Subscribe here: Apple Podcasts | Spotify | YouTube

What happens when the majority of content on the internet tips over into AI slop? On this episode of Galaxy Brain, Charlie Warzel talks to Max Spero, a co-founder of Pangram, an AI-detection company. They discuss how AI-detection tools work and how effective they can be at identifying what’s made by humans and what comes from a chatbot. They explore the cultural concerns around authenticity in the large-language-model era, and whether detection can keep up as models improve. The pair discuss how the speed of AI development and synthetic content threatens to degrade the quality of human writing and pollute the internet—and what, if anything, can be done to stop it.

The following is a transcript of the episode:

[Music]

Charlie Warzel: I’m Charlie Warzel, and this is Galaxy Brain, a show where today we are going to talk about the flood of AI writing online, and a tool that claims to help us determine what’s human and what’s been written by a chatbot.

One of the big fears of the generative-AI moment is that we are entering this kind of slopocalypse. Chatbots, audio, video, and photo-creation tools—they all make it extremely easy for anyone to churn out synthetic content quickly. Especially text.

By now you know the story: Students are using ChatGPT and other tools to write their essays for them. A rash of hastily self-published books on Amazon are clearly not real. Search-engine-optimization marketers and content farmers are flooding the internet with articles written by chatbots in order to game Google and make a quick buck. The web is just shuddering under the weight of all this slop. And there’s some research that suggests that half of all new articles generated online are coming out of large language models.

There’s all kinds of issues with this, right? A lot of this content is forgettable, soulless pablum. It drowns out quality, human-made stuff. But one of the bigger problems is social. On an internet where it feels like every other person is passing off generated-AI text as their own, this distrust forms. Some journalistic outlets have embraced chatbot-assisted stories, while others, like us at The Atlantic, have strict rules about AI use and pride ourselves on the very human craft of writing. For example, in March, the publisher Hachette cancelled a horror novel called Shy Girl for suspected AI use. Now, the author denied using AI in a statement to The New York Times and suggested that an editor inserted AI writing into a draft. But, semantics aside, there are fears about the industry’s ability to keep up and detect AI-generated text inside manuscripts. But the reputational risk isn’t limited to book publishing.

In 2023, Max Spero co-founded Pangram. It’s an AI-detection company that uses machine learning to try to distinguish human writing from AI. Pangram is being used by publishers, academic institutions, and other places to try to ferret out suspicious work. The company boasts of having a false-positive rate of one in 10,000—meaning one in 10,000 times it identifies human work as AI-assisted or generated writing. Pangram has been at the center of unveiling a lot of AI-generated work. When suspicions began to swirl about Shy Girl, Spero ran the text of the novel through his software and declared that it was 78 percent AI-generated. He posted that research to his X account and spoke to The New York Times about it.

The company now has a browser tool that lets people scan Reddit, LinkedIn, X, Substack, and other social-media posts to see what writing is human. And it’s already made news.

Wired recently reported that, according to Pangram, an April thread from the Pope’s X account—ironically, a thread warning about the dangers of AI—was, itself, likely written with the assistance of AI. The Vatican didn’t respond to Wired’s request for comment.

And recently the technology journalist Taylor Lorenz used the tool to scan thousands of posts on Substack from popular newsletters. Lorenz found some newsletters in the top rankings were “publishing 100% AI-generated content, according to Pangram, seemingly with no human editing whatsoever.” Online, you can see the contours of an AI-detection arms race forming; coders are already building tools to introduce errors into AI writing or to strip out AI conventions in order to make it seem more human.

Spero jokingly refers to himself as a “slop janitor,” and he says that his mission with Pangram is to increase transparency and help keep the internet from becoming less and less human. But the stakes are high—Pangram is quite good, but it isn’t perfect. And AI detection is getting harder every day, as the models get better and as the internet that it has to train on becomes more and more synthetic.

So how does Pangram work? How can we distinguish between the more convincing AI text and the real stuff? Are humans starting to write more like AI now? And is all of this detection helping us preserve our humanity, or is it making us more paranoid? I brought Spero on to talk through all of it.

[Music]

Warzel: Max, welcome to Galaxy Brain.

Max Spero: Hey Charlie, thanks for having me.

Warzel: So let’s talk about Pangram. It is AI-writing-detection software; something that, before I was aware of you, I had heard of it. The first, I think, time it really came up for me was hearing about educators and people like that using it to run papers from students through, and things like that. But then I started seeing it in a lot of different places. Specifically, most recently, I was talking to a book agent who was telling me that, like, everyone is leaning at publishing houses, really hard, on it. And basically anything they can latch on to that may be able to detect: Is this human-written prose? Is this AI-generated prose? Is it assisted in some way, sort of in the middle ground there?

But I wanted to start, before of all that—why do we need AI-detection software? Like, why do we track this stuff? What is the need to know this, in your mind?

Spero: Basically, I think AI enables just such an incredible, large volume of content. Without any way to differentiate between human- and AI-generated content, we just lose any semblance of signal-to-noise ratio, right? It’s like, if I’m just like reading texts on the internet, you know, it takes me one minute or five minutes to compose a tweet.

Well, in that time, a large language model can pump out hundreds of thousands of tweets. Obviously there’s some onus on platforms to help also filter out like AI-generated or spammy content and get you to like the real stuff. But I kind of believe that it’s going to be increasingly difficult for platforms to deal with this deluge of AI content, which is why we’re building Pangram.

Warzel: Well, and also, with the platform part of it—it’s one thing if it’s spam. If it’s just, you know, absolute garbage, it’s very clear that this is like, No human is going to want to post this.

But so much of what comes out, it’s not just that it’s believable human writing. It’s also like—people are using the large language model–generated text to say very normal things, right? They’re using it in Twitter posts; they’re using it in LinkedIn posts to announce their jobs. They’re using it to comment on whatever. I have seen many horror stories of people accidentally leaving in the ChatGPT “Would you like me to make this funnier”—or whatever your adjective—“for you?” on Tinder profiles.

Spero: Incredible.

Warzel: People are using this stuff in ways that isn’t spam. They’re using it in their actual “I wanna communicate with human beings,” in life. And I feel like that is also just a very big shift to why the platforms are not going to be able to go after this, right? Because if you do want to crack down on it, you’re going to be cracking down on people earnestly trying to communicate with other people.

Spero: Well, personally, I think we’re in the middle of a really big culture shift. People really haven’t settled on like what is acceptable or not. I think there’s a lot of people who really get the ick when they read AI-generated writing. Like, you can tell this sort of prose and the style, like “This came from ChatGPT.” And so it feels to me, like when I’m reading something that came from ChatGPT, I don’t love that. It’s also just an authenticity thing. If I’m sure, like I guess if I’m on Tinder, I see someone with a like AI-generated, kind of long-prosy profile, like whatever. What about the messages? Do they have their, like, OpenClaw, you know, swiping and, you know, messaging to try and get a first date? And then they show up for real. And then we talk, and they have no idea like what the conversation was like. Or they, you know, skimmed it ahead of time. Like, I think as soon as you let in a little bit of AI-generated content, you open the door for like this truly inauthentic behavior.

Warzel: You all participated in a study or a survey, I think it was, where you guys note that in 2025, 35 percent of the newly published websites on the open internet were AI generated or AI assisted. And the next part is really interesting to me too, which speaks to what you said. Which is internet users are overwhelmingly cynical about this; 75 percent of people polled felt like an AI-dominated internet would be less accurate. And 83 believe that AI will collapse unique writing styles into monoculture. Is that ultimately what your fears are, or your reason for taking this up? This feeling that there’s something that needs to be protected here?

Spero: I think as humans we’ve built so many spaces on the internet around trust and, you know, believing that there’s a real person with real effort on the other side. And then AI kind of turns this all on its head in a way where, yeah—we could just have somebody flood the internet with misinformation. There are hallucinations that LLMs repeatedly make, and if we bake it in to the fabric of the internet, then at some point we’re not gonna be able to work our way out of it anymore.

And I do think the collapse of writing styles is one of the biggest problems. Especially among writers or people who talk to ChatGPT a lot, I hear this fear from them as well. They’re like, “I don’t wanna talk to Claude too much, because I’m afraid that I’m gonna start to adopt his writing style.” Which is crazy to me.

Warzel: Yeah. It’s very interesting to me, as someone who is a writer. I find it very fascinating because I understand why people want to automate the drudgery. I don’t understand why, at this point, why people want to automate the actual expression part, right? Like, I don’t know why people want to automate the thinking—the part of this that is actually creative. Especially when it comes to text, I feel like there’s so much fun that you can have playing around with language. And just watching that creativity leech out, to me, feels obviously extremely troubling, as a writer myself.

But also just—I don’t know. I feel like you get the most thinking done when you are actually trying to craft something like this. When you struggle a little bit.

Spero: When you struggle a little bit, you don’t have an LLM telling you exactly, “Hey, this is the next token; it’s the next word.” I think what we’re seeing in the writing industry is the drop shipping–fication of writing. People see it as a get-rich-quick scheme. I see there’s YouTube tutorials online of “This is how you make $1,000 a month publishing AI-generated books.” And they walk people through these full AI workflows.

I don’t think they’re actually real. I don’t think people are making tons of money off of this. But people are making money selling courses and telling other people to go pollute the space.

Warzel: I want to talk about how Pangram works. I have gone onto your “how Pangram detects AI-generated content” page. And there’s a lot of stuff here. I’d love for you to walk us all through it. Broadly, how does this detection stuff work? And then I guess we can get into the specifics of it.

Spero: Sure; yeah. I can maybe cover two things. One is like: How is the machine-learning model trained? And then two is like: What is it picking up on?

So I think the first part—how is it trained—is pretty interesting. I think the way Pangram works is pretty different than basically any other AI detector out there. What we’re doing is we are taking millions of human-written documents, and we’re actually making calls out to the large language models, the AI models, and asking for a synthetic mirror. That’s what we call it.

So, for example, I have a 500-word essay on Moby Dick written by a seventh-grader. Then I’ll ask Claude: “Hey, also write me a 500-word essay in the same style.” And then, so we have two documents. One is human written; one is AI. And then we’re training this large machine-learning model contrastively, to learn the difference between the human document and the AI document.

Warzel: Is that all happening in the moment, like when I’m running the thing? Or is this just how it was trained?

Spero: No, so this happens during the training process. So we’re training this model. It sees millions of documents. And then, at the very end, it’s a fully trained classifier. So any new document, it’s not putting it into ChatGPT, but it already knows the stylistic tells. And so the way I kind of like to think about it is: When you’re writing a document, every word, every token, every sentence is a decision that you’re making. So, over the course of a document, you are basically following a single path on this very large, very wide decision tree. And the longer the document is, the wider the decision tree gets.

And something interesting about large language models is they actually tend to make the same choices over and over. Experts call this mode collapse. But so what it looks like, in terms of documents—if I’m saying, like, essays about Moby Dick, all the possible essays about Moby Dick that LLMs make—it’s a much narrower decision tree than all the possible essays about Moby Dick that any human would make.

And so we’re learning to understand what these very narrow decision trees are. And how then, when we look at a new document, we’re looking at how many of these decisions align.

Warzel: It’s a little bit like when you and I are all prepped for a podcast conversation, and I’ll have this whole, like, “This is where it needs to go.” And then you’ll get another person in there, who’s another human, and it gets really messy, right? You go on these tributaries, and these things. It doesn’t unfold the way.

And so I guess what you’re saying is like: If I was behaving like a language model, I would just constantly be routing it toward a center. Right? You wouldn’t go on those different tangents and do that kind of thing. That’s a very—

Spero: Exactly. Yeah. You would kind of follow the same path. If we took you and ChatGPT and put you together in a conversation, you’d follow more or less the same path a hundred times, versus you and me. There’s a lot more chaos. We’d have a broader range of conversations.

Warzel: Okay. So then once you have the model, and you have the mirrored prompts, you then retrain the model, right? Can you talk to me about that process of retraining?

Spero: Once we have the first model, it’s usually pretty accurate. It can get to 99, 99.5 percent. But then we do something called active learning. So we take this model that we’ve trained, and we have it search over a large corpus of data—and find it goes and finds examples that it’s uncertain or incorrect about. And then we take these examples and bring them back into our training set to do a second training run.

And the reason we do this is: These uncertain examples are more valuable in the training process, because they’re closer to the boundary between human and AI, or at least the perceived boundary that our model sees. And so when we bring these in, our model is able to learn more effectively and have a much lower false-positive rate.

Warzel: So let me see if I’m getting this right. You essentially, you have the model. And then the model will have, by nature, some false positives, some false negatives, right? That’s like—

Spero: Yeah, correct.

Warzel: And when you get those cases, and you’ve identified, “Okay, this is a false positive; we have flagged this as AI when it really is human”—you then run that through the machine and tell it, basically, “You made a mistake.” And then…

Spero: Yes. And just to clarify: So a false positive is, if we flag something that’s human written as AI generated. Then we got that wrong; call it a false positive. An average AI-detection model, like our first pass, will typically have around 1 percent or half a percent false-positive rate.

But then we go collect these examples. We bring them back, and we actually don’t continue training this model. Instead, we restart from scratch on a new model. But the new model has these additional difficult documents in its training set. And then it’s this new model that we retrain from scratch that is able to have a one-in-10,000 false-positive rate, which is much lower.

Warzel: How often are you going through that process where you’re starting from scratch and sort of rebuilding the model to make it more effective?

Spero: We retrain our model from scratch every three to six weeks. Part of the reason we have to do it so often is the field of AI is moving so fast. There’s always Claude Opus 4.7; that’s the new one. And then, now ChatGPT 5.5. They’re always a bit different. The decision trees slightly skew from ChatGPT 5.4. So we do have to collect more data and retrain.

Warzel: So let’s just go to this. Okay, you said false-positive rating of one in 10,000. This may come off to you like a dumb or obvious question, but the thing that just pops in my mind is like: How do you know? I’m sure there are instances where somebody is able to prove, “No; this was me, and I did it in here, and I can show you.” What happens in that sense? Is that just logged away, as like, “Okay; this is a really great example of a false positive that we can talk about, and we can use it when we work to retrain the models”? What do you do when one of those things is brought to your attention?

Spero: Yeah, so typically, we don’t train on these, but we keep them in an eval set. So when we’re training a new model, we’re able to look at our past false positives and see, “Hey, does our model improve on any of these?” And if so, that’s a good sign. But then we also still need to calculate our ultimate base rate. Because we don’t want to have some improvements here and then other regressions somewhere else. But yeah; we do collect them. We look at them and try to understand—“What is our model picking up here?” And use that to make the model better.

Warzel: I went through, and I fed a bunch of my stuff into Pangram, as I’m sure anyone who’s using it does to test, And it was like, “This is human stuff.” But if I had found a flag of, you know, something I’d worked on, I imagine I’d be frustrated or whatnot. Do you have those frustrated conversations with people when there are the false positives? Do you interact in that sense with them? And what do you say to people when that does happen?

Spero: Absolutely. I mean, like one in 10,000 is not zero. And then, you know, if this means that if a thousand people all put 10 documents through, then maybe one person is going to be like, “Hey, I put 10 documents through, and one of them came up as AI. So, you know, I think your false-positive rate is higher than you claim, because I only put 10 documents through.” But I don’t think they realize that on a population level, there are going to be these, like, anomalies. And so sometimes I just have to talk people through it and just explain that you really have to look large scale to fully understand it.

And sometimes I think we get other misconceptions about what a false positive is. Like sometimes we get somebody who writes, and they try to sound as AI as possible. They say, like, “As a large language model, I cannot do whatever.” And I’m like—well, of course, Pangram is going to say that’s AI, because you’re trying as hard as possible to make it sound like it’s AI. And so I have had these conversations as well.

Warzel: This gets at something I wanted to ask, which was: Have you guys found in some of these edge cases that there are people out there who write like AI? Who just kind of have that, like, “I have synthesized a lot of relatively boring Wikipedia, newspaper-style articles,” and are able to output sort of the mean when they sit down to write? Is there something different about, you know, that decision tree that we’re talking about, and that the software is really gonna just catch that most of the time?

Spero: Surprisingly, no. I think most humans, people, are just not as mode collapsed as LLMs are.

Warzel: Thank god for that.

Spero: Yeah, yeah, thank god. So when I go back and I look at the false positives, I might see a Yelp review from like 2018, and I’m like, Wow, this reads exactly like how ChatGPT writes Yelp reviews. But I have to remind myself that, yeah—this is just a single isolated incident. And we actually have gone in and done profile-level analyses. Like if we have a false positive on one case, do we also have false positives on their other pieces of writing? And we found no, not really. People are just diverse enough that we don’t really have this issue.

Warzel: That’s interesting. So, all right. One thing that you said, speaking in another interview that you gave, was that from all of this—and this gets to the decision-tree nature of how the models work—that you’re learning how frontier models make their decisions, right? This process of diving in to detect has taught you a lot about that.

Can you tell me a little bit more about what you have learned about some of these models, or how the models are progressing? Like, you are in there in the actual machine-learning weeds of it all. And I think to a lot of us, it’s hard to know. Yeah; this just gets better. Or like, you know, “Claude used to be thought of as a little more literary than ChatGPT was,” or whatnot. But I feel like you probably have some concrete insights into how these models are evolving. And I’m just curious what you’ve learned there.

Spero: Yeah; I’m happy to tell you some things that I’ve learned. I think this isn’t super well fleshed out, because this has just been our experience over time. But my experience is that since 2023, large language models have gained much stronger preferences. And I’ve seen personally that there’s a very high correlation between a model having strong preferences and a model just being highly capable in general. So I think this is something interesting where like the early ChatGPT would kind of like say whatever you want it to say. people would call four very like sycophantic and today’s large language models, they will push back. They will say like, no, you’re wrong. They, they have like pretty strong preferences. Like if you ask, for example, Claude to write an essay on like a topic that’s choosing, it’s often going to choose consciousness or it has a few favorite topics. And so I think a part of what we’ve learned is just like, how do we do prompting to still get like a very diverse data set, not just like 10,000 essays about consciousness by Claude.

Warzel: Why do you think that is so emergent? That idea of these models having a kind of subject-matter preference or something like that?

Spero: I do think having preferences is an important part of intelligence. If you just think of a human with no preferences, then they’re kind of just like a zombie. Have no agency. They’re not going to go out and do something. Like it, you must have preferences on like, “This is the best way to do it.” Something, right? To be able to do things right more often. And I think, like, this has been baked in explicitly through a lot of this reinforcement learning, in terms of coding and making these models stronger coders and more agentic in their harnesses, like Cloud Code and Codex. But I think this has the side effect of giving these models strong preferences in other fields, like writing, where they may have not had them before.

I think the least-preferenced models are like GPT-2 and GPT-3, where it’s—they’re truly trying to simulate the distribution of human text.

Warzel: What else are you all looking for—or I guess your models looking for—when they’re evaluating a text for AI? Like, what are some of the other tells that are in there?

Spero: This is an active area of research. I think, when you train a machine-learning model, like a classifier model, especially, it’s fairly black box. It will tell you what it thinks about the label, and it’ll tell you its confidence. But what it’s not going to tell you necessarily is, like: I think large language models would put a period here. And so because there’s a semicolon instead, this is likely human writing.

But I do think that is like a lot of what it picks up on, all of these like micro-decisions—where for a human, it might be a 60/40 decision, but for an AI, it’s really like a 90/10 decision. And so we look at just, like, the aggregate, cumulative effect of a bunch of these micro-decisions. I will say there’s some like really major ones that I pick up as a human looking at AI text. Like the “it’s not just this, it’s that” formulation, which I think a lot of people have picked up on now. I think older large language models really overuse, like, delve or testament, or just had way too many em dashes.

But I think the newer models, the tells are getting much more subtle. And so it’s getting harder to pull out really explicit, just like phrases or words that they overuse.

Warzel: Something you said there that’s really interesting is that you’re working with models that at the same time have in the reasoning component of it, or the justification component for the decision, a black-box element. Does that, for lack of a better … does that freak you out a little bit?

Spero: Not really. I think the way that people seem to be adopting technology like Pangram is very human in the loop. It’s not just like a professor is going to go, “Pangram said your paper was AI, so I’m just going to fail you.” But instead…

Warzel: I mean, how do you know? I guess, right?

Spero: I think they’re starting with a Pangram result where it says, Hey, this looks like it’s AI. And then they go in; they look at the citations. Or like, you know, does this match the student’s prior work? Especially in the academic-integrity councils. But also I think, more generally, I’m talking with publishers too. And it’s kind of the same thing. Of like, how do you use Pangram as a starting point, for opening up this conversation about, “Did you use AI? How did you use AI? We need to disclose this responsibly if you did, and go from there.”

The best AI models don’t really display their reasoning in a way that generative AI does. Even if you ask ChatGPT, “Why did you do that?”—it’s just going to give you text that sounds plausible. It’s not going to actually be looking into its weights and truly understanding what it did.

Warzel: Does it worry you that there are people out there who will adopt the software and not put a human in the loop? And, you know, because I can see in a lot of senses, like if I’m a professor, right, and I want to use this stuff. And like, the responsible way to do that is to put the human in the loop or whatnot, right?

But I think one thing we see with—and I’ve heard a lot of it anecdotally from people in higher education—of like, first, it started out with all these professors and people being really mad that the students were shirking their duties of writing, doing the actual writing work. And just submitting, you know, auto-generated stuff.

And then I noticed some of those same people, a couple of months later, just started like grading, you know, using models and things like that. And that side of it. And so, you know, in a perfect world, yes: Everyone’s using this to start a conversation. To flag something that they think is suspicious. To go in and then do the work.

Does it worry you that there are some people that are gonna be out there that are just gonna be like, “Okay, I see that it says pretty high confidence that that’s AI.” Like, no questions asked, the student has failed on the paper, or that there is some kind of consequence. Do you worry about that happening?

Spero: I think I typically have a lot of faith in people and in humanity, and people making sure that they are using technology responsibly. That’s being too optimistic. So I think there are good ways to use this tool automatically without humans in the loop. If say, let’s think about the problem of, like, X—where they want to stop AI-generated replies.

Warzel: On X, yeah.

Spero: Right. So if yes, on X—so if a reply, they check with Pangram, looks AI generated, they can maybe reduce the visibility of it. And then if they look at an account, and of 10 replies, nine or more are coming back as AI generated? Then they might shadow ban or ban the account, because it looks like it’s a reply bot.

And so I think like things like this makes sense. And then, of course, where the human in the loop comes in is they have an appeals process. If we go back to some of the other, like, higher-stakes cases, like publishers and academic integrity, we have been working really closely with the academic-integrity groups and universities. And I think everybody I’m talking to is putting together really responsible policies.

Warzel: Are you worried at all that so many people using these tools—publishing, you know, AI-assisted writing or stuff that’s been sort of muddied by the use of large language models—does that interfere with the data set? Is it going to be hard to tell after a while, when people are using all these tools, what is the human writing, versus what is not, if the writing kind of starts to fuse?

Spero: Part of training a great machine-learning model is having really clean labels. And today, I don’t think we can really trust internet-crawl data to have clean labels in a post-ChatGPT world. There’s just going to be some latent AI level in all of the text that we find. But I am worried also about, like, we want to have modern text in our training examples. Especially as language changes, even over the course of like three years, I think slang and, you know, there’s modern slang that didn’t exist before.

Warzel: Yeah. We’re all looksmaxxing now, man.

Spero: Exactly, exactly. So this is a very difficult problem. It’s an active area of research, but I think the very first step, for us, is collecting really clean human-written data from 2026. And so we have a couple initiatives for this. I think one part is just looking at people who we know are trusted writers. Somebody who’s been blogging for a long time; their text hasn’t come back as AI. And then also looking at younger generations, trying to get writing samples from students and kids who have grown up with ChatGPT, and very likely have had their writing and thinking influenced by AI. And the first step here is just to measure whether this impacts our false-positive rate at all. And I think the next step is to work on our training process.

Warzel: Well, what’s wild too is, you know, it’s not even at this point, people who are like, “Let me feed this into ChatGPT.” It’s like—I try to disable these things on my own Google docs, but it still wants to auto-complete every other sentence of mine, right? Just like, as I’m writing an article, it wants to do that. And you have to, you know, like get rid of that.

But there’s tons of people, I’m sure, who are actually embarking on trying to write themselves. And they’re like, Oh yeah, I’ll just, you know, I’ll let it finish that sentence for me. Right? Like, that looks like a good thing. You hit enter, and you’re done. And then that’s technically AI-assisted text, even though it’s tough.

Spero: It’s everywhere. It’s pervasive. I’m not a big fan of that. I think it is creating this writing monoculture, and that’s why personally I don’t have ChatGPT help with my writing at all.

Warzel: You all announced a Chrome browser extension that sort of allows you to passively turn on this detection for certain things. Like LinkedIn, X, Reddit…

Spero: Substack and Medium.

Warzel: Substack, Medium, yeah. And I got access to this a little bit. So let it run in the background as you’re scrolling, then you can basically take a … it shows the number of posts scanned. Human, AI, AI assisted. And gives you sort of a grade on your feed, right? Like is it 80 percent AI?

Spero: What’s your feed?

Warzel: Honestly, okay, so here’s the thing. Not that I was hoping that it was gonna be like a nightmare. But I’m at like 85 percent [human], I think I want to say. It’s not as much slop as I would have thought. And the stuff that is, is like not super interesting. Same on X. Like I was hoping for just, Oh, this person is insufferable. I can’t wait to see. And it’s like, no, it seems like there’s a lot more humanity in there.

Spero: That’s good.

Warzel: I did notice when I turned it to the For You page on X, the numbers, the ratio changed significantly. Because I think I do tend to follow, as a journalist, other journalists or people who are—

Spero: You’re following real people.

Warzel: I try. But what was the impetus for that? For this idea of, “Let’s see it in real time”? Is it more of a—and I don’t mean gimmick negatively—but is it more of like a proof-of-concept thing? Like, I want people to be able to sort of see the impact of this broadly across the ecosystem in real time? Versus, like, there is huge utility to being able to, like, call this out as I see it.

Spero: Well, I think, to me, right now, I’m looking at the trend lines. So if you asked me a year or two ago, like if I see an AI-generated Reddit post that’s super viral, I’d be like: Wow, wait, that’s cool. You know, It’s happening. And now I’m like, I’ll roll my eyes. Because, you know, I only have to scroll for like a couple of minutes to see it.

So even if it’s still like, you know, 85, 90 percent of my feed is human—that 10 to 15 percent AI? That didn’t used to be there. And I think if we look back a year from now, if we’re not actively curating our feeds and we’re just letting the algorithm take over, then that 10 to 15 percent is going to be 40 to 50 percent And so I think this is really a proactive tool before the internet gets bad and more full of slop.

Warzel: So this is the way that you would like someone to use it if they choose to sign up for this and do it. Would you like to see people be like calling that out? Like, if you see, you know, something in your feed that is not advertised as [AI], and it is “slop,” what is the behavior you’re hoping to get from it? Or is it just simply raising awareness? Of like, “Okay, it’s the Wild West out there right now”?

Spero: It’s funny, we actually started out before we did the Chrome extension. We just had a little Twitter bot where you could tag at Pangram Labs, “Is this AI?” And then it will respond; it’ll scan the parent posts and then respond with a verdict. They found it was a fun game. They were using it to call out these big slop accounts. They knew these slop accounts were posting slop, but they just didn’t have any way to, like, prove it or call it out before Pangram.

And so I think that’s one use case. But the one I’m more excited about is just the passive and manual curation of your own feed. If I see something that’s AI generated, I’m going to mute or block the account. I’m going to choose not to engage. I’m not going to spend my time reading it. I’m not going to spend my time commenting or arguing against their points. And I think overall, this is just going to give AI-generated content a lower reach than something that was actually authored by a human. And that’s my overall goal.

Warzel: Are you worried that this could become a weapon of sorts? I’m of two minds on this, completely. Because on one side of it, there’s so many reasons why I don’t want the slop in my feeds. I also am a 20-year veteran of Twitter. And I know that, like, these tools can very easily be turned into screenshotting stuff. Do you worry about them getting like caught up in the culture war at all?

Spero: Yeah. This might sound weird coming from me, but I’m really not a fan of call-out culture. That’s never been like the primary goal of mine. My hope is that it’s overall a tool for good, right? That’s why we give you this feed-health score. So I mean, maybe some people are gonna go … like, they see something AI-generated, and then they’re gonna go interact with it more, and then screenshot and call them out. But that’s gonna kill their feed-health score. If you want a high feed-health score, you should just non-engage. Mute and move on.

Warzel: You did, however, engage—or Pangram engaged—with the pope. You took shots at the pope.

Spero: You talking about the Wired article recently? Yeah.

Warzel: Yeah, there was a was a Wired article of using this detection tool, and seeing that there were a number of tweets from the Pontifex account. Which is the pope’s official, you know—it passes to all of the different popes that we’ve had since we’ve had popes using Twitter. And there are some concerned statements. The pope has a lot of deep concerning thoughts about AI and synthetic content and things like that. And the lack of humanity. And at least one of them, according to your detection software, as Wired reported, seemingly written, assisted by AI, in some capacity.

Spero: Yeah, I scrolled through the pope account. I think among like 15 tweets, five of them came back as AI. So like my kind of understanding and verdict here is like: the pope’s not writing his own tweets.

Warzel: Yeah.

Spero: He has a social-media team, and his team is probably, you know, AI enabled. They’re probably savvy, Gen Z; they’re using AI to some degree. They probably didn’t think that they’re gonna get caught, or that people are going to be able to notice. I don’t think it should be a huge criticism, like, My god, you know, the pope is a huge hypocrite. I think it’s more just like cultural commentary: It’s really interesting. Like, everyone’s using AI now to some degree. And so I think this is also kind of a direction where we’re really interested, in terms of research, is better understanding the degree of AI assistance.

So right now it’s not binary, but it’s ternary. So we’ll say something is human, AI, or AI assisted. And I really just want to like expand and fully understand the scale of AI assistance. Because I think this is where the world is heading. There’s no shame in asking ChatGPT for a little bit of help, but there is shame in asking ChatGPT to generate a whole novel or post for you.

Warzel: What would you like—in that spectrum of “AI assisted”—to be able to see? And how do you think you could get there?

Spero: So I would … like, if I’m going to use AI to translate something, or I’m going to use AI to help me make it more formal or help me fix my wording, we can catch that. And we can also say, “We don’t believe this is fully AI generated, but we believe this is AI assisted.”

Whereas I think the part that I really want to make sure we’re catching and separating from the rest is the things that are just—here is a research paper on archive. To ChatGPT, like: Please make a viral Twitter thread and explain the main points. And then it does that. And I want to be able to say, “That’s AI generated.” Like, I don’t think Pangram is quite there yet. I think a lot of times if I take a tweet or article, and I ask ChatGPT to rewrite it to make it better, then it might come back as AI assisted. It might come back as AI generated. I think today we don’t have the greatest understanding of even measuring how much ChatGPT changed my initial text.

Warzel: I’m really curious, given your involvement in this space. The last few years of generative AI and this agentic stuff—how do you feel about it? Does this stuff excite you, this AI moment that we’re in?

Spero: I do think we’re in a really exciting AI moment. I think AI capabilities have gone through the roof in the last couple of years. We’ve gone from AI being like, It’s a silly chatbot. It kind of sounds to me like this is actually doing real economic work.

But I think it’s also made me increasingly aware that there’s a lot of harms that are going to come from AI, and it feels like nobody’s really working to mitigate them. Like, we’re doing some of it, which is we’re helping mitigate the proliferation of AI slop and AI spam. And we’re giving people tools to, like, help understand this and then also address it at a larger scale.

But I think there’s a lot more than that. All the AI CEOs are warning about job loss. And I don’t think they’re saying that because it’s good for fundraising. I think they’re saying it because they actually believe it. And then there’s also people on the other side that are talking about like the great environmental harms. Yeah: like huge data-center buildouts and the electricity and power cost of running these huge language models at scale. And these big agentic loops.

Warzel: Yeah. Well, I ask because what’s interesting to me is that—you’re someone who plays in these fields. You can build these things; you interact with these things; this is your industry. And then you would have to be excited about the advances of artificial intelligence to do it.

Spero: I absolutely am. Yeah.

Warzel: And yet, at the same time, what I think is so fascinating is that this is a reactionary, like a bulwark-style force against this too, right? Like, not against progress necessarily, but against, as you sort of said, the misuse of it. And I think the speed of it to me is what it feels like it’s a bulwark against. Like the problem—

Spero: Yes. Yeah, we’re a little bit of a speed bump to this oncoming tsunami. It sometimes feels like.

Warzel: But I find this to be the problem. So much of the conversation, the discourse gets thrown between these two. Is it a stochastic parrot, or is it mostly sentient? And it’s not a helpful discussion of what is actually happening on the ground.

But the thing to me that’s terrifying about this moment is the speed, right? You have these new models coming out. Like, I can step away for a long weekend, not look at the internet, and come back and genuinely be like, I need to take like five hours to figure out what’s happened to this industry that I’m ostensibly supposed to be monitoring at a given time. And there’s something about that that is like … you know, some of that is PR, some of that’s hype, some of that is actual iteration. It’s actually what’s happening. It’s how people are using it, how the culture is changing around this. And I think the speed is what is really terrifying. I mean, are you kind of terrified by the speed of all this?

Spero: Yeah. This is all happening so fast, and the world is changing so fast, and a lot of it still feels to me like we’re not ready. If you ask me, like, what am I doing? Why am I here? It’s like I’m really trying to shape the world in the direction that I want to see it go. Like I want to see people using AI to cure cancer and, you know, make senior care easier and make all of our lives better. And I also don’t want to see AI polluting the internet. And making public spaces, public internet forums becoming completely unusable and allowing cheaters and people who want to just not do their job to use AI to replace that. So sort of like: There’s these two sides, and I want to see the good side of AI flourish, and I want to help mitigate the harmful effects of AI as much as possible.

Warzel: What do you think happens next to the internet? I mean, what we see now that’s not even hypothetical is this, you know, influx of slop. So much synthetic content, perhaps, you know, by some estimates tipping over into over 50 percent of the internet being synthetic. You know, the idea that basic things like Google are are less useful. All these different websites where user-generated content is supposed to be paramount. It’s less useful. All the SEO hacking. All the stuff that takes these things that I think all of us love, or at least found great utility in. All of this, like the internet itself, just feels like it’s really at risk right now. Where do you see this all going?

Spero: We’re still very nascent, but I think we are in the birth of this adversarial industry. So I think what’s going to happen is the problem is going to get worse, everywhere. So we’re going to see more slop, more AI agents, people using AI and basically turning compute into either influence, political influence, narrative influence. And then I think eventually, we’re gonna see people utilizing more tools like Pangram to help stop and mitigate this.

But I think maybe a good comparison is like very, like early computers. Where, you know, viruses were starting to become a thing. And then eventually there were anti-viruses and cybersecurity companies, and there was like this adversarial back and forth. I think we’re at the very beginning. This is the ground zero of that today.

Warzel: But just to push a little bit on that. I think that that’s probably true. But just from, again, the speed and the humanity-centric thing. There’s just something, to me, that is worried about the speed at which the humanity is being leached out of this. Do you worry a little? Like we could lose this fight, a little bit?

Spero: I mean, I think, yeah—I think there’s a very real risk that things could go wrong. In my view, things going wrong is not the same as like most people in my industry think. A lot of people, when they talk about things going wrong, they think of, like, AI doom: Claude becomes super-powerful and then takes over the world and kills all of humanity. But the internet has for the last 20, 30 years really operated as a high-trust society, and we are at risk of losing that.

Warzel: I hate to end on, like, the downer note on that. But I also thank you for coming on here, and also for trying to demystify some of this, right? I think it’s a really fascinating exercise. So thank you for all the time.

Spero: Of course, yeah. Thanks so much for having me.

[Music]

Warzel: That’s it for us here. Thank you again to my guest, Max Spero. If you liked what you saw here, new episodes of Galaxy Brain drop every Friday. You can subscribe on The Atlantic’s YouTube channel, or on Apple or Spotify or wherever it is that you get your podcasts. And if you want to support this work and the work of my fellow colleagues, you can subscribe to the publication at TheAtlantic.com/Listener. That’s TheAtlantic.com/Listener. Thanks so much, and I’ll see you on the internet.

This episode of Galaxy Brain was produced by Renee Klahr and engineered by Dave Grein. Our theme is by Rob Smierciak. Claudine Ebeid is the executive producer of Atlantic audio, and Andrea Valdez is our managing editor.