I think if people were purely using LLMs to convert their own writing into a publishable post, that would mostly be fine, but in practice people seem to be outsourcing the entire writing process (“Write me a post about X”). I predict that if someone actually had an LLM do an hour-long interview about some topic and then write a post about it, or wrote a post-length prompt and then had an LLM turn it into something publishable, the results would actually be fine (likely still running into LLMisms like going on tangents and focusing too much on style, but that’s minor).
It would be interesting to try a policy of “you can submit an AI-written post as long as you attach either a Google Doc with an entirely-human-written prompt longer than the post, or a transcribed interview where the human portion is longer than the post”.
I agree that these excessively-outsourced posts are bad and that the quality of thoughtfully-prompted posts would probably be fine. I suspect that the more tangential considerations are more important, like how intellectual culture would be affected and how much moderators should be regulating the writing process of authors (the “how” rather than the “what”).
I think your “prompt similar in length to post idea” is decent, but I think there are more sophisticated measures we could create that are less easily gamifiable, for instance, time spent on LessWrong as the primary browser, multiple edits on LessWrong across days.
I’d worry about false-positives there, since some of the most prolific posters are copying posts from their external blogs, and I frequently write in Google Docs and then copy and paste.
Fair enough. I think you could probably create a carve-out for cross-posting, as I believe cross-posts are a small minority of posts. For GoogleDrive, I am not sure if this would be feasible, but allowing users to export a version history and having a tool to verify proof of work could be a solution.
I think if people were purely using LLMs to convert their own writing into a publishable post, that would mostly be fine, but in practice people seem to be outsourcing the entire writing process (“Write me a post about X”). I predict that if someone actually had an LLM do an hour-long interview about some topic and then write a post about it, or wrote a post-length prompt and then had an LLM turn it into something publishable, the results would actually be fine (likely still running into LLMisms like going on tangents and focusing too much on style, but that’s minor).
It would be interesting to try a policy of “you can submit an AI-written post as long as you attach either a Google Doc with an entirely-human-written prompt longer than the post, or a transcribed interview where the human portion is longer than the post”.
I agree that these excessively-outsourced posts are bad and that the quality of thoughtfully-prompted posts would probably be fine. I suspect that the more tangential considerations are more important, like how intellectual culture would be affected and how much moderators should be regulating the writing process of authors (the “how” rather than the “what”).
I posted a top-level shortform to test this theory.
I think your “prompt similar in length to post idea” is decent, but I think there are more sophisticated measures we could create that are less easily gamifiable, for instance, time spent on LessWrong as the primary browser, multiple edits on LessWrong across days.
I’d worry about false-positives there, since some of the most prolific posters are copying posts from their external blogs, and I frequently write in Google Docs and then copy and paste.
Fair enough. I think you could probably create a carve-out for cross-posting, as I believe cross-posts are a small minority of posts. For GoogleDrive, I am not sure if this would be feasible, but allowing users to export a version history and having a tool to verify proof of work could be a solution.