In practice, LW's ai policy is more about quality than literally "was ai used". Nobody really cares if you use AI skilfully. They care about weeding out low effort slop posts. Just post what you want to and hold yourself to a high standard of quality, and it'll be fine.
I read the whole policy change as meaning this old policy (filtering for quality, not LLM use) is about to end.
As a new member of LessWrong I was endeavoring, as we all should, to familiarize myself with the rules, principles, and etiquette of the community. While taking a break from the suggested core reading, I thought that a good way to start to pattern what was not acceptable on the site was to look at the moderator comments on denied posts. I was surprised to see so many posts rejected as they were (or suspected to be) LLM generated. While posting entirely LLM created content is understandably forbidden, my mind quickly went to the obvious edge cases and problematic implications of such a rule. Don't get me wrong, at its core, I support both the rule and the premise behind it - that original, genuine human thought, and the expression thereof is valued, and should be elevated above the all-too-easily produced AI elaborations on simple kernels of ideas. And this isn't even getting into the mountains of AI slop posts that could overwhelm the site should the moderators abandon this principle completely. That said, this is a site whose core principle is rationality, so let's apply this principle introspectively and explore the potential tensions this rule manifests.
The first problem with this approach is that the detection tools are far from perfect. We all know the basics at play here; while these tools are doing some heuristic analysis looking for em-dashes, structured lists, etc. they are primarily doing statistical pattern detection to measure how predictable the text is. In other words, how consistent is the sentence length and structure, and how predictable is the next word(s). These are pattern indicators, not evidence of AI writing, and this is the reason that dozens of the top universities in America have disabled, disallowed, or institutionally restricted their use. Most other notable schools which haven't implemented restrictions have cautioned their faculty to not rely on AI detector findings alone when asserting wrongdoing.
To further quantify this point, many AI detection vendors claim < 1% error rate, however, independent studies often find significantly higher false positive rates - upwards of 10%. I can personally (albeit anecdotally) attest to this false positive issue. In testing both QuillBot and GTPZero, I have had three of my papers flagged as AI generated - 55%, 60%, and 70% probabilities - with "Likely AI" and "Mixed" classifications. None were authored or edited by AI. This may seem difficult to believe given the informal style of this letter, but when attempting more serious works I adopt a more disciplined form characterized by a structured, deliberately coherent flow, written in the third-person objective. Combine this with my liberal use of dashes, a habit from as far back as high-school, and these detectors seem to love to flag me. Clearly this in not because I'm some literary savant, but simply because I happen to match certain patterns.
Forgive me if I'm wrong, but what I could infer from the moderator comments I read, there is a relatively hardline approach to denying posts that are even suspected of being AI-generated. I will admit, however, I'm not aware of the exact mechanism or threshold involved in such determinations. Regardless, based on the commentary I read, it was clearly something more than just the cleanup of obvious AI slop posts.
So here's the rub: from what I can see so far, LessWrong values structured, long-form reasoning and analysis from people who can articulate it in balanced and well-formed prose. Ironically, authors who do this naturally are the ones more likely to get caught up in detection sweeps, along with those who actually did use AI of course. What we end up with is an inversion of Blackstone's Ratio where "It is better that ten guilty persons escape than that one innocent suffer". We are willing to toss out some of the best innocents so long as we sweep up the guilty. This seems to me to be something worthy of further reflection.
Again, I'm not against the premise of a clean and quality-oriented site dedicated to rational exploration. The question is in how we achieve it. At the end of the day, I'm perfectly comfortable with the moderators making a judgement call when something is just blatantly AI generated or the detection kicks back a 95% probability. But if the topic and content of the piece is particularly informative or thought provoking, isn't it worth our community seeing, even if possibly AI assisted in its creation? I believe we should be striving for quality control, not a purity test.
Alas, I am but a lowly newcomer here, and I accept that LessWrong is a benevolent dictatorship. Let's just not forget "benevolent" is in the description too.
Sincerely,
Richard