@shaneparrish: “We shape our tools and thereafter our tools shape us" — John Culkin, writing about Marshall McLuhan One problem with A…

“We shape our tools and thereafter our tools shape us" — John Culkin, writing about Marshall McLuhan One problem with AI watermarking on writing is that humans are learning to write like AI as quickly as AI is learning to write like humans. My kids write their school essays by hand in class, without access to any devices. Yet when I read them, they look as though they were written by AI. In fact, one of the reasons they write them by hand is that teachers can no longer tell the difference. How is that possible? It's nearly impossible to prove or disprove to a third party that something was written with AI. Try it for yourself. I once submitted a few pages from an early book to an AI detection engine. With near certainty, it told me that what I wrote was AI. The only problem was that the book was from before AI even existed. Watermarking can make some words and patterns statistically more likely to appear in AI-generated text, possibly allowing a very weak signal on a long enough sample. But as people absorb these patterns, their own patterns shift. If my kids are any indication, it won't be long before they update their own mental prediction algorithm and start watermarking their own words. This was Culkin and McLuhan’s point: the tools we create shape us. Even if we could reliably prove that AI was used (and we can't), that would tell us little about the quality of the output. A carpenter who refuses a hammer doesn't make the house any better. He has only made his work slower and harder. Judge the work, not the tool.