| I don't think the Stephen Hawkings example is compelling. No one argues against AI for transcription or to help paralyzed people communicate. Stephen Hawkings still wrote every word of his book. But it does raise the question: what if he hadn't? What if Stephen Hawkings had secretly used an LLM to write A Brief History of Time? Imagine that we find a letter in which he confesses that he used an LLM to write the book. Would that lessen the value of the book? I don't think so. I think we need to distinguish between two arguments: A. You should avoid AI writing because it is bad.
B. You should avoid AI writing because it isn't human. I find myself in camp A. I think a lot of AI writing is bad, but (1) it gets better when a human spends time working on it, and (2) AI is improving over time and its writing is getting better. For me, an AI writing detector is about curation. It is hard to figure out what's worth reading, so filtering out low-effort AI writing is a good heuristic. But what I really want is a Good-Writing Detector. I don't care how a piece of writing happened--all I care about is if I'm going to value reading it. In contrast, camp B is never going to be happy with AI writing. It doesn't matter how good it gets; even if it is better than human writing, camp B will not approve. I get that, even if I don't agree. There is virtue in sacrificing to support an ideal. |