Pangram Labs' AI slop detector goes viral — 25x user growth and Quora as a major customer

much for taking the time to hop on the show. How you doing?

I'm great. I'm great. Yeah, just cleaning up slop on X.

It seems like it. Uh, when did this project start? I feel like I've been start I I've seen most of the viral stuff uh in the last few weeks, but have you been building this company for a while? Give us a little bit of the prehistory.

Yeah, so the company's been around for about 2 years. We've been building models to detect AI generated content and just building doing a lot of the research.

Yes. [snorts]

And trying to be the most accurate in the market. But it wasn't until like end of 2025, December 29th actually was the first day where I launched the Twitter bot and it's really just blown up since then. People are using it on articles on basically all these long form posts that people are just posting slop.

Yep.

Online.

Are are you already monetizing it? Because I imagine you have a ton of attention, ton of people that wind up on the tool. Uh I interact with it mostly through I see the screenshots that people post or I see your post. Um, but are there businesses who have an incentive to partner with you? Are you talking to big platforms, small companies? Like, uh, have you have you started like what what does the business look like today?

Yeah. Yeah. So, individuals can either sign up for a free account on Pangram. They can get a Chrome extension and check things uh in the browser or you can pay $20 a month to get access to um, a lot more checks.

Uh, but we also work with internet platforms. So, one of our biggest customers is Kora. So they send most of their content through Pang just to understand like is it AI generated or not because that's against their policy and they want to keep the platform high quality and so yeah we've had a number of API users a bunch of different internet platforms

large and small people basic like either like officially rling on you or like basically DIY reinforcement learning by just uh generate me an essay go into Pangram it's at 100% make some changes and remove the m dashes. Okay, it went down to 70%, rewrite a couple sentences and just keep iterating until I get it down to 0%.

People are trying. I think that's part of the fun is like that this is such an adversarial space. And so like part of keeping up with it is just making sure Yeah. We're robust to these gpo attacks and this like RL

Yeah.

Uh on the model.

Yeah. Uh, does it make you mad when people are sharing an article on X and it's obviously written by AI, but they're just like, "This is the be this is the most

well-written essay I've ever studied I've ever read on this topic. Changed my life.

This changed my life.

Oh my god." Infuriating. And the person probably spent like 15 seconds just typing in chat GPT. Make it really deep. Make it really good.

Yeah.

The worst one that I read was

Yeah. Yeah. Tell us. was this like um the one that went viral about the like supposed Uber Eats or Door Dash whistleblower. That's right. It was like they wrote on Reddit. Their entire post was AI generated and they're talking about how like you know the company's so evil. Um and it I think it went viral just because people wanted to believe that it's true even though it's like obvious slop.

Wait, were you able to clock that as AI just personally or with Pangram? because when I read it, it didn't jump out to me as AI generated, but there were other flags that just seemed like factually in uh inaccurate or uh just inconsistencies and like logical

to me it immediately read like fanfiction. It was like I'm on a I'm in a public library laptop. It's like who does that?

Uh like you can just post on Reddit anonymously. Like Reddit is not going to come and like arrest you. I don't know.

Anyway, uh yeah, but yeah. How did you clock it personally with the tool both? What were you thinking?

Yeah. Yeah. I was reading it. I'm like this reads like a sto this reads like a story and not like real life somebody writing this and then I checked it with Pangram said 100% AI. I went to the comments and there already other people were saying that you know this is obviously AI.

Okay. Yeah. And and then uh yeah talk about the the the impact like are you is this an AI safety project? Do you think that there's uh there's like a societal uh you know disaster that comes from AI slop? Is it something we'll adjust to? Is there a white pill good outcome here? How do you like process the the the like philosophical part of your work?

Yeah. So, I mean I think a lot of the big labs are already doing a great job on the internal AI safety side, which is making sure Claude doesn't go out and like autonomously decide to kill everybody. But I think there's this like application layer, the like external AI safety of like can people use AI for harm? And I think this is a great harm that people just like aren't really like thinking or even trying to tackle. I think like this threatens um the internet as we know it today. I I don't if you're aware of like dead internet theory of just the idea that it the internet is just going to become bots talking to bots and it's like a ghost town where people can't participate anymore because their voices are drowned out by bots. And I think we're very much at risk of this if we don't have any technology to solve it.

What do you think about the inverse? I I posted that uh you could probably drive someone insane by just quote tweeting everything that they say with the 100% AI panggram screenshot because so I I I write a daily newsletter and I write about 500 words and I always I never use AI to write it or even proofread it. There's typos in there. Um, so people will still accuse me of using AI and so we're like maybe we need to live stream me typing it or something and I don't know, it just doesn't matter. But it is sort of infuriating when you take half an hour, you write out some thoughts and then someone's like, "Oh, this is AI or share the prompt." Even if they're joking, it's like kind of gets under your skin. What What do you think about that?

Yeah. No, it's it's annoying. And I mean I think that's another reason why like we're building this technology is there are other AI detectors out there that have 1% or 5% false positive rate. And so they'll like incorrectly flag a lot of things even like the Declaration of Independence. And so we so Pangram has a 1 in 10,000 false positive rate. So the goal is we are almost never flagging something incorrectly. If we say it's human like we're we know it's human.

Yeah.

Are you working uh with any of the big dating apps yet? I imagine that uh this is something they'd want in kind of like actual like on on the user side as you're uploading because they want to not have AI generated content because that would create a bad experience for users broadly.

I cannot say I' I've heard some like about some issues from people who work in dating apps. Sure.

And I think um the issue is growing. I heard an interesting story from a a former uh dating app engineer who said that there they had bot problems 10 years ago. People would just go and write a Python script to just swipe on everyone and just send like basic messages. Um and he said that when they detected bot activity, if they ban them, the the the botter would just create another account. So what they did was they put them in like bot hell basically where they would only talk to other bots, which is a funny solution because then their system says, "Hey, I'm still on the app. It's still working. I'm still seeing all the API requests. Everything's working. Um, but they're just only interacting with other bots. But you can imagine getting false positive into that and just it being the worst experience. You're like, I didn't meet anyone. Everyone out of this app's a bot. So, there's a whole bunch of funny funny nuance things. I do want to ask about uh image image uh generation, image detection. Uh, that somewhat relates to

Yeah. Yeah. So yeah, what I was trying to get at is like where I I assume over time like most AI most written content will have been run through some type of model before it's published almost like a spell check even. Uh so I'm trying to think about like where where do you really not want AI content, right? Like I don't care if somebody uses

an uh a Gen AI to make like a head shot, right? Like if they just want a head shot that looks somewhat like them,

they might go into Photoshop anyway and add some color correction, increase the contrast,

pimple or something.

Not that big of a deal. But where where I do think is like the probably the most pressing issue right now. It's just like any anything news related. So if like if you're if you're sharing things that are like news or or any content that people are using to understand the world

like I could imag like like are we moving to a world where you'll have to add like a panggram watermark to the to the post like

like like is that the kind of behavior that we want because like if you if if an AI generated image goes up two million people see it and then six hours later it gets community noted and says like okay this is this was uh generated by AI, it's too late. Like 2 million people like, you know, may thought it was real.

That's a huge problem. And actually like this is happening today. So if you know like all the like the Russian troll farms where they're like posting comments and like uh news articles, they've basically Russia has expanded this to basically just use LLMs that they've fine-tuned themselves. And they've we've seen news websites come out of Russia that just have 50,000 articles that are completely AI generated. Most of them are like normal news and then some of them are really pushing the like pro-Russia anti- Ukraine stance and they're just posing as like random like Oregon local news and

Yeah. And these are crazy and and they're popping up faster than uh we can take them.

Yeah. like we can actually the white the white house shared an AI generated image

AI modified image yes

AI modified but

got community noted on axe got quote tweeted but uh yeah I mean it's an increasing increasing issue uh what uh what is the state of potential bots on other platforms uh it's great that you've got the bot working on X but uh it seems like we really need it on Facebook I I you know if there's a couch the size of a room that is in the shape of a gorilla There should be panggram in the comments saying this is AI generated. So the uh that the elderly generation perhaps is protected from slop.

So next up for us is Reddit and LinkedIn.

Okay.

So I think Reddit there's a bunch of viral posts that are AI generated. I think people know that they can karma an account and then sell it to a bot farm or something. And on LinkedIn, I think it's just funny because like people are like genuinely posting slop. Like they're not bots. They're just like Oh, yeah.

posting AI slop every day.

Yeah.

Uh are what is the state of of wiring up a bot on a modern social network these days? I feel like a lot of the APIs have closed down a fair amount. uh is it something where you need to have a relationship with the business to actually run a you know a useful bot that I would say I as a user I want this it's not malicious activity but you're still getting API access you could slop yourself if you wanted to you have the access so what's involved in actually spinning up a bot and keeping it online and keeping it posting regularly without hitting rate limits

so for Twitter is actually pretty easy um I mean besides having to pay $200 a month And then if we pass some limit um well now they have usage based API but otherwise before this the next tier up was $5,000 a month which is absurd but um

I for a for a startup that has funding if you're growing like it's it's not you it's not

prohibitively but you can't do that as a hobbyist.

Sure.

But yeah it was not too bad with X and then the other platforms seem to be a lot more locked down. So, LinkedIn, I've been waiting for API approval for two weeks now.

Um, and Reddit kind of same thing. But, I mean, Reddit seems to be pretty bot friendly, so I think we'll get there eventually.

I've never seen a reply bot on LinkedIn, but

have you been surprised by conversations with any of the platforms? Because in some ways, they just want more content that's engaging. And so, there's kind of like this dilemma where they might say like, "Yeah, we're looking to crack down on on, you know, fake imagery." But in reality, they're like, "Okay, this is driving engagement.

This is not like, you know, it's a it's it's like even with X, people have always said like, do do they really want to crack down on bots or like do they just say they want to crack down on bots?" Because in some ways, like it's just driving notifications for people that might not otherwise be having a lot of uh activity.

One man's slop is another man's fame on [laughter]

the the LinkedIn trust and safety folks seemed so resigned. It seems like they're having such a bad time. Because in every single text box there's a little like write with AI button. So like in I want to compose a post I write 20 words and then have it expand it out into like a long essay. Or same with like inmails. They'll AI generate inmails that are customized for the person's profile. And so like they can't actually do anything around like AI slop or using AI as a signal because it's so integrated into the platform at this point. And I think to the detriment of the platform, like the quality of content there is so low now.

How many words how many words of text do you actually need to detect if something's a

Yeah, I've seen some people tag uh Pangram with like someone will post like one sentence and it's just [laughter] like I don't know. I don't know if it's a hallucination or something, but it's it's a funny funny meme. So, so our our official line is 75 words, but I made it smaller for Twitter. Yeah. Because obviously there's so many uh short posts and typically

what is going to happen is just going to say it's human like

unless it's like really obviously AI or like as an AI language model,

Sure. Sure. Sure. Yeah.

Yeah. The main it it almost doesn't really matter because

like I really want to know if something's AI if it's really long. Yeah, totally. I only want to read this for the most part if I know that somebody put in a lot of effort to like, you know, actually

say something.

And there is the flip side which is like I I'm if if if someone does a really cool deep research report in one of the LLMs and they just shared a link to it and they're just like, "Hey, I was interested in this particular research report. I generated it. You can just click here and read exactly what I what what the prompt result was." Uh, I might be down because I like reading deep research reports sometimes if if I prompt them and I might be interested in what you prompted. Uh, the funniest uh the funniest short Panggram uh uh request is always someone will say, you know, is this article AI? And then Pangram will come back and say, yeah, it's 100% AI and then someone say, is that is is the request to call Pangram AI? And Pangram will say like [laughter] no, it's not. It's human. No,

but uh anyway, so so so what's next? Uh what's next with the business? What's the state of the company? How large is the company? Where where do you see this all going?

Um, so far we're I mean we're pretty small. We're still seed stage nine people. Um, but uh since July we've grown 25x in terms of users and number of queries.

There we go. We've seen a lot of we've seen a lot of interest basically in this um consumer level interest just like I want to know what's slop or not. I I'm not like a teacher. I'm not a writer. I just want to know that what I'm reading is authentic. And so

what's next for us?

In my opinion, look forward to a Chrome extension that's like ad block but for slop.

I like that.

It's going to automatically slop block everything that's slop on your on

on the internet. You own sloppern.com.

I do not.

You should.

It's not for sale. Somebody else somebody else beat you to it. But uh

somebody has it.

Oh well.

Well, congratulations on the progress, the virality. Uh I love seeing it and thanks for all the hard work. Uh it's

really really doing uh good for the world

for sure

in the slop era.

Yeah.

Thanks so much.

Have a good rest of your day. We'll talk to you soon.

Talk soon.

Bye.

Let me tell you about Plaid. Plaid powers the apps you use to spend, save, borrow, and invest. Securely connecting bank accounts to move money, fight fraud, and improve lending now with AI.

Next up,

next up,

actually

we our next guest will be in the reream waiting room.

Still waiting on Russ from

In the meantime, uh we have some massive news. Johnny Carson's former LA estate just listed for $40 million. Little teaser. The mansion section. We will be getting to it later in the show, but uh the Los Angeles estate that late night TV legend Johnny Carson once called home has hit the market for a dash under 40 mil. All proceeds from the sale have been earmarked for three charities. Cedar Sinai's medical center, the David Gaffin Foundation, and Share, which supports disabled, abused, and neglected children. Uh Compass is sharing the listing. He's a longtime host of the Tonight Show, which and he died in 2005 at age 79. an absolutely legendary uh talk show host. Uh big inspiration for me. If you haven't listened to the Johnny Carson uh episode of Founders, you absolutely should. It's fantastic. Uh also, we we re we previously reviewed his other house uh the Malibu compound that recently hit the market for $110 million. Um

still sitting on the market. So, originally built in 1950 for Mvin Loy, the late film producer and director, best known for his work on The Wizard of Oz. Have you seen The Wizard of Oz?

YES.

WOW.

BUT I think I was like forced to watch that in school at some point.

Yeah. The Yellow Brick Road, Dorothy. It's Toto. It's good. Uh the the estate rests, this particular estate rests on 1.5 acres in the prime East Gate Enclave. A gated driveway bolstered by a security booth empties at an expansive motor court with two garages with room for four vehicles with sprawling stucco and stone accented structure at its center offering six bedrooms and 11 bathrooms. It's 9,000 square feet across a single level. It's introduced via a marble clad foyer with access to an elegant wet bar equipped living room that wraps around to a cozy wood panel den. I like that the sound of that den. You love a wood panel tent.

I love wood paneling generally. Ideally, mahogany, the official wood of business.

We need We need a mahogany sponsor. Truly

huge missed opportunity. They've been mahogany strongest soldiers.

We love mahogany. Uh from there, a spacious living room sports a fireplace and pocketing doors spilling out to a pillared loia. Loia, I'm not familiar with that word. While a formal dining room connects to a kitchen with an eat-in island and a breakfast nook, a soaring rotunda flows to a posh primary estate, which has a fireside sitting area, and dual walk-in closets and baths. It has a tennis court,

of course,

a basement with a wine celler, detached two-story guest house that once housed Carson's office. Uh, the amenities continue outdoors. The manicured grounds host a sund deck encased pool and spa flanked by an open air cabana cabana with a fireplace as well as a lighted tennis court and a viewing pavilion holding a kitchenet and powder room so you can view the uh the tennis that's played. Per the Wall Street Journal, Carson had to convince Joanna to relocate from the east coast to the west coast. She was truly a New Yorker at her core, so it was hard to move her out to California. Uh the nephew said who's of

one and a half acres in Bair. Will that do it? Not a bad not a bad option. Uh well there are some other there are some other stories in the uh in the mansion section which we should go through. Uh first let me tell you about railway. Railway simplifies software deployment. Web apps, servers and databases run in one place with scaling, monitoring and security built in.

I loved hanging with Jake yesterday.

He is potentially a Joe Rogan CEO. Fantastic conversationalist. Fantastic. uh the way he breaks things down, tells stories. I I really enjoyed that conversation. He'll be back. He's he's on a generational run and congratulations to him again on the good news. Uh so, the Call of Duty creator Dave Anthony has a custom Bair mansion. It just sold for $22 million. Uh and we