Pangram CEO Max Spero on why human touch is about to become the scarcest thing on the internet

0:10Welcome to Mixed Signals from Semaphore Media. I'm Ben Smith, the editor of Semaphore, joined by our media editor, Max Taney. This week, we're talking to the CEO of Pangram, Max Spiro, whose technology is quite accurately detecting AI-written text across the internet, across media, leading both to some small-scale embarrassment and to some big questions about what is the role of humans and of machines in our lives and particularly in the media industry. We're going to talk to him about how Pangram works, about the escalating game of cat and mouse between increasingly sophisticated AIs on both sides, and about his broader vision of the role of humans in the media of the future.

0:51We'll be right back with Max Spiro after the break.

1:07driving innovation and reshaping the marketing landscape today. Whether it's AI's impact on the industry or creators' role in culture, Josh is getting into the weeds with the people who are actually shaping what's next. These are your notes from the frontier. You'll hear from CMOs, agency legends, brand visionaries, and brilliant thinkers. Search for Frontier CMO from Think with Google wherever you get your podcasts or watch on YouTube.

1:38Max, really excited to have you on the show today. I think I want to start out by asking you a very simple question. I think a lot of our audience has used Pangram or has checked it out. But for those who haven't, can you explain what Pangram is and how does it work? Yeah, Pangram is technology to detect AI-generated content. So we train models that will look at the content itself. So text, image, and soon video and tell you, is this AI generated? What is the degree of AI versus non-AI? And I don't get a sense of where it's from. Can you explain in a little more depth just how do you do that? Yeah, yeah, exactly.

2:25Sure. So it's a machine learning model. So it's essentially trained on tens of millions of documents where the Pangra model sees a human written document and then an AI written, we call it a synthetic mirror. and so it will kind of, the AI document will mimic the human document. It'll look similar, but the pan-gram model is going to be able to pull out basically, yeah, signals from the style and word and sentence choices of the document to try and figure out, is it AI generated or not or AI assisted or human written? And when you say sort of word and sentence fragments, is it just like it looks for the word load bearing or is there something more subtle?

3:15It's really mostly operating on these really subtle features. So it's rather than like looking for a single red flag, like load bearing or like it's not just X but Y, it's really aggregating a bunch of weak signals together over the course of the document. You mentioned, I've heard you mention on another interview that basically you guys have a one in 10 ,000 false positive rate. Where does that number come from? Yeah. So what we do is we actually look at pre-2022 data. So text and writing that we know is not AI generated, but it's also not in our training set. And so we're able to look at a few million of these documents, run Pangrom on them, and see how many of them we classify as AI, which is almost certainly incorrect because it was pre-Chat GPT.

4:16And so that's how we measure our false positive rate. What makes writing, to you, AI-generated versus human-generated in terms of what it looks like or what are its ticks? I've heard you talk on some other interviews about how you just kind of get the sense. When I read certain pieces of writing on the internet, you mentioned the it's not X, it's Y kind of trope. But what are some of the most obvious hallmarks of AI-generated writing? There's a couple that come to mind, which honestly, it's like if chat sheet pieces end honestly or genuinely, I need to push back on that. The other one that I see often that really annoys me in AI text is it's called a cataphoric teaser.

5:15Someone says, here's the part that nobody tells you, or here's what most people get wrong. And I absolutely hate that. But I think overall, AI really wants every sentence to be the most important sentence. It's like there's no thinking ahead to the end of the essay or something. It's just trying to make every sentence sound impressive, which makes it very tiring to read. And it kind of makes my eyes glaze over. Yeah, that's such a subtle, interesting thing. That's so interesting. Why does it do that? You've paid a lot of attention to the hallmarks of AI writing. Why does it use some of those tropes or cliches?

5:57When you train a language model, you start by doing next token prediction. So you're going to say, given the last N words, what is the next word? And so I think these models, these were like GPT-2 and GPT-3. I think these models actually, they didn't have these tells. They had maybe a little bit of mode collapse, but they could mimic human writing pretty well. But I think what happened is they started doing this assistant training, which would make it do things like overuse the word delve because that was just in the assistant instruction fine tuning a lot or fairly frequently. And then I think they started doing RL on these models and that's where they really picked up on these signals of what good writing sounds like.

6:48So I think a lot of times it's either the instruction tuning or the RL that gives these models... It's reinforcement learning. Yeah. The reinforcement learning gives these models a reward when they produce something that is considered by the reward model good writing or engaging. And then that signal, for example, like the negative parallelism, it's not just X but Y. I think this is a very reasonable construction in normal writing, but then it just gets turned up to 11 through the reward signal and gets overused to the point where it's annoying. How many users does Pangram have? Oh, we have a couple hundred thousand users.

7:30It's kind of a wide variety of people, students, educators, writers, lawyers, just random internet users who want Pangram to tell them what on their Twitter feed is AI generated. It's kind of a really broad variety of people. Who's the most frequent user of Pangram? What's the archetype for the most frequent Pangram user? I would say the most frequent are probably the internet users, people who use the Chrome extension. It's a really sticky product. And so what it does for those who don't know is you download this Chrome extension and anytime you are on any social feeds, so Twitter, LinkedIn, Reddit, Substack, Medium, all of these, you'll just get little proactive labels on every post that tells you if it's AI generated or not.

8:28And I think that's the thing that people love. a dashboard where you can go in and paste text and check if it's AI. And that is used frequently, but it's not as frequent as people just scrolling their feeds. And as you think about building this into a business and talk to your investors about building this into a business, who's the real customer base? I think this is a pretty open area of debate, not to scare my investors and tell them I don't know. You must have just raised$9 million. I assume you told him something. Yeah. Well, so I'm a strong believer in this consumer motion. I think that there's going to be more internet users who need something like this to navigate the internet as it's just increasingly filled with bots and AI slop.

9:19And then I think eventually, over the course of the next few years, we're trying to build up this enterprise motion. We want Pangrub to be integrated in every internet platform, every law firm, every school. I think there's a lot of different areas for us to play in enterprise. Why did you decide to start this? And what did the first version look like? Yeah, it's kind of funny. So ChatGPT came out and I was like, this is crazy technology. It's really different. And I had been in the AI space for a while. My background is training machine learning models. I was at Google. And then I was working at Neuro, which is a self-driving car company.

10:02And so I feel like I was very deep in the weeds of machine learning. And AI as a field covers machine learning, but also the generative AI side, which we just call that AI these days. But anyway, I was like, this is crazy technology. I just had this intuition that it's going to be important to be able to detect it, especially because at the time, I had no intuition for detecting AI text by eye. So I was like, man, this stuff is just indistinguishable. This is like, we need a machine to do this. Or maybe it's impossible. So I talked to my researcher friends. They all said, it's impossible. Don't try.

10:46Or if it's not impossible today, it'll be impossible next year. So don't worry too much about it. But we're kind of interested anyway. So we started off in September 2023. We were training our first AI detection models, and we were just looking specifically at college admissions essays. We had some college admissions essays we found online, and then we had AI create the mirrors of them. and we train a model. We're like, hey, this is actually doing a lot better than we expected. And so we were trying to figure out why and eventually built this intuition for detecting AI text ourselves. And eventually we realized, okay, we can actually detect the preferences that are latent in these models.

11:35And if we could do this and we could tell how these AI models prefer to start and end their admissions essays. But can we do this for all texts? Can we generalize? And so that was sort of the next step over the course of the next six months into early 2024, is like building a text detection model that generalized beyond essays to reviews and blog posts and poems and recipes and general internet content. What were some of the first ticks that you realized in these college admissions essays that kind of made you both understand how AI was writing and then help you develop Pangram so that it could track some of these tells or detect them?

12:22So maybe I'll tell you about reviews because that was where we ended up spending a lot of time. Like we shifted from essays pretty quickly into reviews. And at the time, I thought like this is like the next big thing is like fake review detection through like AI content detection. So we were starting to notice like all of these reviews, they have the same structure. They're like, you know, nestled in the heart of Brooklyn is, you know, this restaurant. It's like, and then they'll talk about like, from the moment I walked in, I could tell it was like going to be amazing. And then they would like, you know, go through and like, talk about like the steak was so juicy and like, like really generic.

13:00And then it would end with like, I highly recommend a visit. And so we realized like, it's really like, like these reviews are incredibly mode collapsed in a way that like people who are not. What does mode collapsed mean? Oh, yes. So there's a wide distribution of reviews of text in general, but let's just talk about reviews. It could be anything from like, I hate this one star to like, you know, a long, long glowing review. And then so AI could produce like this whole range of, of reviews. But if you just ask it to write a review, it's actually going to produce something that's kind of default.

13:37It's like right in the middle. It's going to produce something that's descriptive and warm and in line with its assistant tuning, but it's not going to stray too far from the center of the distribution. And so that's what I mean by mode collapse is like this, the output is not far from the center of the distribution. And so it looks reasonable when you look at one. But when you look at 10 ,000, you realize they're all the same. And so we were noticing that there were people who are getting Yelp elite status by just writing and publishing an AI-generated review every day. And these reviews were actually bubbling up to the front page because to all of the algorithms, it looks high quality and legitimate and a very descriptive, helpful review.

14:27But I think the algorithms didn't realize that it's fully AI generated. So these tools right now, Max, they're being used online for a kind of occasional witch hunting, essentially, or at least really like shaming people who are caught using AI. And I'm curious, were you coming from a place of sort of personal outrage? Like, are you angry about this, these reviews? Is the sort of motivation of the tool that you want to put a stop to this? Yeah. I mean, originally I was like, oh my God, I can't believe like all these reviews are AI generated and no one can tell. And like, is this like, is there something underlying?

15:05Just like, no, no one can tell that there's, you know, the internet's dead. And then obviously like, it wasn't. Why are you upset about that? I think it's just inauthentic. At least for the reviews, it's clear that there's review farms behind this where some business will pay for hundreds or thousands of reviews. And then there's also these accounts are also just putting AI-generated reviews out elsewhere to essentially seed the account with authentic seeming behavior. here. So I'm like, man, there's all these bad actors out here that can now use AI to scale up what was essentially somewhat cost limited.

15:49They have to hire review writers in the Philippines and they're going to be kind of shitty. And now the best reviews of the restaurant are just written by ChatGPT. So I mean, that bothered me initially, but I think we had really only hit the tip of the iceberg in terms of AI content on the internet. And do Claude and Chad and Gemini and Grokt, they have different voices in these contexts? Can you tell? Oh, absolutely. Yeah. What are some of the differences? I think load-bearing is definitely a Claude-ism. It's not a ChatGPT. I think Claude tends to be just so incredibly verbose and it overhedges.

16:35Basically, Claude's always saying, like trying to hedge what it's saying. Whereas I think ChatGPT has, they've trained it to be a lot more curt. So it's going to speak in a lot shorter sentences. It will try to be concise and then it will kind of fail. But also in prose, like ChatGPT has this strange preference for using these like really like short staccato, like one to three word sentences, which I think is a little bit unusual. When it comes to Pangram, when it says, okay, this is 99 % chance this is AI-assisted or whatnot. Does it know that it's Claude versus ChatGPT versus Grok versus Gemini?

17:18Or is it distinguishing? So we found it actually can distinguish pretty well. So what we did is we looked into the weights and found that one of the later layers in the model was actually able to separate out the the AI family, the LLM family pretty well. So it can distinguish Claude from Gemini from ChatGPT, but it actually can't really distinguish Opus 4.5 versus Opus 4.6. These sound very similar to the Pangram model. And has anything surprised you over the last several months as Pangram has really blown up into the public's imagination. Yeah. I'm surprised at how big we've blown up, how much people care about AI.

18:10We've had a series of literary scandals, which honestly, I didn't really expect. I think I expected a lot of people to just throw their hands up and accept that AI is just going to be writing all the books and that's fine. But I think I've been very surprised and impressed that publishers are taking a more principled stance against AI text. But maybe that's to be expected because they're also like, many of them are in the process of suing OpenAI and Anthropik over pirated books and using their material without their consent. So I think it probably makes sense that they don't want to also be publishing text that was produced by these companies.

18:52So this is a media show. We really have been really interested specifically in using Pangram to help us understand unlabeled AI editorials or journalists using AI without disclosure. And I'm really curious, some of your either clients or customers are media companies. You have a deal with Substack. Is that a place where you are hoping to grow the business? What's your relationship with media companies like? Ultimately, all media companies should be using Pangram. I think that, look, there's nothing wrong with using AI in a way that's acceptable, but I think it's really important to know where the AI is being used and also where it's detectable because that's where writers and readers and listeners are also going to be able to pick up on it.

19:37So I think Pangram is partially helpful as a tool for telling you how the content is going to be perceived and received by others. But I also think it's really important to just help keep people honest. It's so easy in the age of AI to just say, I'm busy, screw it. I'm just going to use AI and not do this work myself. And I feel this myself as well as a CEO where sometimes I need to write investor updates. I need to write memos to the team. And it would be easy to just just let Claude write it. But then I have to think to myself, whenever I've tried to do that, it doesn't say exactly what I want to say.

20:22So then I have to go back and rewrite most of it myself anyway. And I think Pangram is a really good check on that. There is another school of thought on this, which is that basically what matters is kind of the intent. If you endorse this message, I don't really care if you wrote all the words. And Paul Gigo, the kind of legendary editor of the Wall Street Journal editorial page, expressed this after the kind of financier, Stan Druckenmiller, wrote this very tough op-ed about his protege, Scott Besant, that was just like, obviously, obviously, AI composed. And Jago said, the question, you know, is whether a contribution reflects the author's original argument, and if the author has the standing and credibility to make it.

21:02Nobody can doubt that this op-ed is his genuine opinion. And kind of from a news perspective, like Druckenmiller taking a shot at Besant was huge news. Whether he wrote the words, who cares? What's your response to that? Yeah, I mean, I think some people can get away with that. If you're famous and influential, then I think people are going to care a little bit less, you know, as long as you stand by it. I think that - Well, it's more just like, the point is the point. The point isn't like the arrangement of the words and the weights and the probabilities that one word succeeds another. I understand that argument and I don't fully agree just from going through the process of me, myself trying to write with AI, I find that it's so easy to let AI put words in your mouth and say things that you maybe wouldn't have necessarily actually said, but seem reasonable.

21:55And I think there's a little bit of a slippery slope here. And then I think there's another argument, which is when I'm writing with AI, I can do two things. One, I can write a few bullet points and then ask AI to expand it into a full op-ed or essay. The flip side is I can write five pages and ask AI to compress it into an op-ed. And I think there's a much better argument to be said for allowing the AI compression of my thoughts and turning it into something that is readable and concise. Whereas I think if AI is just going to expand my bullet points, just show me the prompt, just show me the bullet points so I can read it and move on.

22:46We'll be right back with Max Spiro.

22:56In this week's segment from Think with Google, I talked to Google's VP of Marketing, Josh Spanier, about how marketers are rethinking retail. You were just at Rethink Retail in New York, which is an event for like 500 retail marketers and CEOs. What were your takeaways from the gathering? You know, retail marketing is really, really tough. It is a game of inches. One of the things I like about it is retail marketers, although they're excited about agentic commerce and the future in AI and all that sort of stuff, They really care about what's happening today and building foundations for all that new stuff coming along in the future.

23:28So really, it's a lot about how do you win in the short term? And what do I do now to build my foundations going forward for when agentic commerce comes along? And what are those foundations? So product feeds are kind of the new digital storefronts. If you can connect all your product inventory, everything you sell down to a granular level, the price, the unit size, the color, the shape, the fit, everything. and connect that into Google's Merchant Center, all of a sudden you can connect what you actually have in your inventory to the whole world. As people discover and search for unique, specific things, connecting your product inventory into Google's Merchant Center, it's a great place to start in this agentic coming world.

24:07Second, friction gets in the way of conversion. Every extra pop-up, every extra click of a button, every password reset gets in the way of someone actually buying your product. Google has invested in a number of technologies such as the Universal Commerce Protocol and Universal Wallets, which allow for seamless transactions. So every customer you could potentially reach can just click and buy your product then and there. In an agentic world, that's going to be really, really valuable. And third, you really need to just optimize for growth. You need to look at your total marketing strategy and really turn on the AI embedded in YouTube, the AI embedded in Google Search, and actually take off some of the limiters and some of the biases we bring as marketers about, well, I'll give so much money to this channel and so much money to that channel.

24:50Just let the AI work it out. And what we see is higher performance. And that, again, will set you up for this agentic coming world in the years to come. Where can people find out more about this? We've actually published all the keynotes and articles from the Rethink Retail Week on YouTube and also on thinkwithgoogle.com.

25:17So, Max, I want to return to the kind of the media side of this. You guys did this Substack partnership recently. Substack basically said, hey, you know, we don't want our platform to be overrun by, you know, people publishing every day or whatever. And it's just, you know, kind of mindless AI garbage. And it ends up filling up feeds and inboxes and it degrades the product. Can you talk a little bit about both how the Substack tool works and also what the response has been, both positive and negative, to the partnership you guys have with Substack? Yeah. So SubSec took a really principled approach, I think.

25:52They said, I think Chris Bass posted a note against clawed fishing. Basically, I don't want people posting AI text as their own and getting away with it, essentially. But he recognizes that there's a lot of legitimate uses of AI. For example, there's just aggregations of a bunch of research or financial news that the readers don't really care that much that it's AI generated. So what they did was they brought Pangram in as a tool that anyone can use on any Substack article. You can click the little three-dot menu and click Scan Text with AI, and you'll get the full pangram score, which will say, how much is it human AI or AI assisted?

26:40On top of that, sub stack writers can also add an AI disclosure for their column. They can say, this is how I use AI. It could be anything from, I don't use AI at all to, I talk into a voice notes app for 15 minutes and then ask Claude to turn it into a newsletter, something like that. So I think there's a wide variety and people can disclose however they want in their own words. And I think it was received really well. I think a lot of people thought it was executed thoughtfully and it was also, I think, just necessary to combat this rising tide of AI slop. On the other side, I think we did get a lot of hate from people who use AI to write and compose the writing.

27:37Have you gotten any signal about readers clicking on little three dots and caring? Because I think there's obviously the sort of mob mentality of Twitter definitely lends itself to like, we must find the AI. But I have not seen a really clear signal that readers of these products care that much. Have you? Yeah, I mean, I think definitely it's I think it's a really important signal for someone to choose like to decide, like, do I want to like spend my money and like subscribe here? Max, a few minutes ago, I've heard you talk about this a little bit. You've said, you know, that you are you have you have to train kind of the detection tools on stuff that was produced before, you know, basically the rollout of ChatGPT, whatever model it was in 2022, because anything afterwards might be tainted or a lot of stuff created afterwards might be tainted by some sort of AI.

28:34We're not really exactly sure. Ben, you feel like at some point it's going to just become indistinguishable. You know, Ben's writing from Ben's writing assisted by, you know, Ben's agent who's helping him kind of manage it. The increasing blurriness between the two is just going to eventually mean that there is no distinct, meaningful distinction. Is that an accurate, am I describing the feeling? Yeah, totally. I'm kind of sort of wondering if Max is like, it's like 1457 and Max is like a monk running around yelling about Gutenberg Bibles, you know? Well, yeah, that's, so I want to ask, Max, what do you, what do you think about Ben's, you know, Ben's idea that these are all just going to essentially, and I think this is what Paul Gugot was saying as well, is that eventually these are going to kind of just be the same thing and why are we trying to stand in the way of that?