Oliver Kuperman's Shortform

Oliver Kuperman

This is a special post for quick takes by Oliver Kuperman. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

LLM-assisted writing is a positive good, not a necessary evil, and the LessWrong LLM moderation policy should be changed to allow it without requiring the scarlet letter of AI content blocks.

Well crafted LLM-assisted writing is a good thing, and this community should embrace it instead of shunning it with unenforceable rules. I have already talked about the reasons elsewhere, but I don’t think I have emphasized the benefits of LLM-assisted writing enough.

First of all, LLM-assisted writing likely helps people with writing anxiety/ perfectionism. LLM writing allows you to focus on content instead of form by mostly taking care of grammar and flow for you, which is a significant part of writing, and gives you a “good enough” baseline that you would lack otherwise.

LLM-assisted writing helps with writing blocks, as if you are having trouble expressing something you can give an LLM a rough idea and see it iterate.

LLM assisted-writing facilitates instant feedback. If you have an unclear idea, LLMs will likely flag it and help you improve your argument.

Finally, LLM-assisted writing simply makes the writing process more enjoyable, as the aforementioned productivity gains as well as LLMs transforming writing from a purely solitary activity to something more resembling a conversation really make a positive difference.

All of these benefits far outstrip the downsides of allowing high-quality, LLM-assisted writing (which are mostly illusory) and I would be willing to debate this proposition with anyone. LLM-assisted writing has already been implemented elsewhere without complaint, and the harms of allowing it are largely hypothetical and can be counterbalanced by the upvote downvote system already used by LessWrong.

4Viliam5mo

It is a useful signal from inside -- if you feel more anxious writing about X than about Y, it is a signal you probably know about X less -- but it is not useful across people. Some people are okay generating arbitrary amounts of text, other people get anxious whenever they imagine having an audience.

7gwern5mo

All of that sounds like good reasons that heavily LLM-written text should be banned for being especially likely to be bad, misleading, forgery of real experiences/emotions, superficially appearing to be high quality but missing the ungradable essence, etc.

This conclusion does not follow from the premises I laid out.

I am not claiming LLM writing hasn’t improved since 2023, just that it’s improvement has not tracked with the improvements in math and coding.

Very few people (using LLMs or not) are going to match the writing quality of the very top writers. That doesn’t mean their writing is worthless or should be banned.

Pure LLM writing might not be great at capturing emotion/first person perspective, but there is nothing preventing LLM-assisted writing from capturing this.