[Edit: This was a test to see if anyone would call me out for writing a post with AI. Good job Eye You for spotting it and calling me out! After 17 people voted, you were the first one to confidently call this post out for the slop that it is.]
I definitely want to either strong-upvote or strong-downvote this, but I’m not sure which.
For strong-upvote: I believe Rationality in general and LW in particular really really need more tests. And unannounced guerilla tests are more holistic measures of applied rationality than tests called in advance. And being able to tell whether & to what extent something is slop is a useful skill.
For strong-downvote: Insofar as this is slop, it’s spam. And insofar as it’s not slop (you did set the topic, structure the essay, and decide the result was good enough to be a functional test), it’s teaching us “remember to shun anyone who seems like they’re using AI to help get their points across”, which blocks out quite a lot of potentially valuable testimony from our already-pretty-insular community. And while I 100% believe you planned this as a test, “haha I was just testing you” is a classic dodge up there with “this was all a social experiment”, so it’s kind of bar-lowering to not have pre-registered your test with an independent third party & then revealed that once the game is up.
In conclusion, I give this post two thumbs up, but also two thumbs down.
I want to clarify something: I wasn’t testing whether an AI can generate nonsense and the community will eat it up. I wasn’t posting bad ideas to see if they would be trusted. I was posting good ideas, but presenting them badly, because I wanted to know how well they would be received when packaged this way. This sort of experiment is something I’ve done many times before.
In the past, I measured this phenomenon by writing badly. However, my rhetorical skill has advanced to the point where this method no longer works. To solve this problem, I let AI do the bad writing instead of me.
There are two ways that a post can be slop:
It can be slop at the level of ideas.
It can be slop at the level of sentences.
The core ideas of this post are the fruits of a project I started 15 years ago and have been working on intensely for 9 of those 15 years. The main ideas of this post are the opposite of slop. The “Courtesan” framing device is fictional, and entirely of my own design (not AI’s), and therefore does not constitute slop either.
What is slop is the individual sentence structure.
it’s teaching us “remember to shun anyone who seems like they’re using AI to help get their points across”
I believe this post is precisely the opposite.
As for the spam concern…
Spam
Here’s how I think about spam: Even when I don’t use AI, I feel like >5% of what I write is slop. Not AI. Just really bad human-generated slop. The reason for this is that deliberately trying out weird and risky things has a long-tailed payoff curve that’s solidly positive in the long run, even though it results in some slop in the short run. (Why can’t I just write it and not post? Because posting weird experiments is how I get signal on them.) <1% of what I write is experiments like this one. With ratios this low, the risk of “spam” doesn’t seem significant to me. Honesty, trust and integrity is what’s important to me.
Tests
And while I 100% believe you planned this as a test, “haha I was just testing you” is a classic dodge up there with “this was all a social experiment”, so it’s kind of bar-lowering to not have pre-registered your test with an independent third party & then revealed that once the game is up.
You are correct to call me out on this. If this post had been well-received, I would have 100% continued using AI. Not in the sense of writing posts with disregard for truth (which I wasn’t doing here in the first place), nor in the sense of producing spam (which I also wasn’t doing here). Rather, I would have offloaded the writing of boilerplate sentences to AI the same way I offload writing boilerplate software to AI.
I already use AI in my posts in the following ways:
Research.
Error checking. (Not error responsibility. It’s just one layer of defense-in-depth.)
Anticipating likely counterarguments.
Checking spelling and sentence flow.
I have not been and do not intend to use AI for:
Abnegating ethical responsibility for the truth or signal of my posts.
This was a test of:
Do I have to write every sentence myself or can I spend many hours talking to an AI first and then have the AI do the drudge work of writing the individual sentences? Would readers prefer this? In 2025, the answer is no. I wasn’t intentionally testing the community’s skill level. It is simply that there is a variable I needed to know the value of, and it produced a Rationality test as an unintended side effect that I acknowledged retroactively.
I believe Rationality in general and LW in particular really really need more tests.
I have several ideas for proper Rationality tests that I’d like to try out. However, all of them take significant time investment. That’s why efficiency in writing is so important. If I can find efficiency gains, then that frees up bandwidth for things like deliberate Rationality tests.
I definitely want to either strong-upvote or strong-downvote this, but I’m not sure which.
For strong-upvote: I believe Rationality in general and LW in particular really really need more tests. And unannounced guerilla tests are more holistic measures of applied rationality than tests called in advance. And being able to tell whether & to what extent something is slop is a useful skill.
For strong-downvote: Insofar as this is slop, it’s spam. And insofar as it’s not slop (you did set the topic, structure the essay, and decide the result was good enough to be a functional test), it’s teaching us “remember to shun anyone who seems like they’re using AI to help get their points across”, which blocks out quite a lot of potentially valuable testimony from our already-pretty-insular community. And while I 100% believe you planned this as a test, “haha I was just testing you” is a classic dodge up there with “this was all a social experiment”, so it’s kind of bar-lowering to not have pre-registered your test with an independent third party & then revealed that once the game is up.
In conclusion, I give this post two thumbs up, but also two thumbs down.
I want to clarify something: I wasn’t testing whether an AI can generate nonsense and the community will eat it up. I wasn’t posting bad ideas to see if they would be trusted. I was posting good ideas, but presenting them badly, because I wanted to know how well they would be received when packaged this way. This sort of experiment is something I’ve done many times before.
In the past, I measured this phenomenon by writing badly. However, my rhetorical skill has advanced to the point where this method no longer works. To solve this problem, I let AI do the bad writing instead of me.
There are two ways that a post can be slop:
It can be slop at the level of ideas.
It can be slop at the level of sentences.
The core ideas of this post are the fruits of a project I started 15 years ago and have been working on intensely for 9 of those 15 years. The main ideas of this post are the opposite of slop. The “Courtesan” framing device is fictional, and entirely of my own design (not AI’s), and therefore does not constitute slop either.
What is slop is the individual sentence structure.
I believe this post is precisely the opposite.
As for the spam concern…
Spam
Here’s how I think about spam: Even when I don’t use AI, I feel like >5% of what I write is slop. Not AI. Just really bad human-generated slop. The reason for this is that deliberately trying out weird and risky things has a long-tailed payoff curve that’s solidly positive in the long run, even though it results in some slop in the short run. (Why can’t I just write it and not post? Because posting weird experiments is how I get signal on them.) <1% of what I write is experiments like this one. With ratios this low, the risk of “spam” doesn’t seem significant to me. Honesty, trust and integrity is what’s important to me.
Tests
You are correct to call me out on this. If this post had been well-received, I would have 100% continued using AI. Not in the sense of writing posts with disregard for truth (which I wasn’t doing here in the first place), nor in the sense of producing spam (which I also wasn’t doing here). Rather, I would have offloaded the writing of boilerplate sentences to AI the same way I offload writing boilerplate software to AI.
I already use AI in my posts in the following ways:
Research.
Error checking. (Not error responsibility. It’s just one layer of defense-in-depth.)
Anticipating likely counterarguments.
Checking spelling and sentence flow.
I have not been and do not intend to use AI for:
Abnegating ethical responsibility for the truth or signal of my posts.
This was a test of:
Do I have to write every sentence myself or can I spend many hours talking to an AI first and then have the AI do the drudge work of writing the individual sentences? Would readers prefer this? In 2025, the answer is no. I wasn’t intentionally testing the community’s skill level. It is simply that there is a variable I needed to know the value of, and it produced a Rationality test as an unintended side effect that I acknowledged retroactively.
I have several ideas for proper Rationality tests that I’d like to try out. However, all of them take significant time investment. That’s why efficiency in writing is so important. If I can find efficiency gains, then that frees up bandwidth for things like deliberate Rationality tests.
You want more tests? Here is a test.