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.
Being Pro-Choice Seems Epistemically Arrogant, Thoughts?
When it comes to the issue of abortion, both sides seem to speak with a sense of certainty I do not believe is justified. However, the moral implications of a government embracing the wrong policy under a pro-life framework are far worse than the moral implications of being wrong under a pro-choice framework, so it seems to me it is far easier to support the former.
Note: I think exceptions in the case of rape and the health of the mother are morally permissible, contraception should be freely accessible to all citizens, and public schools should teach the importance of being careful about sex.
I am vegan, but for non-vegans I think it’s more ignorance and self-interest than epistemic arrogance, as the vast, vast majority of most societies are non-vegan, whereas there is at least an order of magnitude more people who support pro-life positions, and people don’t personally benefit from non-veganism in the same way pro-choice folks benefit from their own position.
However, I would claim most non-vegans who have thought deeply about vegan arguments, and don’t make arguments from the angle of nematode welfare, are epistemically arrogant.
I think it's totally in scope to be correctly confident about an issue like abortion even if ~half of people agree with you.
But, I also think your point basically stands: there's an asymmetry in costs to being wrong. (Don't take that too far though, a lot of harm is caused by forcing people to have children that they don't want, and aren't prepared to handle.)
The issue is I think the asymmetry is so great that you have to essentially be like 99% confident in your position to be justified, especially when the compromise position I laid out should prevent a good portion of unwanted pregnancy, especially if you subsidize elective procedures such as vasectomy and provide more childcare support. I think the best arguments for the pro-choice side are consequentialist in nature, but the consequentialist calculations are very complex in this case, so I think it is better to rely on a deontological framework, which is what most people do anyways.
1I think we should be careful about reasoning from "if X is true, the consequences are so much worse than not-X, so we should act as if X is true", since there are many ideas that are in fact exponentially unlikely (eg Pascal's wager, and analogous situations). The first line of attack should always be just trying to figure out what's true on the object level.
But yeah, I think this line of argument is basically valid.
This is indeed how I think people should think about the question of animal consciousness. I feel less strongly about the abortion question for various reasons, but I am much less sold on "pro choice is obviously correct" than most liberal Americans.
I mean, I think the real issue with the originial Pascal’s Wager is it’s actually impossible to actually calculate probabilities or assign utilities to different outcomes, as there is a literally infinite amount of conceivable deities (including deities with opposite pay-off matrices) and also infinite utilities kinda break decision theory.
For most realistic scenarios though, I think Pascalian logic is more or less necessary to some extent, and without it you get pretty bizarre conclusions. Like imagine a genie offers you a button which has a 99.9% chance of saving a life but a .1% chance of killing everyone, I think it’s pretty clear that pressing the button is a mistake.
The pro-choice and pro-life debate is a lot more complicated than the button example, but no one seems to want to do the actual work of justifying their viewpoints despite significant opposition being present, a status I would like to call epistemic arrogance.
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.
Can you name 3 authors, on or off LW, whose outputs have been dramatically improved, let's say at least >3x in constant-quality quantity or constant-quantity quality, in the past year due to LLM assistance?
Not sure about the 3x bar, but there are numerous people who have either professed the considerable usefulness of using LLMs to assist their writing, or have used AI to produce highly reviewed works. Besides myself, there are Vauhini Vara (Pulitzer Prize finalist), Stephen LaMarche, and also politicians, such as the Lord Mayor of Cork, Fergal Dennehy.
However, I think this exercise is flawed because there are probably a lot of people who use AI to assist their writing without outright disclosing it.
Those are weird examples given how 'numerous' you claim your examples are. Like your chosen #2 is... a 2023 novel? So you are claiming that even back in 2023, LLMs were so good that they plausibly pass the bar of considerably increasing writer productivity? How good are they now, in mid-2026, then, after extraordinary capability increases in other areas (such as coding or math), and why is it so hard to see this gain in thinking/writing?
because there are probably a lot of people who use AI to assist their writing without outright disclosing it.
AI writing is highly stereotyped and easy to spot, even without Pangram. If there are so many, can you give 3 examples where they look like they are using AI to assist their writing and etc etc?
I think there are multiple things at play here. First of all, I never claimed that LLMs alone and without substantial human feedback could write on the level of top authors. The 2023 book is an example of a passable book written primarily by AI, which demonstrates how AI can provide a decent level of baseline writing quality.
I suspect AI writing skills have significantly improved since 2023, but (speculatively) several factors have limited it including:
- It is likely harder to get AI to be better at creative writing using RL as opposed to math or coding as what makes a particular piece of creative writing “good“ is harder to formalize than whether or not a program works or a math proof is valid.
- There has been less incentive for top AI labs to focus specifically on writing, as the market for AI writing tools was always going to be smaller than the market for AI programming tools, and being able to write a great blog post does not clearly move you closer to RSI in the same way that being able to code or do math does.
- The people working for and running these AI companies have more expertise in math and coding and so are better designing targeted training for those fields.
- Finally, much of writing is about relating real experiences and emotion. AI often lacks a human touch because it is simply not human, and while AI might be able to overcome this, this is a barrier that is simply less prevalent in other fields.
I think the first three bullet points apply decently well to “thinking” too.
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.
I think this is wrong. Helping people be less anxious in churning out words is bad. Perfectionism is neutral, but reducing people's anxiety when writing is bad. "These words are written poorly" is an incredibly useful proxy for "this person doesn't know what they're talking about."
Treating writing as a staple crop (like wheat) rather than a complex product (like chocolate milk) is pretty flawed: More isn't always better. The steps taken to create it matter, and the work done to produce it materially changes the final outcome. You can absolutely ruin the final product by changing the process or substituting components (chocolate milk with artificial sweeteners tastes uncanny even to people who like artificial sweeteners; LLM text and text written predominantly by people who read LLM text feels bad to read, and is generally shallower).
Even if LLM text wasn't worse aesthetically, which at the present time it seems to be, the end result of allowing people to use LLMs in writing would be worse text from authors who had less to say and lower standards. "Enjoyability of writing" is not something sane to optimize for, and LLMs actively harm most of the things that people value in reading. The fact that a human took a substantial portion of time out of their own life, to convey something that meant something to them to others, is a way to mitigate DDOS attempts on people's attention, and the act of doing the labor of writing yourself implies you understand the topic in a way that isn't demanded of you when writing with an LLM.
I agree with the anxiety being a useful signal. I frequently have trouble writing something and realize it's because I don't have a clear thesis, or I'm trying to argue something I don't have good evidence for, or there's some other problem with the post. I could probably get an LLM to write something but it would be hard for it to write something good from a muddled idea.
(This can go too far, and forcing functions like Inkhaven are helpful, but presumably the goal is to get enough practice that you get better at coherent writing, not to just ignore your instincts and publish slop)
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.
This all reads as a very elitist and privileged perspective. The ability to write well is not the same as the ability to think well, and what constitutes "high-quality writing" is highly subjective. There are numerous studies linking writing status to socio-economic status (example), and it is quite obvious to me that writing skill, which is largely an acquired skill, is correlated more with socio-economic status than raw intelligence (Opus 4.8 agrees with me, ask any chatbot, and it will probably agree too).
At the end of the day, I guess it is inevitable that we need some additional tools besides our own brains to decide if something is worth reading, as even without AI there is only so much writing we can read. However, the year is 2026, where there are more moderation tools than ever before, so why should we still use "writing quality" as the primary metric to judge effort? There are probably far better gauges of this that are available that would allow both human and cyborg writing to flourish.
On your other points, I do not think LLM writing is measurably worse stylistically (after a bit of prompting/editing) than the text produced by most people on this forum, and I do not think most people can reliably distinguish LLM text from human text after the most obvious tells are removed (overuse of em dashes, it's not just x, it's also y sentence structure, overly formal, impersonal writing, etc.)
I do not see why the enjoyability of writing is not important, but the enjoyability of reading is (you suggest that LLMs writing might be less "aesthetically" pleasing, but the only purpose of aesthetically pleasing writing is to make the reader). If writing is more enjoyable, better ideas will likely be produced, and it will drive more people to engage on the site than before.
Finally, I am not advocating for one-shot "LLM write me an essay on x" posts; I am advocating for allowing writers to use LLM assistance in a manner that still requires deep reflection and thought. How we can allow the latter without the former is a challenge, but not one that is insurmountable.
It isn't elitist nor privileged to say that writing skill matters. Raw intelligence isn't a useful proxy metric for whether someone's thoughts are valuable, but writing is, and this wouldn't change if it were correlated with socioeconomic status. You just cited a chatbot's output instead of backing up your own point. This demonstrates shallow thinking. The bot has diminished your ability to effectively argue for your position, and worse, it has gotten rid of something that would have helped you think better: Being forced to confront the existing data available yourself.
To quote a Richard Guindon comic strip:
Writing is nature’s way of letting you know how sloppy your thinking is.
Requiring people to have acquired the skill to write prior to using a text-based discussion site is the correct approach. There are virtually no benefits to allowing LLM-assisted writing and an endless series of cons.
Writing skill past a certain point is pretty useless in most people’s everyday lives, especially in our current age. It’s a fine skill to encourage, but I don’t feel like it should be used to gatekeep unless it’s obviously necessary, especially in the light of the long history of people using poor correlates to intellectual capacity and effort to judge others (Examples: Latin, Cursive).
On my use of claude to elucidate my point, I already provided justification for my view that writing skill correlates more with socio-economic status than intelligence or effort: writing is more of an acquired skill shaped by the taste of society than an innate property (like intelligence). I don’t wish to spend much time thinking about a point that is rather trivial.
How long you spend thinking about an idea isn’t based on societal preferences in the way that writing skill is.
1
1It scares me. The writing quality of the LW feed is exceptional. The only other venues that aggregate similar quality writing are fancy magazines that gatekeep and pay authors and publish much less.
If you look at other "group blog" style venues online, LLM writing has hurt several of them substantially. Hacker News and Lobsters and numerous subreddits have been suffused with LLM-written posts that are superficially good in the LLM style (so it gets upvoted by people who don't notice that it's LLM-written or who decided they are willing to just trust what LLMs or unspecified centaurs say.) If you don't find LLM output reliable then that sucks. Fortunately, the LW moderation effort is also exceptional, and aggressively hunts down LLM-written posts that skimp on quality, which seems to be preventing this from happening.
Maybe it would be OK to open the floodgates further and rely on karma-weighted voting, and on the administration's continued adaptability, to keep quality high, but we have a good thing going here so I would be really scared to kill the golden goose.
I have no problem with moderators taking down posts they find to be low quality, I just don't want a blanket ban on LLM output unless you put all your LLM-Assisted writing in distracting content blocks.
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 the current LLM detection system is flawed, so I proposs a bet. In the proposal is accepted I will write 10 3 paragraph less-wrong styles post, 5 AI generated and 5 normal. If the LessWrong LLM detection system correctly assesses 8/10 or more, I lose the bet. If the detection system correctly fewer than 8/10, I will win the bet. The LLM posts will be fully AI generated, but I will be able to provide it a style sheet and instruct it to mimic my style. Additionally, I can give up to 2 prompts of feedback to modify the post. The human posts would all be generated by me, but would try to follow the same style sheet and use grammarly for spelling check. Finally, the LLM system must run without human intervention. Even if a human judge can “feel” something is LLM, this is not what the experiment is designed to measure. I would be willing to wager up to 500$ on this.
I think you could probably succeed at this, but you're testing something different from where the system's value comes from. While the anti-LLM policy is trying to prevent all LLM-written posts, most of the value comes from preventing low-effort LLM-written posts. If you can get a high-effort LLM-written post past the filter then that's sort of bad, but doesn't mean the filter isn't doing its job.
I think you could easily fix this by ensuring you adhere to the same non-LLM style guide and (if both sides are paranoid) having a third party to judge if sandbagging occurred.
The policy should be changed to target what can actually be enforced then. Unenforceable policies on morally neutral behaviors only punish good actors and promote bad ones.
It's not a neutral behavior though. It's much easier to spam the site with LLM-written posts, and obviously LLM-written posts are bad for reasons that are related to why it's obvious they're LLM-written. If you prevent people from spamming the site with low-effort slop, and some people still use LLMs in a high-effort way to write better posts (even though this is still against the rules), then that's still a win.
I think we agree that low-effort, unreviewed LLM posts are bad, but this should be targeted instead of high effort cyborg writing. I don’t like it that posts which lightly use LLMs need to use these weird LLM boxes.
Not sure if you saw this, but the post in my AI writing experiment is officially 100% human-written (although low-confidence). This was surprising since I told Claude that I think AI style is fine, that the experiment was entirely about content, and it could just write naturally.
I did’t see that, but yeah I think all but the most blatant examples of AI writing are difficult to spot reliably, especially without substantial false-positive rates. The thing no one seems to get is the current policy doesn’t meaningfully change LLM use for anyone but the scrupulous.
This isn’t that important, but the “term Bay Area” has always annoyed me (as someone from Tampa), as it appropriates a generic geographical feature that many cities contain. Should we start calling Denver the “Mountain Area”? New York City the “Island Area”? Phoenix the “Desert Area”? Oklahoma City the “Plains Area”? The only reason the “Bay Area” sounds normal is because people are used to it.
What are other similar cases?
There are 75 islands in the world bigger than "the Big Island" (Hawai'i); but it's the biggest of its archipelago, and the nickname disambiguates the island from the state.
In petroleum-driven geopolitics, "the Gulf" often means the Persian Gulf; e.g. the 1991 "Gulf War". In this usage, "the Gulf states" are Saudi Arabia and its neighbors. But I'm guessing that in your area, "the Gulf" often means the Gulf of Mexico, and "the Gulf states" are Alabama, Florida, Louisiana, Mississippi and Texas.
"The Mountain West" means Colorado and its neighbors, not the French Alps, even though the French Alps are in Western Europe and are mountainous.
"Sahara" is just Arabic for "desert".
For your examples, they are either used by locals informally ("The gulf" as a shorthand for gulf of Mexico is rarely used by national publications), by foreigners (who often give illogical or simplified names to things because the original would be too complicated), or apply to very large regions (which in my opinion, is less annoying because it is less self important). The fact that the San Francisco area gets its own special name that is frequently used despite the fact that there is nothing especially impressive about that particular feature of the area is slightly annoying imo. If people in the Bay Area want to colloquially call it that, I wouldn't mind, but the fact that national news refers to it as such is slightly annoying to me.