The new editor we’re working on has “AI-written-section” as a first class type of paragraph block, with an intended norm that all AI content needs to go in said blocks. (expecting that posts that are entirely-AI-content-blocks usually won’t meet our quality standards and get downvoted and/or mod-delisted on a case-by-case basis).
As an alternative, maybe you could require users to supply whole-post-level metadata about AI assistance when they’re submitting a post?
I’m imagining some sort of structured input area in the post editor that has a checkbox for “This post is unambiguously 100% human-written” (wording could be improved but you get the idea), and if you check that box then you’re good, you don’t need to do anything else. But if you don’t check the box, then you have to fill out some more info (maybe including some free-response text fields?) that clarifies precisely how you did and didn’t use AI.
The advantages I see of this, versus the new block type:
It makes it harder to lie by accident.
If I’m a new user submitting an AI-written post, it may not be obvious to me that I must wrap the entire post in a specific block type that most block-based editors don’t even have[1], or else I’m effectively making a false claim to having written the text myself.
But if there’s a checkbox in front of me with a clear textual label, and I have to decide whether or not to check it in order to proceed, it will be much more obvious which actions in the UI amount to lying.
It provides the opportunity to “nudge” users away from making types of posts which tend to be low quality, by making it a bit onerous to submit them (without lying), while being more flexible than a total prohibition on those kinds of posts.
Like, if the metadata form-filling experience in the AI-assisted case requires the user to write out a bunch of info about exactly how and when the AI was used, that might be enough to discourage people who are submitting relatively low-quality “slop.”
And it could defuse the cheap thrill of saying “look, I’m doing cutting-edge research despite having no background in the field!” when that actually means “I asked Claude to do something I only halfway understand.” The thrill, I think, involves the conceit that it’s fundamentally “your” work; having to spell out Claude’s extensive role ruins the fun (in a good way).
It gives you something to look over when making moderation decisions that’s informative but more concise than entire posts. And you could potentially create new policies over time that take this metadata as input, allowing you to adapt to AI progress and adoption without changing the technical design.
It potentially extends to types of AI assistance that aren’t literally “the AI wrote these specific blocks of text.”
So if you wanted to, you could discriminate AI writing assistance from AI research/coding assistance, and so forth.
(Related: my sense is that the main problem with the “novice does research” posts is not that they’re written by AI—although they generally are—but that the underlying research was also conducted by AI, and has not been sufficiently vetted by a human with relevant expertise. This seems very different from, say, a seasoned researcher asking Opus to spruce up their rough draft because they need to hit a deadline[2].)
I have zero familiarity with the details of LW moderation so it’s possible I’m barking up the wrong tree with some or all of the above. Just thought I’d throw it out there.
(which might interact badly with the markdown-like formatting that LLMs typically expect to be able to use, but I’m sure you guys have thought about this aspect already)
I mean, I tend to think that that is also usually a bad idea, for reasons that are (sort of) illustrated by Dagon’s experience in another comment on this thread. But in any event, people do it, and it’s not the same as the novice-researcher-slop thing.
I do think this is a good alternative to at least consider, although I think some of your details are off.
In particular, re:
And it could defuse the cheap thrill of saying “look, I’m doing cutting-edge research despite having no background in the field!” when that actually means “I asked Claude to do something I only halfway understand.” The thrill, I think, involves the conceit that it’s fundamentally “your” work; having to spell out Claude’s extensive role ruins the fun (in a good way).
Most of these people are pretty excited to share that it was coauthored by Claude.
As an alternative, maybe you could require users to supply whole-post-level metadata about AI assistance when they’re submitting a post?
I’m imagining some sort of structured input area in the post editor that has a checkbox for “This post is unambiguously 100% human-written” (wording could be improved but you get the idea), and if you check that box then you’re good, you don’t need to do anything else. But if you don’t check the box, then you have to fill out some more info (maybe including some free-response text fields?) that clarifies precisely how you did and didn’t use AI.
The advantages I see of this, versus the new block type:
It makes it harder to lie by accident.
If I’m a new user submitting an AI-written post, it may not be obvious to me that I must wrap the entire post in a specific block type that most block-based editors don’t even have[1], or else I’m effectively making a false claim to having written the text myself.
But if there’s a checkbox in front of me with a clear textual label, and I have to decide whether or not to check it in order to proceed, it will be much more obvious which actions in the UI amount to lying.
It provides the opportunity to “nudge” users away from making types of posts which tend to be low quality, by making it a bit onerous to submit them (without lying), while being more flexible than a total prohibition on those kinds of posts.
Like, if the metadata form-filling experience in the AI-assisted case requires the user to write out a bunch of info about exactly how and when the AI was used, that might be enough to discourage people who are submitting relatively low-quality “slop.”
And it could defuse the cheap thrill of saying “look, I’m doing cutting-edge research despite having no background in the field!” when that actually means “I asked Claude to do something I only halfway understand.” The thrill, I think, involves the conceit that it’s fundamentally “your” work; having to spell out Claude’s extensive role ruins the fun (in a good way).
It gives you something to look over when making moderation decisions that’s informative but more concise than entire posts. And you could potentially create new policies over time that take this metadata as input, allowing you to adapt to AI progress and adoption without changing the technical design.
It potentially extends to types of AI assistance that aren’t literally “the AI wrote these specific blocks of text.”
So if you wanted to, you could discriminate AI writing assistance from AI research/coding assistance, and so forth.
(Related: my sense is that the main problem with the “novice does research” posts is not that they’re written by AI—although they generally are—but that the underlying research was also conducted by AI, and has not been sufficiently vetted by a human with relevant expertise. This seems very different from, say, a seasoned researcher asking Opus to spruce up their rough draft because they need to hit a deadline[2].)
I have zero familiarity with the details of LW moderation so it’s possible I’m barking up the wrong tree with some or all of the above. Just thought I’d throw it out there.
(which might interact badly with the markdown-like formatting that LLMs typically expect to be able to use, but I’m sure you guys have thought about this aspect already)
I mean, I tend to think that that is also usually a bad idea, for reasons that are (sort of) illustrated by Dagon’s experience in another comment on this thread. But in any event, people do it, and it’s not the same as the novice-researcher-slop thing.
I do think this is a good alternative to at least consider, although I think some of your details are off.
In particular, re:
Most of these people are pretty excited to share that it was coauthored by Claude.