomson has suggested, Anthropic might actually
wantto watermark so that it doesn’t scoop up is own output to train its models, thereby degrading the results. In fact, the company would probably love to be able to figure out other companies’ watermarks and have them not be able to figure out its watermark.
For the record, my opposition to Claude’s plan to begin watermark-adulterating all generated text, worldwide, has nothing to do with it ostensibly being to comply with a regulation from the EU. I do think the EU regulation is deeply ill-considered (and, having been drafted in 2024, already technically out of date), but I’d have written everything I wrote (and will continue to write) about the matter even if the EU had no AI-related regulations in place. In fact, while writing my initial piece, I really wished this had nothing to do with the EU, because pointing out the folly of the regulation (and the ridiculousness of company supposedly worth $2 trillion being technically unable to comply with an EU regulation only within the EU itself) seemed a distraction from my point, regarding the ramifications of watermarking text by way of detectable word choices. Moltz’s aside here confirms that. Also, more and more, I do think this has little to do with the EU regulation and everything to do with Anthropic and Google wanting to watermark the text that Claude and Gemini generate for their own purposes, to identify text to avoid training future models on to avoid model collapse/AI inbreeding. I suspect blaming the decision on the EU “Code of Practice” regulation is just scapegoating on Anthropic’s part.
(As a bonus, Moltz has a deliciously sharp (no cheese pun intended) take on the schadenfreude-rich saga of MacStories returning to Twitter/X.)