Another common pattern I'm noticing is that AI writers love inventing their own jargon. They may or may not also define the jargon.

I vouch for the argument and welcome criticism of it.

Thank you for this. This is the right way to do AI-assisted writing.

I think the math is unnecessary since this is a conceptual argument, especially given that you yourself don't understand the math. A proof isn't worth much because the real question is whether the premises of the proof are true.

As for the conceptual argument in the abstract:

1. I don't expect that a misaligned AI would keep humans alive for instrumental reasons. If you're correct that human brains contain irreplaceable information, then the thing to do is scan humans' brains and save the scan data before killing everyone.
2. Even if the argument is correct, it is (as you say) scoped. I take approximately zero consolation in knowing that ASI might keep us around as slaves for a little while before exterminating us.

FWIW this type of argument has gotten some prior discussion, and the counter-arguments I gave are not original to me. You might search for writings on the question of "would AI keep humans as pets?" or similar.

Update: I have noticed a suspicious pattern with AI-generated user profiles. Profiles that fit this pattern are not always AI-generated, but it's a red flag:

• profile picture with a face
• bio line always says "[senior or fancy-sounding title] @ [company/org]"
• full bio has multiple paragraphs

like they're treating EAF as if it's LinkedIn

This gap isn't too substantial IMO. Compare to the gap on something like immigration (78% of Republicans support a border wall vs. 21% of Democrats) or abortion (20% of Republicans say it's morally acceptable, vs. 78% of Democrats).

Thanks! I wouldn't describe this as a negative result, more just that it's not obviously a great idea, and there are some important downsides. Further investigation might reveal that it is, in fact, a great idea. (Probably not, but maybe.)

Yeah I was surprised in pretty much all the same ways as you!

The trouble with this is that AI-generated posts tend to be very long, which makes it a huge time sink to assess their merits.

Data is not contaminated, as only the first attempt is counted. Feel free to take it as much as you want!

How many bullets were you able to get? I know it's possible to get at least 12 but there might be a way to get more.

Couple reasons why I browse /allPosts instead of the front page:

1. Front page turnover is very slow—posts stay there for multiple days. I check EAF multiple times per day most days.
2. I want to at least read the title of very post. Many posts never show up on the front page.

On completing the quiz with no conflicts I was invited to take the quiz again to explore possible conflicts. Did you exclude repeat quiz takers from your data analysis?

Yes, if you take it multiple times, only the first one is included in the analysis.