The AI Slop Problem Nobody's Talking About | Substack CEO Interview

Hey, my name is Chris Best. I'm the co-founder and CEO of Substack.

Yeah. You know, I guess in some ways slop defined broadly as, you know, content that nobody believes in is, I kind of think of as the foil for Substack. The thing that we exist to do in the world is help people make things they believe in and make money doing it. And so, you know, defined broadly all the way from spam to clickbait to, you know, generic soulless stuff. That's sort of the stuff that Substack, you know, exists to provide a real alternative to. And then I've had the experience, I think many people have, you know, over the past 12 months on the internet, having this feeling of more and more places that I spend my time and life online feeling kind of like inundated with things that I perceive or I feel as slop. Um, things that are, you know, thoughtlessly made or that are presented as sort of deep work by somebody but are actually kind of recognizably like clawed copy-paste. Um, so I sort of as just like a consumer of the internet, I kind of had this feeling. Um, and then there's been, you know, I started talking to the guys at Pangram who are one of the sort of leading, you know, AI text detection things and they have, they had launched a Chrome extension where they could, you know, you can kind of turn it on and see for yourself what you're reading online. And they put together some stats of kind of like how prevalent is this on a bunch of major platforms. And the stat that kind of blew me away was for long form writing on LinkedIn something like 40% of it was fully sort of generated according to their measures. And that's not a perfect correlation. It doesn't mean that, you know, I think of it as not all AI is slop and not all slop is AI, but if you just think of, you know, the share of text that's human written changing dramatically in the places where we spend a lot of our time online, it just seems very obvious to me that this is going to be a major, uh, a major change and a major thing that platforms and people get to navigate.

Yeah, that's the real dividing line to me. You know, I have no interest in policing what tools people use. You know, we're sort of anti-authoritarian, pro-freedom, give creators the means to do the thing. But I do think there's a dividing line between I'm using tools to make something I believe in versus I'm kind of gaming a system cynically to skip the step where I have to believe in anything at all. You know, the extreme version just being, yeah, you know, hey, Claude, write a thousand viral posts, make no mistakes, off you go. And the problem with that is it's effectively like a denial of service attack against the public square, right? It's not just a problem of, oh, I made a thing and it was bad. Even if it's good, the problem that happens is then as a reader, as somebody who's spending time online, you have to be constantly wondering, is this real? And under those conditions, it's hard to discover and find value in new things. Like as I've been having these conversations, I've talked to people that have told me things like, I stopped reading new writers or I had to stop going in my comment section at all because a few times it happened that I would get sucked into some long conversation with somebody and it slowly dawned on me that it wasn't a person at all. It was kind of somebody had done this thing that said, hey, you know, do some replies and max out my account. And I think, you know, there are some people who are sort of anti-AI overall and who think, you know, only humans should ever do anything. You know, we have space for those people, but I'm not coming at it from that perspective. You know, the way I see it is like, hey, AI gives us new superpowers, but it also gives us brand new ways to be annoying on the internet, and we're just going to have to deal with that.

Maybe the biggest thing is I do not feel like we have all the answers or even should have all the answers. This is kind of like a new world we're all kind of entering this new world together. And at the same time, you know, because things are moving so fast, I don't think we have the luxury of waiting around indefinitely and kind of like we'll just see how this all shakes out because I think you could, you know, you could wake up and have the Substack app feel like LinkedIn very quickly. Um, and so the thing that we're, you know, I think this will be out by the time this airs, the first step we've taken is to build an integration into the Substack app with the Pangram tool. And so if you're just using the app as you would, there's no change. But if you're somebody who's curious, you can go to a long form piece of text and basically do a scan that says, hey, according to the Pangram estimate, do you think this text was, you know, AI generated or do you think it's written by hand? And crucially, that's not the same thing as was this made with deep care or not. And it's not was this even AI used in various parts of this. The thing that the technology allows us to do with like a reasonably good accuracy is kind of just be like, you know, did this text go through an LLM as it was generated? And my view is, hey, we're going to kind of just start with like put that on the table. Let people, you know, let people see that fact for themselves and have an opinion about what that means for them. You know, my experience using it has been, you know, I think there are some times when it's almost everybody thinks it's totally fine. It's like, hey, I'm looking up some information from ChatGPT. You know, almost everybody does that. Almost everybody finds that useful. And then there's kind of like the other end of the spectrum. I don't know if you saw this when the OpenClaw thing was really hot, but there was a guy that was having his OpenClaw bot respond to text messages from his wife.

And I think you know that's the other end of the spectrum. I think like almost everybody sees that as maybe not being the kosher use and then the thing is like I think so there's kind of this spectrum of in what spaces online is the implicit contract between reader and writer. Hey, I'm going to make something for you that's carefully crafted and I'm going to use a bunch of tools. And then are there other spaces where you know in the comments section or in your DMs or in these other places like I think part of my theory here is that the places that are responsible AI use should stand up to transparency, right? You should be able to say, yeah, I used this, this is how I use this, and the reader should be able to see that and say this is cool. We're also giving a little option for you to add like a how I make this statement. So, if you want to provide extra context about, hey, here's my process or here's what you can expect from me. You know, we're kind of starting with just giving some transparency on both sides of the relationship, giving writers the tools to check these things for themselves. And I kind of see it mostly as a way to sort of open this conversation. We're not coming here and saying what you can and can't do. We're not coming here and shutting anything down or doing anything like that. We're kind of as a first step. We're just introducing a one tick of transparency so people can see what's happening.

And yeah, that would be bad for everyone I think.

And how hard do you find you have to push it both on the prose and on the ideas themselves?

Right. This is where people might in the past have said that writing is thinking.

Yeah, this has been one of the interesting experiences I've had as I've talked because I've talked to a whole range of people across the spectrum of AI use all the way from like I would only ever type things with my human fingers all the way to, you know, people. To me, the thing you're describing is, you know, what I would think of as like sort of like high perspective use of the tools, right? It's like, hey, I'm actually trying to get my perspective out into the world. I care a lot about the outcome. I'm using this tool. And an interesting thing I found is that everybody across that spectrum has had like a really strong worry about the AI slop quotient because the problem is there's a low effort version of that that can be hard to distinguish at first blush, but that can kind of take over the comments, can take over like it can make it harder for that stuff to be differentiated and break through. And this is part something we talked about that I think might be interesting to get into here. I think there might be a sci-fi technology that I wish would exist that doesn't exist yet that would be the perfect thing because, you know, the part of the thing with the Pangram thing we're giving you the ability to put down some notes of your thing because the thing that we're going to find is if you're somebody who does 17 drafts and thinks about it and bounces it back and forth, you may end up with something that is still like the text of the final thing is still AI output.

Right. The text according to Pangram or whatever is generated by the thing but it's used to be the case that basically when you read something on the internet, if you read a long thing, it's almost by itself a proof of work like somebody has put enough thought into this to at least write it out. Maybe terrible thought but at least somebody spent this much attention. And now it's actually like it's hard to differentiate that from the person who just says, hey, write me 10,000 plausible sounding AI takes. Make no mistakes. I'm curious like how you think about that and what you think if you think there's like a technological root thing that we actually want is there.

I think that would be useful because you know one thing it's a tempting thing is for people to be like oh there's this thing that's average to the AIs. It doesn't mean that that thing is bad, right? There's a, the way I sometimes put this is like Claude is great. I love talking to Claude. I talk to Claude a lot. I get a lot of value from Claude, but like when I want Claude's opinion, I'll go talk to Claude. I don't need you to be a thin wrapper on Claude for me, right? And so there's humans to be thick wrappers. We want humans to have really strong opinions and to drive what they want out of AI, not the other way around. And that might get more true as the models get better, not less, right? It's not the case like, oh, well, if the prose was just a little better or if the ideas are better. The model can get arbitrarily better and more powerful. And what that means is in the places where the model is deeply valuable, it is going to be like that's not what we're going to turn to humans for anymore, right? Anything where it's like I could just ask Claude, people are just going to ask Claude and that's going to be great for a bunch of things. And then the place that's going to have continued value is in the places that are not in the central distribution. A lot of which may just be, you know, maybe noise or maybe wrong, right? If you expand that landscape of ideas into higher variance, a lot of what you'll get is stuff that is mistaken or stuff that is like, you know, whatever. But that's also where you'll start to get like the real data of new things. Almost like if you think about where the alpha is, the alpha is going to be sort of definitionally in things that are not that you can't just easily get out of the thing.

Yeah. We're a social species. And even if you can have kind of like superhuman media in every part of latent space from the model, people are still going to care about discovering and caring about and paying attention to those things together. There's like another layer that happens that is not, you know, even if you had a super strong AI that could create any work of fiction or any piece of thing in the most convincing, most artful way, people are still going to care about other people experiencing and reacting and having real lives around these things I think.

Very cool. Well, I should ask you this. You're a user of Substack both as a writer and a creator and I hope as a reader a bit. What if you could have anything for your experience around this topic, like what would you want? What would make this, what would be the best way to serve human creativity on a platform like ours?

Yeah. It's like what is the opposite of slop and how could we find and reward that? How could we reward the stuff that's real, which is different, you know, because even AI aside, standard attention-seeking social media algorithms will tend to kind of sand off and homogenize things and find the things that best push certain base buttons or whatever. And the thing you're saying here is kind of like, okay, what's the thing we actually want? What's the thing that's deeply valuable? And can we get a stronger and more reliable signal of that as a benefit to the reader directly, but also to the creator who wants to make those things and can feel like, hey, if I actually take the time to make something deeply thought out that kind of pushes the discourse forward, I'm not just going to get totally drowned out by 10,000 people writing clickbait stuff, whether they use ChatGPT for it or not. Which in turn increases confidence that this is a platform for that kind of work. And this is a platform where that kind of work is highlighted and I think that produces a sort of positive feedback loop when you're looking for fresh voices and you want to find new perspectives etc. I think that would be really interesting.

So I love that question. I think that the best signal that I've seen is you're looking for an ability to develop a coherent thought that acts a little bit like a virus. Like it's a net new thought. It causes you to pay attention to it. You can't get it out of your head. It sparks new connections here, there, and everywhere. And I actually think that one of the good examples of this, outside of Substack, is Paul Graham's blog. And he just writes these things that when he writes them, they sort of stick in your head. Naval does this, too. And you just look at the world a little bit differently as a result. And I'm hungry for more of that, right? Because what I'm hungry for is challenging my own thinking, challenging my own assumptions, not making the same mistake twice, making sure that I read the world as it actually is and not necessarily as I want it to be. And those kinds of viral ideas are not always viral in the sense that they make it mainstream and the algorithm always rewards them, but they're viral in the sense that when you find a reader for them, they kind of stay in your head all day and you kind of can't get it out because it's so densely expressing a perspective that puts a fresh coat of paint, a fresh sensibility on something that you may have stared at a thousand times, right? I think about this because I think one of the moments I had for this was actually from the Anthropic engineering blog when they wrote about JSpace a couple of weeks ago. And I feel like that blog post got read and went viral in ways that I don't think were super relevant because it got sucked into the consciousness conversation. But to me the most interesting part of it was actually thinking about how brains work and how LLMs work as computational solutions to problem spaces. And essentially very narrowly what they're suggesting is that it may be a natural evolved response to difficult problem spaces to use a workspace approach where you evolve essentially higher order bits of knowledge and compute with those bits of knowledge rather than the base tokens that you're working with. And that also happens to be one of the sort of approaches to how people think we do consciousness. I didn't think about the last time I took my last breath. It just seems to be working well. And I found that really sticky, right? Because that helped me to think about LLMs as essentially computing reality in ways that was emergent. And so I've just been kind of chewing on that ever since. And so that kind of a thing like it went viral. I think it went viral in a weird way, but I think the value was still there.

Right. You're almost looking for an indicator of how much the reader has changed.

It is also interesting too, you know, I think some people read the anti-slop stance as an anti-AI stance, but it is really interesting to me that you don't, you know, you don't see Sam Altman or Dario or the Anthropic engineering blog posting Claude slop. You see writing a lot of really thoughtful stuff. And I think that's illustrative that it's not a question of you don't have to be anti-AI. In fact, I think anti-slop is a pro-AI stance understood correctly in my mind.

I guess, you know, my real hope is to help spark a conversation here and to help kind of bring the topic into the open and have this, you know, we can sit here and think about this at Substack all day, but there's this amazing community of people who are thinking about these things a lot and I kind of want to get their ideas on it. You know, we have all perspectives on the platform. All the way from, you know, Sam Chris wrote a piece, the title of which was, 'If you let AI do your writing, I will come to your house and kill you.'

That's right. All the way to, you know, people like you and others who are kind of, you know, doing these like, frankly pushing the frontier of what's possible with these things. My hope is that we can start to have a sense of like what the new norm should be and what the like what are the things that are actually going to solve for the underlying question of value and how are we going to defend against the cases where you know low effort stuff threatens both human and thoughtful AI authorship. You know what is the version of that that can sort of durably not just survive but reinforce and help the real stuff win? That feels like the end goal to me.

And if you look at stuff forwarded by people of my parents' generation, we're already there for some fraction of the population.

Yeah, I think it'll be very interesting. And I think we'll need new norms around a bunch of this stuff, right? We'll need norms where, you know, we never used to have a norm that says you don't have the AI text your wife for you because that was never a thing. And then the minute it becomes technologically possible, there's kind of this moment where it's like, oh, there's a norm for that. Like it can argue with the customer service person and that's great. But like, you know, if I found out right now that during this whole conversation you weren't really here and I'd been talking to your AI avatar the whole time,

I would be upset. And it's not because I'm anti-cool avatar. I'm just like that would be a mismatch of my expectation.

Yeah. Have a conversation with a person. And it would feel like, you know, what I thought was like a mutual investment of effort and energy in these things turned out to be like, I sort of think of this as the writing version of this as claude fishing by analogy to catfishing where it's like the problem is not that you use the technology. The problem is that you betrayed the expectation. And I think the thing that's happening with writing, it's going to happen with video is we have a bunch of expectations. We have a bunch of sort of like implicit philosophical beliefs that we don't realize we have because it's never come up before. Then all of a sudden,

The norm is challenged and then it's like, wait a minute, can you use Claude to text me? Can you use Claude to write a comment? Can you use Claude to have a video conversation with me? Like when is it okay? Do you have to tell me? These are just I think they're going to have to be cultural outputs and expectations. These things through a process that I think is quite hard to predict and I don't think the answer is going to be nobody ever uses AI to make video. I think it's going to be an incredibly powerful tool in all kinds of ways and the thing we're going to have to find is kind of like what are the ways that remain valuable. As the people talk about Jevons paradox a lot, but I also think a lot about Baal's cost effect. You know this thing?

Where it's sort of like, okay, as some things get much cheaper, what remains valuable? Like what's the equivalent of a string quartet? And how do we guarantee that the what are the norms that support those things existing and being findable and being supported so that you know that it's good to invest your time in them because you're getting some version of the real thing. I don't know the answer for video, but that feels to me like the right question. At least.

It's the last scarce resource.

As a small discretion. This has kind of always been the theory of Substack and we thought about this. You know, one of my the story I tell myself about the early internet and social media is that it used to like you used to be able to get bored.

There used to like when I was a kid, you could have this problem of like, man, I have two minutes right now and there's nothing to do. This sucks. And we've kind of like even before AI, we have solved that problem. Any time you spend, you know, and before it's like, oh, give me something free to kind of like slightly distract me. It's like, oh, that's a great deal because like I was bored and now I have something. And now we live in a world where every piece of my attention that I could possibly want to spend, I can spend. And the question is like, how do I spend it well? And that's a very deep question, but it's a worthwhile question because I think as a parent, that's what I think about for my kids is am I teaching them how to spend their attention well? Am I teaching them the habits? Am I modeling the habits of how to spend your attention well that they can learn from and understand essentially what it means to be human by learning how we allocate attention toward one another in positive and constructive ways.

I think you'll see more in personal words.

I was going to say any final words.

Subscribe to my Substack. Check out the new feature. I always tell everyone else to plug their sub. I'm cv.substack.com. I tell everyone to plug theirs. So, I got to live up to that.

Fantastic. Thank you very much. This has been great.