What is the future of writing in the age of AI? And what does Medium want to be? We'll cover that right after this. Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation of the tech world and beyond. Today we have a great conversation for you because we are being joined in studio by the CEO of Medium, Tony Stubblebine. Tony, it's great to see you. Welcome to the show.
Thank you for having me in person. This is great.
It's great to have you here. We're going to cover a lot of ground today. I think one of the things that's fascinating about Medium is you're sort of ground zero for a lot of the questions we ask about AI. Do we want AI writing? How should AI be crawled by generative AI agents, AI engines? All the above. We're going to get into all of it. Let's start right away with AI slop, because why start slow? There was a story in Wired talking about how AI slop is flooding Medium. You took issue with the story. Let me just put the premise of it out there and give you a chance to respond. Wired spoke with this company called Pangra Labs. It looked at more than 270,000 stories. It found that something like 47,000, 47% were AI generated. Is AI slop a growing problem on the internet? How is Medium dealing with it? Is this a fair study that they did on your content?
Oh, these are multiple questions. Yes, slop is an important thing going on right now.
Should we define it? What's your definition of AI slop? Because AI generated stories aren't necessarily slop, or are they?
Right, there's a spectrum of it. So let's start with like every platform is mostly deleting and hiding the things that get posted to it. So like this whole idea that censorship is bad is sort of like because if you run a platform, the majority of what people try to post on your platform is spam. And this is like no one would debate whether or not we're doing the right thing by getting rid of it because it's like Viagra ads or just like outright financial scams or like content that's literally made illegal for good reason, CSAM and whatnot. So like we're already in the business of trying to prevent certain types of content from making it out into the algorithm. And what happened in the age of AI generated content is that there's this slightly more human form of content or at least it looked more human, but it's really nothing. It's just a new form of spam. And so we treat it the way we treat all spam. If we can catch it and block it outright, we catch it and block it outright. If we can't be 100% sure that the user should be blocked or removed from the platform, we'll allow it onto the platform, but really try hard not to let it make its way either into the Google search index or into our own recommendations. So this article from Wired, I think my favorite thing about it, for what it's worth, is the way that they created a meme. Like just this phrase AI slop. I had not seen it anywhere until that one article and then they wrote some follow-up articles.
Yeah, it was already a pretty common term. It's actually surprising to me that you hadn't really, you were already using that. It just felt like that term became the de facto term in that moment and I just like, yeah, I like to see kind of the curve of a meme take off.
And it's a good term. It's the right, we should all be using it. What I took issue with honestly was that it was asking the wrong question. Does it matter how much slop is on Medium or does it matter which is the question they were asking, and does it matter how much slop our readers are seeing, which is the question I'm asking. So we do all of this work to prevent readers from seeing the slop, and I think we're doing a really good job. Like for the most part when I get complaints about AI content on Medium it's what's showing up in the comments, it's not what's showing up in the posts, because our recommendation system is quite good. So I think that was the first thing I took issue with. And then there's something just suspect about the service provider that they used where it's like, that at the same time this article is coming out, that service provider is using the article to try to get me to hire them, and it just felt like a conflict of interest.
Oh, so by the way, AI slop, just to define it, just low-quality written content, sometimes images and videos, but in your case I think written, that's just being posted on the internet for engagement hacking.
Yeah. Right. I mean you could just think of it as like a prompt, like, oh, explain AI, and you go to ChatGPT and say write 5,000 words about AI for me and then you just cut and paste that into Medium. But I think a lot of it is also following the incentives of the internet already, because there's a ton of content out there on the internet that is being generated in really low-quality ways, whether it's outsourced writing or whatever it might be, that's responding to these incentives to get stuff in front of search engines.
And how different does this feel to you than pre-AI slop? I mean those incentives already existed.
Not different. But the question I have for you is the scale, because you're seeing this on the provider end. Are you seeing a real increase in this type of content showing up on Medium because of the scale? We saw like a 10x increase in what people were trying to post on Medium. But also at some level it was kind of a non-issue because the tools we use to battle it are the same tools we were already using to battle spam and to filter out the good stuff into the network and leave the bad stuff out. So we got a huge increase in volume, but not in what the readers see. And I still, I'm like somewhat perplexed that that article made it to print because it was really a nothing burger for us. It was a little bit of extra work, but the same type of work we were already doing.
It's funny you say we got an increase. Is that increase still being sustained, that 10x increase?
Yeah, it didn't turn out to be exponential. So it just came and I think we tightened up. It was just a step function in the change in price. There's not as significant of a change in price to create that slop anymore. It gets incrementally cheaper but it's still like, it was somewhat cheap and then it became very cheap, and that night and day I think just created a one-time change in the amount of trash that was getting thrown at platforms. And I think that's true for all platforms.
You said something very interesting that the provider, the tech company that went to Wired with this data, then came to you and tried to sell you their AI detection services. I hadn't known that.
I mean, the reporter knew. This was happening while the story was underway.
Did they include that in the story?
I can't quite remember. It was like, I don't know. The story is a little bit sus just because of that. But the reporter did at least give our case that what gets read matters more than what exists.
Yeah. In the deck, it was Tony Stubblebine says it shouldn't matter at all. But let me give a counterpoint here about why it should matter. You already brought up search engine indexing and you're trying to make sure that this stuff is not indexed via search engine. One thing that Medium, as someone who writes on Medium, I also write on Substack. We're going to get into the difference between Medium and Substack in a bit. But as someone who writes on Medium, one thing I see is Medium ranks really high in search. I wrote this story, this profile on Dario Amodei. You cannot find it on Substack, but I posted the same post on Medium and it ranks way up there when you're searching for Dario. And I imagine that if I was someone who was in the business of putting up AI slop trying to get traffic and revenue from search engines, even if Medium is not recommending that to readers, it's actually a great venue to try to get in front of people because of the search ranking.
I mean, I feel like I'd rather just lean into, for your listeners, SEO on Medium is great and that's a reason to write on Medium. As far as fighting the AI slop, the thing that we do for them is the same as we do for spam, is we literally remove them from the Google index. We spend a fair amount of time actually making sure that the quality of what we index on Google is high. And I think what we've heard from Google actually is sometimes they're confused by what we're saying no-index to. And I've gotten emails that have brought up stories that literally ask why these 10 stories aren't indexed. And we look at them and we think, yeah, because we don't think they're real enough. We don't think they're authentic enough. We don't think they're deep enough.
It's fascinating because it gets right into this conversation about whether all AI generated content is slop. And I even let off this conversation asking what's the future of writing or blogging in the age of AI, and it seems like AI is getting better at writing. Is there going to be a point where people in the position that you are in say, you know what, it's better than most humans at writing, let's just let it rip?
I am always confused by when people worry about the future of writing. So let me just start with first principles. Writing is thinking and smart people like to think, that's not going away. When you have a thought and you put it onto paper or type it out, suddenly all these subconscious feelings about that thought, all of your life lessons get articulated in a way where just as the writer you understand it better. And so the role of AI in there is not to replace you. If you're someone that thinks there's value to your life and being smart, you're not going to stop writing. And it's the same with reading. We've been down this path with is it valuable to get a summary rather than read the whole thing. This is the CliffsNotes story. The CliffsNotes did not replace the book because there's something about the way our brains work that learn through story and a lot of what we're learning is the human components of that story. It's not just a set of facts. It's the humans involved in that story. What did they learn? What did they think of it? If they're giving advice, why? All of that stuff ends up mattering and it's why self-help books are as long as they are. The advice is usually a page, but then half the book is social proof and that's for a reason, to help you understand why you might make the effort to take that advice. And so I just never have any doubts about trends about writing because I don't think there's ever going to stop being a market of people who think their life gets better if they get smarter. That's not going away. And so where people are getting confused is it just turns out there's a lot of writing that's only done for reasons, like I'm required to give a report or I have to pass some writing up to some bureaucracy. And so you're just checking the box on having provided any writing at all. And yeah, AI can do that, but that was never about thinking or being smarter in the first place.
Yeah, I think we'll both agree that there is real virtue in writing. I mean, I do it as a living. You've done it as a living for a long time. That idea of going from something that you think you know and then you start writing it and you realize you don't really know it at all. There's great value in trying to connect those ideas on the page. That's obviously, you know, we're a big technology podcast. That's why Amazon does their six pages, to try to, although maybe those are going to be AI generated and that's a real question. They do it so they can, instead of having a PowerPoint where your ideas are disconnected, you write it down so you actually know what you're talking about. And I think that's been a big driver of success within that company for a long time. Similarly, when I'm trying to get to the bottom of concepts or think I know certain things and I'm writing it down, I'm like, oh god, I got to make another phone call because I have no idea what happened in the middle of that story that I'm reporting.
That anecdote is so powerful, right? It's that you think you understood what you were talking about until you had to force yourself to put it into words and then you realize there's huge gaps. That is the proof of how important writing is to thinking.
Yeah. No, I think we're both in agreement here that writing is good. Let's preserve it. The other side of it though is that I think, actually I don't even think, I know that both of us believed that AI wasn't going to be or was not very useful in writing and now we both believe that it is becoming more useful. This is something that you wrote in August. Is there a way to use AI to deepen understanding or help writers tell their human stories? Two years ago, we thought the value to writers and readers was less than zero. The AI companies had leeched value from your writing without offering consent, credit, or compensation. They enabled a wave of spam that tried to replace your writing with hallucinated slop. As we've continued working on making Medium the best place to read and write, we've noticed and heard from our readers and writers that some use of AI is starting to be useful. So, it is starting to be useful.
I thought I was going to regret having written that when you started reading it, but now having heard it back, I love it. Like I pulled no punches on that. I think what I was saying there is just a blunt commentary of the state of AI at any given moment. Not that AI could never be valuable. I thought in that moment it was literally just creating problems. And I mean that was true. It was not good enough to really help the writers. It was providing more spam and more slop. And I mean it wasn't a ton of work to fight it back, but it was some and it was creating a real sense of unfairness in the community. I mean our community thinks that their content got stolen and not for a legal reason but for a reason of society. Societies run on exchange of value and for these AI companies to train on the content and offer nothing in return just breaks that sense of fairness that well if you got something I should get something too.
And they crawled and trained on Medium content.
Oh yeah, for sure. This is a lesson that I'm bringing up more often because I think if they don't get their act together, we're going to do it again. We learned really early on that Medium is a big enough corpus of data that we can poison the results of any large language model. And the way we learned this is because my boss and the founder of Medium, Evan Williams, when he started Medium, he was a huge fan of em dashes. And so he's the one that built it early on into Medium that if you do a double dash or dash in certain points, it gets automatically converted to em dashes. And so as a result, em dashes became automatically added into Medium content for a decade. And then also culturally part of Medium, like writers who'd never heard of em dashes are suddenly seeing em dashes. And so the Medium corpus is so heavy on em dashes because that's what Evan Williams thought, he just thought they were beautiful among many typographical opinions of his.
Wait, this is why ChatGPT writes with em dashes.
I'm saying it's because of Medium. Absolutely. In part we popularized it, in other parts of the internet too, but the Medium corpus is very, very deep in those. So when you hear, oh this must have been written by AI because it's got so many em dashes, it's because the AI is trained on Medium which does have a lot of em dashes naturally. So I think that if we can't find some fair exchange of value, if we can't get something going, if they insist on continuing to train and work around our blocks and continuing to avoid even paying for what they're training on, I think we're just like a lot of other platforms are going to quietly catch their crawlers and poison the content that they're training on. And who knows what weird hallucinations we can fit into their training set.
So you would play dirty like that? I mean, this is a prisoner's dilemma, right? You start collaborative. If you're matched with collaborative, you stay collaborative. If you're not matched, then you have to switch too, right? Effectively they're antisocial is what I would say. All of these companies, OpenAI, Google, Anthropic, they could have started with, hey, we want to do this thing. We think we're going to get a lot of value out of it. We want to train on your content. How can we help you get value out of it? That could have been the starting point and instead their choice, completely their choice, they took the starting point of we're going to train on your content, what are you going to do about it. And it's taken the industry a while to react but we are reacting right now and I wish we had reacted as an industry earlier. But now that we do, I think we're going to have to see which of these companies want to come to the table and which are going to continue to be antisocial, and if they are, they're going to be met with antisocial behavior, of course.
Okay. So we have some news about the steps you're taking here to fight back. We're going to get to that in a minute, but I just want to keep following the thread that we started on, which is this idea that this was not very useful for writing. AI tools were not very useful for writing a little bit ago. Now they're becoming useful. So where have you found the useful areas?
Oh, yeah. I think the thing we're most excited about is the idea of an AI agent acting as your assistant as you're writing. So as we're building, we're building writing tools right now. We're about to actually announce that we've broken ground on a new writing app for Medium.
I mean, come to Medium Day on Friday.
This is going to go live after that. So, it's a companion to Medium for all of your writing, for your notes, for your drafts, for things you're planning to publish, for things that you're sharing, and containing all of the things that feed into that writing. Saved articles, your reading history sometimes. And we found a couple things. One, AI applied to your own writing is such a helpful way to search and organize it. Because now instead of a keyword search, you have this free-form way to say, like, oh, you know, I've got all these deadlines coming up. Can you summarize all my notes about this upcoming project and list out the deadlines in date order? You can just tell an AI that now. But we've only been able to look at it through the lens of the entire internet. ChatGPT will do that for everything that's on the internet. We haven't really seen a lot of that applied to everything that you've ever written. And there's a lesson I had learned in the past that comes kind of from the world of productivity. I spent a big chunk of time there, which is productivity nerds are always making complicated systems that are fragile and break down and the ideal system allows you to be messy and I think that's what AI lets you do there. And then there's this other piece of AI as the writing assistant. Our view there is that actually you can kind of put the AI to the side and what it'll allow you to do is stay in your train of thought. So this is not AI replacing you. This is actually AI letting you be more human, like get your thoughts onto paper, and then when you're coming back around for the edit pass, AI has a lot of really smart suggestions alongside of it. So that's what we've figured out as we start to actually fold the current state of AI into helping people write more.
Okay, so let's talk about a couple of these things. So first of all, AI allowing you to be messy. Am I reading it right that for these old productivity tools you'd have to really use their systems and yeah, I'm thinking about Roam. With Roam I tried to use that, it broke my brain, it was so complicated to use. It was like you have to make headings and subheads and subheads of subheads and subheads of different children and parent things and some people whose brain works that way could use it, I couldn't use it. And so what you're saying is what AI will allow you to do on the productivity end is maybe just dump all your notes and writing and emails into one thing and then you can just query it and it will be able to handle that unstructured data.
Well, yeah, there was a word we were using which is not the app name but we've been using the word bucket for a while, which is just the idea of what does it take for you to feel safe to think like, this is a bucket that I can throw anything into it and I don't have to spend time organizing it and I know I'll be able to get it out when I need to. Like what's required to design an app to do that? And I think it's possible now. And before it was just like you would have to, goodness gracious, the amount of work that went into building, it's called building a second brain.
It's like your entire first brain is being used to build this second brain.
I'm sure it worked well for some people. Not for me.
It did. And if you're extremely motivated. I would say I think the second brain concept is mainstream, but the current implementation is not because it's too hard.
Right. Sorry to all the productivity folks that I've offended with that, but having a better tool and I think you're right, generative AI could enable that type of thing being built. I'm looking forward to using yours. And then the writing side of things, this will enable the writer to basically have an AI editor. It's funny because I use that now with ChatGPT. I'll drop my drafts in, my interviews in, and I'll ask, what did I miss? Did I capture the tone of the interview faithfully? And so you're going to build something purpose-built for that.
Yes. And some of the things you've seen already are good and we'll include those. But we have some unique takes on it that I think reflect that we are actual writers and we've written a lot over the years and we have a deep understanding of all the writing processes from idea to publication.
Okay. So how AI changes the future of writing. We've talked a little bit about the production of writing and why that is important there. The other side of it is, since we're here with the CEO of Medium, is how a public publication or a platform handles this writing. Now you've talked a little bit about how you don't want low-quality AI content on Medium and we'll treat it as spam. But also in that post where you talked about how AI is becoming more useful, you clearly say we're going to limit the distribution of AI generated content and you'll say no to AI generated content in your partner program. But then you had something interesting that you said after that headline about not including AI generated content in your partner program, which I think pays some humans to write on Medium. You wrote, 'Even if ChatGPT could generate a perfectly valuable story, we still want our partner program to incentivize human storytellers.' And that 'even if' is interesting because you really went from a place where you thought that there was no use here to now acknowledging the fact that ChatGPT could write a valuable story. So, and we both see this stuff improving and making more useful content. So what does a world look like where ChatGPT could write as well? Again, I'll go back to it. Forget, like we both agree there's a virtue to writing, but forget the production side of it. Now I want to talk about the publication side of it to you.
I wonder if you have like a go-to good research example. The one I've been using is I spend most of my time a little bit north of New York City and I have a go-to research prompt, which is go find events in my area in the next two weeks. And then I go on to say specifically look for these sporting events. I'm in the boonies, so my sporting events are minor league baseball, minor league soccer, and roller derby. But that's just what I want to know, are they in town? What's going on at the county fairs? What's going on at art openings? What's going on at the movies? Like just find all of that and give it to me. And so this is a customized event calendar for me essentially that is probably relevant to thousands of people. So why wouldn't I publish that, right? I think that would be for me an example of valuable AI generated writing. The human element is my own taste in the prompt. And then the result is fairly factual. And let's be completely honest, it would be rude to publish that without at least double-checking that it was factually true, because I've many times thought I was getting good research out of one of these AI tools and then only to find that it had hallucinated the whole thing. But let's say the human effort is in the prompt and then in the fact-checking, and otherwise that's AI written top to bottom. Yeah, I think that should be published and I think someone publishing it would be doing a service. What we're saying in this really kind of niche part of Medium, which is 80% of what gets published on Medium is by people that are just using the tool and the Medium network for distribution. The vast majority of the internet is non-professional writers, like just people who are looking for a way to communicate with the world. And that's what I, that's like the part of Medium, the part of blogging that got me excited in the first place. And a lot of what they're communicating is just a life lesson. This thing happened to me and I want to share it with you. And those life lessons are not in the AI training sets yet, right? And those life lessons, if they're not published, won't show up in the AI research because that is just scouring the internet. So where are we going? How are we going to get people to continue sharing the lessons of their life? There has to be some incentive systems, and to date it has been Google, right? The idea that you can just write something on the internet and traffic will just show up out of the blue. That's kind of like Google made the public internet. And so a question that I wonder is, is Google taking that away? If they take the incentives away, is the public internet going to go away? And the small part that we play is that there's a section of Medium where we pay the writers and we pay them essentially to make our subscription business work, to get some of the best writing, not all of it, but some of it, behind the subscription, and that's how we've made a business. And so that's what we're saying is our partner program. For them, we were just taking a stand and saying the thing that we want to pay for to make sure it exists on the internet is your real life lessons. And this idea of paying you to spend 15 seconds generating a ChatGPT transcript just doesn't even make sense to us as a worthy thing to pay for. But that's different than saying whether or not it's valuable. We're saying we're paying to make something exist on the internet that otherwise wouldn't exist.
I see. So what does the web look like if we're going to have a lot of writing
that's generated by AI, like thinking about your event calendar, sitting alongside and probably outnumbering human content, human-generated content.
Yeah. If it doesn't already, probably does already. But the quality question—it will keep getting better though.
My head originally went straight to relationships. Are we going to get tired of having relationships with each other? Are we going to prefer an AI for a relationship too? And of course that's the weird edge of AI right now. You have AI girlfriend, AI boyfriend. The degree to which that shifts is one of the other reasons that people write and share online—because we're trying to create connections with each other. I don't think that will ever disappear. I hope it never disappears. I still like connecting with people, so I think that'll continue to drive a lot of activity. But the question I've been posing to friends and advisers is: if Google Gemini reduces search traffic enough that there isn't an incentive to post in public, will people retreat into private spaces? And the example I usually give, because we already see this happening, is that we think there was a Twitter exodus. Most people think it went to Blue Sky and Threads and originally a little bit to Mastodon. I would make the case that a lot of it went to Discord, and that's an example of the public internet retreating into private spaces where it's safer to be yourself and to be weird.
I've never seen a Discord that didn't make my head explode.
Well, you should join. We have one for Big Technology subscribers. I think it's excellent. I've seen exactly the same thing that you've seen—a lot of, I could go weeks now without tweeting, or a week, let me be realistic. But I'm in the Discord all the time and it is less noisy. It's friendlier. Not all of them are, but it is friendlier than Twitter, which—easy to be friendlier than Twitter. And it's more information-dense and less madness-inducing because you don't have that home tab driving you nuts.
All right, you've convinced me. Smart people join the Big Technology Discord. You heard it here first.
I'm not worried about the end of civilization or something. I just think the internet is shifting a lot right now. And that's probably one of the shifts that will happen—if the incentives for public discussion disappear, you'll see people shift into different ways to get the same thing that they had been getting.
