We're ‘dangerously close’ to dead internet theory according to Pangram's CEO

The internet has a trust problem, and it’s not just because social media feeds are filling up with AI slop. AI-generated text and images are now making their way into job applications, product reviews, and even insurance claims, leaving platforms and users alike scrambling to figure out what’s real. A handful of startups have cropped up in the past couple of years to become the “trust layer” the internet needs — including Pangram. The startup recently snapped up $9 million for its AI detection system and landed a partnership with Substack, which is now using Pangram’s tech to show readers which of their favorite authors use AI to write their newsletters. Pangram also recently dropped a new AI image detection tool. On this episode of TechCrunch's Equity podcast, Pangram co-founder and CEO Max Spero joins Rebecca Bellan to dig into the promise of AI detection tools and where to draw the line between AI assisted and AI generated. Listen to the full episode to hear more about: Why Spero thinks the internet could be “dangerously close” to the dead internet theory becoming reality within a few years. Why figuring out how much AI went into something is harder — and potentially more useful — than simply labeling it AI or human. The stakes of getting AI detection wrong, especially when false positives involve sensitive images. Why Spero thinks the “bottom tier” of writing jobs may be gone for good, while genuinely good human writing could become more valuable. Subscribe to Equity on YouTube, Apple Podcasts, Overcast, Spotify and all the casts. You also can follow Equity on X and Threads, at @EquityPod. Chapters: 00:00 Intro 1:15 Building Pangram: from binary detector to gauging AI assistance 2:06 Why job applications, reviews, and Amazon are drowning in slop 5:39 How Pangram actually spots AI writing 9:26 Where people draw the line on AI-assisted vs. AI-generated 12:50 The false positive problem, and "America has a Pangram problem" 14:30 Detecting fake images, now that AI photos are scarily good 19:13 Competing with GPTZero, Winston, and building "core infrastructure" 21:43 Will good writing jobs actually come back? 23:36 Outro Learn more about your ad choices. Visit megaphone.fm/adchoices

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  This episode of Equity is brought to you by Schneider Electric. Schneider Electric, your energy technology partner.

• 0:06Rebecca Bellan?

  Hello and welcome back to Equity Tech Crunch's flagship podcast about the business of startups. I'm Rebecca Bala and this is the episode where we bring on industry experts to help us explore a trend in the tech world and dive deep. Everyone has heard of the dead internet theory where we all just stop engaging online because it's so filled with bots and AI slop. But what if it doesn't have to be that way? A handful of startups have cropped up in the last couple of years to detect AI-generated writing and keep the internet human. But the prevalence of AI content doesn't just affect your social media feeds. It can result in all kinds of fraud, from job applications to customer reviews to insurance claims. Pangram is one company that has gotten a lot of attention lately for its role in trying to build the trust layer of the internet.

• 0:48Rebecca Bellan?

  So today, we're talking to Pangram's co-founder and CEO Max Spero about the promise of AI detection tools and where to draw the line between AI-assisted and AI-generated.

• 1:03Rebecca Bellan?

  Max, welcome to the show.

• 1:04Speaker 02

  Hey, thanks for having me, Rebecca.

• 1:06Rebecca Bellan?

  Yeah, really excited to talk to you again. So give us a little bit of the background of Pangram, which has been kind of blowing up as the new AI detection tool that a lot of people are using. This started two years ago. Give me a little bit of your backstory.

• 1:21Speaker 02

  Yeah, so we started the company, man, almost three years ago now.

• 1:25Rebecca Bellan?

  Oh, really?

• 1:26Speaker 02

  Just kind of like as a really interesting problem that we wanted to solve as machine learning researchers. We saw AI getting better. ChatGPT and GPT-4 had come out. We thought like we could try to tackle this problem, which seemed largely unsolved at the time. And I think since then, we've made a ton of progress. We've went from like something that's kind of binary of like, is there AI yes or no? To really more like mixed, trying to understand like, what is the degree of AI assistance versus human input?

• 1:58Rebecca Bellan?

  So you've talked about this being a trust layer on the internet, et cetera. So why do we need a trust layer? What are you seeing that is causing? I mean, it sounds like an obvious question, but what are you seeing that's causing a need for your kind of a business?

• 2:12Speaker 02

  Well, AI just sort of enables people to produce large volumes of text that seem plausible. And I think this actually sort of creates issues with a lot of societal structures. For example, like if you look at something like job applications where people write cover letters, maybe short answers, typically like if somebody's on a job hunt, they could apply to maybe a couple dozen jobs in a day. Now you could use an AI auto-applier to apply to a thousand jobs in a day. And so this creates this just like massive imbalance of effort where it's infinitely easy to produce content or take actions.

• 2:55Speaker 02

  And then the verifier or reviewer on the other side ends up being kind of swamped. But I think there's still a lot of fear that like any AI system looking at resumes could be biased.

• 3:07Rebecca Bellan?

  You know, from a user standpoint, right? We're seeing AI slop take over the internet. It's in our feeds. It's in academic papers. It's in politicians when they're presenting new legislation. And I think it's a, as users, we don't want to see that, right? As people who are like existing and cohabitating space on the internet, you're just like, oh, this is like a low effort slop that I don't really want to see. But like, aside from that, where is the, like, who's losing money off of this? Like, who's finding their businesses are being affected by AI slop?

• 3:45Speaker 02

  I think like there is this issue with like the social platforms where like AI slop is a good thing because it makes all the numbers go up. You know, it's engaging. People are posting more. People are liking more. But then I think we're also seeing this like longer term effects where people churn from platforms that have too much AI slop. You know, people like in the short term, but then in the long term, they're less engaged. And they're like, I don't, I don't want to just see all this slop.

• 4:15Rebecca Bellan?

  I'm curious if you've seen anything about like, you know, reviews. I make a lot of decisions for where I might eat or what clothes I might buy or what have you, based on reviews. I'm wondering if any of your customers are platforms that host reviews so that they can ensure that they're not just posting a bunch of fake bots review, which has been a problem for a while.

• 4:35Speaker 02

  Yeah, I've definitely seen a lot of this on basically every review platform, like anything from like Yelp reviews where somebody wants to get Yelp Elite, that requires like a high volume of reviews. I'm almost certain that they're taking action on this now and trying to make sure that people don't get Yelp Elite just by posting AI reviews. But we also see it on Amazon, just because there's such a really strong incentive to have a high volume of reviews and high quality helpful reviews. So I often see, I feel like every time I'm on Amazon, I scroll into the reviews, I see at least one AI-generated review on the front page, which is shocking to me.

• 5:13Rebecca Bellan?

  Yeah, it's interesting. And then, and you know, of course, all of these companies are doing their own AI-generated synthesis of all the other reviews and they're telling you kind of what reviewers say, which I guess isn't slop. We know what it is, but I don't like it.

• 5:27Speaker 02

  Yeah, but when AI is just synthesizing other AI's reviews, like what are we even doing?

• 5:32Rebecca Bellan?

  Now, tell me, how is it trained? Let's take a step back. Like, how can you tell the difference between AI-generated versus AI-assisted? And what are you using as training data?

• 5:45Speaker 02

  So we have a corpus of licensed human examples. So these are writing examples from the pre-2022, before ChatGPT came out. And then what we're doing is, say, I have an example like an essay on Moby Dick, then I'm going to ask a ChatGPT to also write an essay on Moby Dick. And then Pangram will compare the two and learn the stylistic differences in everything from like the word choices, the, all the choices that a human would make differently than AI. And then we do this on sort of like a large scale. So basically, the total training set is 50% human text and 50% AI text.

• 6:25Speaker 02

  We also have some different kinds of prompts because we want to take, when we decided to take Pangram from being binary to understanding AI assistance, we added a bunch of AI assistance prompts into our training set. So now we'll look at, say, an essay, and then we could say, make it more detailed, or improve it, or polish my writing. And on this scale, we're also able to understand the degree of difference, like from the human text to the AI polished text. How much did it change? And then we're able to teach our model to estimate the degree of change by just looking at the final product.

• 7:08Rebecca Bellan?

  Do you have any like findings as to what it's actually looking for? Or is it kind of still just a black box, right? I mean, we all see the, you know, Amanda, one of our reporters wrote a funny article. It's not just one thing, it's another thing. And it cracked me up because we had like traded, you know, that version of AI slop between Slack messages for a while, just being kind of silly. Mostly she was, she was doing great at it. I'm, I kind of sucked.

• 7:34Rebecca Bellan?

  I'm like, what, you know, I think The Economist recently had a piece that looked at, it's the punctuation, it's the repetition, but like, is there any, are there any patterns that you're able to pull out?

• 7:43Speaker 02

  I think the patterns that I look at are different than the ones that the Pangram model looks at. So for me, it's really like these like fluffy adjectives and metaphors that don't really make sense and the sentence structures, like it's not just next but why. But I think what the Pangram model is looking at is it's looking at a bunch of weak signals for any like given word choice. It might say like, AI prefers this word choice. Humans sometimes choose this, this word choice. And then, but if you take this over the course of 50, 100, 200 words, if you're looking at like every single word choice is the, what AI would prefer to say, then I think we can build up confidence over time through a bunch of weak signals to say that we're actually very confident that this is AI-generated.

• 8:28Rebecca Bellan?

  So it's not, you're not using like watermarking, right? It's not, it's not that.

• 8:32Speaker 02

  Watermarking has its limitations. For example, like a watermark can tell you whether a text was touched by Claude. So it can say like, if you take your text, ask Claude to polish it, then the watermark, it may show up, but the Claude watermark can't tell you, was this a light polish versus was this fully AI-generated? Whereas I think Pangram is much better at that and much more granular.

• 8:58Rebecca Bellan?

  What are you seeing in terms of conversation about what people will accept from AI-generated versus AI-assisted? Like, you know, there's a lot of Substack creators and we can talk about your deal with Substack in a minute, but there's a lot of creators who, you know, maybe don't have an editor and they might put their work through Claude or ChatGPT to polish it up or check it for grammar or, I mean, like Grammarly, I guess, is a version of this, right? So what are you seeing in terms of like what people will accept and not call slop? And just be like, well, I guess this is just like a tool that people can use?

• 9:32Speaker 02

  I think we're seeing a really wide range. Like, I think we haven't really settled on any given norm. There's some people who are like, oh, the ideas are my own. And then I put in bullet points and had Claude write me a whole Substack letter. And I think they're, they get tens of thousands of readers. Like, like this is like a legitimate use case that people are okay with. But on the other side, I think if somebody's writing like very like. Personal pros or like so- something like that. I, I think there's a lot less tolerance for Claude coming in and then like making it just sound like it's AI generated. And then I would say like, nobody really has issues with like using AI to fix your grammar

• 10:13Speaker 02

  or do small edits. But I, I think oftentimes people will under disclose the amount of AI use. Like if, if I put my blog post into Claude and say, make it better, like it, it will change fifty plus percent of the words sometimes, which I think is just erasing my own voice.

• 10:30Rebecca Bellan?

  Yeah, that, that is definitely a problem. Like if you want, as a writer, if you want to develop a voice, you can't, you can't just ask the chatbot to make it better for you because it will take away your voice.

• 10:40Speaker 02

  Yeah, and it, it has its own voice, which is very distinctive. Like people read it. It's like, I, I, I know that writer that sounds like Claude.

• 10:47Rebecca Bellan?

  Right, or, or it's chat, right?

• 10:49Speaker 02

  Exactly, or it sounds like chat.

• 10:50Rebecca Bellan?

  Okay, so talking about content creators, you raised nine million recently. You also partnered with Substack. And so Substack's new tool will tell you who's been writing their newsletters with AI. So readers can, you know, scan any post, a note, or apply a comment that's over a hundred characters. And to see a human versus AI estimate, you know, we talked about how people on social media, they, they want to know what they're interacting with. I know you've also got the Chrome extension, which kind of gives you a little bot next to people's names as you scroll through like your LinkedIn to, you know, if they might have used AI to, to write their posts. What other kinds of platforms do you see using this tool down the road? I mean, Substack makes a lot of sense, but

• 11:34Rebecca Bellan?

  are you working with other similar types of platforms?

• 11:37Speaker 02

  I think to be determined on which platforms we'll work with, but I think.

• 11:41Rebecca Bellan?

  You don't want to break the news right here, right now?

• 11:43Speaker 02

  I, as much as I would love to, we'll see. But the Chrome extension allows us to bring Pangram to platforms, whether they want us there or not.

• 11:53Rebecca Bellan?

  Yeah.

• 11:53Speaker 02

  Which is, I think, really great because not, not everyone is going to have the same aversion to, to AI slop, especially if it's improving their own metrics on their website. So I think that the main thing that I'm really excited about is bringing the Pangram image detection to the Chrome extension as well. We can bring this to Reddit, to Pinterest.

• 12:14Rebecca Bellan?

  Yeah, I do want to talk about the, the image detection model because that's new. But before we get into that, I would want to ask you about this kind of witch hunt component, right? So like The Atlantic wrote in May, America has a Pangram problem. A Wall Street Journal author called it a defamation machine. People have been using it to accuse others of using AI to write. And it's not always perfect, right? You have a really good success score and I use it a lot of the time. I'll just throw shit in there just to, just to see. I'll like mess around with it. And just every once in a while, I check on your product to see how it's working. Sometimes it's perfect, sometimes it's not. How, how are you dealing with that when, you know, we're in this world where people just want to point the finger?

• 12:56Speaker 02

  So Pangram has a one in 10,000 false positive rate. That means text that was fully human written will get flagged as AI about one in 10,000 times, which is very much not zero. That means like, at the scale we're running Pangram, we're going to have a few dozen false positives a day. So I think that's something really important.

• 13:14Rebecca Bellan?

  It's got to be more than that. Anecdotally, I've kind of noticed it a little bit more. And sometimes it changes. Sometimes I'll put a whole piece of text in and it'll say human written. And then maybe I'll put a paragraph and it'll say AI assisted or AI written. And I'm like, interesting, how did that change? You know? And, and why is it false? Right? So it's tricky because it's such a needed tool, but you don't want to go around falsely accusing people like students trying to, you know, submit papers and what have you.

• 13:42Speaker 02

  Yeah, no, I, I mean, I think it is like something we're working to manage. The goal of Pangram is to help enable like positive and like meaningful discourse about this. But I do think it would also be an issue if it was taboo to call someone out for slopping it up and like posting AI content as their own. Because I, I think that's part of what's happening right now. Like people are willing to call out AI content because they feel like there's just so much AI slop that's being shoved in their faces through all channels. And like, we are in this really critical period of setting norms where we need to be able to say like, hey, don't post complete AI slop under your own name. Like, that's not cool.

• 14:24Rebecca Bellan?

  Yeah, I want to pivot back to the photo detection because this is something that was released, what, like last month, I believe, we covered it. And talk to me a little bit about that. Why now? What's cool about it?

• 14:35Speaker 02

  Yeah. So why now? It's because AI images have gotten really good, kind of scarily fast. I think we went from AI images being obvious, you just count the fingers, you look at the text and make sure it's not garbled. And today it's like GPT image can produce a really like photorealistic image that I cannot tell whether it's AI or not. Grok can do this too. Midjourney still produces like these really striking images, which you have some like AI sense of them, but they're all good now. Well, good meaning like realistic. Is it good that we have a bunch of images that we can't tell if they're AI or not floating around the internet? Maybe not so much. So I think that is where Pangram

• 15:20Speaker 02

  can help as well to help flag if something is AI generated before misinformation goes really viral.

• 15:27Rebecca Bellan?

  Yeah, okay. So this is, you know, available on the, the Chrome extension as well. Obviously from, from a social platform standpoint, I want to know if what I'm seeing is, is AI generated or not. But are there any other interesting use cases that are, you know, very urgent or emerging aside from kind of deepfakes?

• 15:45Speaker 02

  I mean, I think we've just seen an increase in fraud that's enabled by AI. So anything from like submitting a fraudulent expense and having AI like produce a receipt for you to like, you know, like having AI put a bug in your food and then sending it to DoorDash to get a refund. Like, I think all of these things are like really possible with AI and we don't really have any defenses against them right now.

• 16:10Rebecca Bellan?

  Yeah, and or like marketplaces catching, you know, fake listings or dating apps. I would imagine them having this as like some kind of layer in between. Like you do, so you, you have like public partnerships like with Substack, but are you also like kind of white labeling the product and, and being this, you know, internal layer that, you know, maybe the, the end user wouldn't see?

• 16:28Speaker 02

  That's the plan for us. We'll see. We work with a couple of social platforms on this. So Quora still uses Pangram. Yeah, they have questions and answers and they don't want somebody to just use AI to answer all the questions because you have a chatbot for that. People are looking for like real human connection and anecdotes.

• 16:45Rebecca Bellan?

  Where, where, again, with like AI assisted versus AI generated, where do you draw the line, right? Like what is, like if it's a real image versus an AI modified image, right? If someone's essentially using AI to do what Photoshop would have done or, you know, cut someone out of a photo or what, or what have you? Like, are you able to tell that kind of thing?

• 17:04Speaker 02

  Depends. I think like, if you're using AI to literally crop the photo, then like, no, we can't tell. But if you're using AI to like regenerate the photo in a way that like one person is now missing, like yes, we can tell that. And I also think like, there's a lot of good reasons to be able to tell this. Like, it's kind of reality distorting if you can just change a picture of what happened in the past. And so I think it's really important for us to be able to say, like, no, this is not an actual accurate representation of, of the past, of reality.

• 17:37Speaker 00

  Every company today is looking for an edge. Better insights, faster decisions, greater resilience. But as AI drives new growth, it also creates new demands on energy and infrastructure. Schneider Electric is connecting energy, automation, and digital technologies to help organizations make smarter decisions and operate more efficiently. Because the future won't be shaped by AI alone, it will be shaped by energy and industrial intelligence. Schneider Electric, advancing energy tech.

• 18:04Rebecca Bellan?

  Yeah, but I mean, the flip side of that is there's, there's a liability question if you do false positives on a photo of like a war crime or a police shooting or an insurance claim, right? Like, you know, is that something you're thinking about, about being liable if you say something is AI generated when it wasn't? And, and maybe it was a necessary photo that was taken to hold power to account?

• 18:23Speaker 02

  Sure, yeah. I mean, in some sense, like, that's why like multiple accounts are very helpful. Pangram image false positive rate is about one in a thousand. So maybe it'll happen where we have 10 pictures of a war crime and one of them Pangram says it's, is maybe AI generated, but neither of them, it's confident are real. Then I think we still like have a good sense that like this was real. If there was only one picture and Pangram says it's AI generated, then I think there's reason to be skeptical. And I think starting with this healthy skepticism because Pangram says it's AI is like a really great place to be. It's, it's a better starting point than like being skeptical of every single image because we have no idea what's AI or not.

• 19:06Rebecca Bellan?

  Right, yeah. And I am currently skeptical of every single image. Okay, so, so you guys are standing out in the space, right? I mean, you've had The Atlantic writing about you, you have profile Substack deal, but you're not the only ones here. There's competition in the space, including Winston AI, Originality AI, Copy Leaks, GPT Zero. One thing about GPT Zero, they were recently acquired by Superhuman. And I'm kind of curious how you look at the long-term, you know, business value of Pangram. Like, is this, will this always be a standalone product? Or does this become part of, you know, someone's stack, right? Can you imagine this being infrastructure and seeing that as like an exit strategy for the company?

• 19:46Speaker 02

  I'm not really thinking too much about exit strategy, but I do see the long-term vision of Pangram is it's not really just a tool that people use to paste things in and get an answer. But we're really looking to be more like core infrastructure for the internet. I think as the internet is more and more agents and bots and AI generated content, it's going to be increasingly important to find what was human and prioritize it. So we're going to be working with platforms. We're going to be building out this Chrome extension to help people navigate an internet full of AI content. And the goal is to be deeply embedded and help kind of like solve this problem that, that AI has created out of nothing.

• 20:28Rebecca Bellan?

  Yeah, I mean, it sounds like you're trying to save the dead internet theory or save us from the dead internet theory.

• 20:33Speaker 02

  Exactly. I think we're, we're actually dangerously close. If, if we do nothing, like dead internet theory will happen within the next few years.

• 20:42Rebecca Bellan?

  What happens if we have a dead internet? Maybe we'll, I'll just go like touch some grass, you know?

• 20:46Speaker 02

  Yeah, I mean, I think it's gonna be a lot more walled gardens. It's gonna be like closed group chats and like, in a sense, the anonymous internet that's been free and open is, has been dying for a while. Like, most of the internet is now under real names and real faces, which I think is a big change compared to 15 years ago.

• 21:06Rebecca Bellan?

  Yeah. How do you future-proof? Because, you know, AI will only continue to get better. They have better resources to stay ahead, to sound more natural. And again, also, people might become more tolerant of AI-generated or AI-assisted content over time. So, yeah, how do you future-proof? What's your moat, exactly?

• 21:23Speaker 02

  I think for us, it's really about, again, it's like, understanding the degree of AI content. Like, if I, if somebody like co-wrote something with ChatGPT and they think it's really good and they put a lot of time and effort into it, then maybe like, that's, that's great and, you know, we should be celebrating that there was a lot of effort put into this, even though there was also some degree of AI assistance. Whereas, if you see a Twitter profile and the last 30 replies are all AI-generated and they're all like a minute apart, then I think we know, like, with very high certainty that this is like an entirely automated account. And that's really, I think, what we need to be solving for in the long run. Like, while Pangram is really good at telling you the degree of AI assistance today, it's

• 22:07Speaker 02

  not because any AI assistance is bad, it's because it's really important to tell the difference between some AI assistance and fully AI-generated.

• 22:16Rebecca Bellan?

  Do you think that, you know, in the future, you know, one of my first gigs was copywriting. I would just write website copy, SEO-focused website copy. And I think all of those jobs have just completely disappeared, right? I wrote blogs for companies, stuff like that, which a lot of it is kind of nonsense anyway, but, you know, this was the name of the game. Do you think that we're gonna place a premium on human content in the future? Like, well, a lot of these jobs that, you know, got demolished, maybe they'll start coming back as brands have, you know, they might suffer from a poor association with having just like nothing but slop.

• 22:47Speaker 02

  I think definitely, yeah. I think slop and like AI content in general, like, is, is associated with low-quality, low-effort outputs. We're hiring writers, we're working with writers directly, like, everything that we put out is human-written. Anthropic is still hiring writers, OpenAI is still hiring writers. I, I think it's very clear that, like, there is still a need for this role. People aren't just outsourcing it to AI at, at the highest levels. But I do think that the premium we're putting on this means that, like, the, the lowest level, like, copywriting and like people churning out like SEO blogs, like, that was kind of human slop already before AI was a thing.

• 23:29Speaker 02

  So, so I think these are the ones that are going to be automated and probably these jobs are just gonna be gone.

• 23:36Rebecca Bellan?

  Yeah, yeah, slop has existed for a long time. And I think like the thing about slop is that it's not just that it's, it's, it's the volume, right? It's not just like that you might be used AI for it. It's, it's like when it's, when it comes out in such a, a low-quality volume, that's when you wanna call it slop.

• 23:51Speaker 02

  Yeah, I think that's a really good point of like, when volume and quantity matters over quality, then AI is just gonna do a better job. AI can produce more quantity at a higher volume, cheaper than a human could. But when quality is what matters, I think that's where we're still prioritizing the human touch and the human voice.

• 24:09Rebecca Bellan?

  Okay, great. Well, hey, that's a great place to end it. Where can our listeners connect with you and, and find your work online?

• 24:15Speaker 02

  You can find me on Twitter at Max underscore Spiro underscore. You can find us at pangram.com.

• 24:22Rebecca Bellan?

  Okay, great. Well, thanks again for joining us on the show. To our listeners, you can find me on Twitter and on LinkedIn. You can find Equity on Twitter and Threads. Talk to you next time. Equity is hosted by TechCrunch senior reporters and produced by Teresa Locan Solo with editing by Cal. Subscribe on YouTube or wherever you get your podcasts and find out what's next at techcrunch.com/events. Thanks so much for listening and we'll talk to you next time.