the internet is full of people who never say anything
INDUSTRY NOTES #6: Max Spero, CEO Pangram
This issue is presented in partnership with Pangram, the AI detection company behind the Chrome extension that lets users scan their feeds in real time and tells you how much of what you’re reading online was actually written by a human being. Which feels fitting, because today’s conversation is really about trust: in language, in platforms, and in whether the internet can remain meaningfully human as synthetic content floods every corner of it.
Good morning and welcome back to as seen on!
Today’s newsletter is another edition of INDUSTRY NOTES, a series where I interview the most interesting founders, operators, and internet people shaping how culture and technology actually move.
Here is something Max Spero wants you to think about. Right now, without knowing it, somewhere between 10 and 20 percent of what you read online was written by a machine. Most people have no idea, and Max thinks that’s the problem worth solving.
His company, Pangram, is fast becoming one of the most controversial and culturally interesting AI companies on the internet. Pangram’s AI detection tools are now being used by publishers, universities, writers, and increasingly ordinary internet users trying to figure out how much of what they’re reading online was actually written by a person. The company’s Chrome extension lets users scan their feeds in real time, flagging AI-generated posts across platforms and giving your internet consumption a live “humanity score.”
In this conversation, we talked about AI slop, authenticity online, why LinkedIn may already be backing away from “Write With AI,” the coming backlash to synthetic corporate voice, where editing ends and ghostwriting begins, and why human intention itself may become a status signal online.
Let’s talk about it in the comments. Enjoy!
Max, you’ve been in AI research since 2016 — Google, Stanford’s robotics lab, autonomous vehicles at Nuro. That’s a very different world from content detection. What made you leave that path and build Pangram?
Me and my co-founder Bradley have actually known each other for over a decade. We were in the same dorm together at Stanford. I’d been in AI research since around 2016 or 2017, first at Stanford’s robotics and AI research lab, then at Google training large machine learning models. Bradley went to Tesla. At some point we both started asking the same question: everybody is trying to build AI. Nobody is thinking about the second-order effects.
That’s really the thesis from which we started the company. We thought detectability was going to be an important problem for a long time. In the early ChatGPT days, a lot of people said it was impossible or that nobody would care. I feel pretty vindicated now. People care about slop more than ever.
Was there a specific moment, whether it was a piece of content you saw, a conversation, or something that made the problem feel real and urgent enough to bet a company on?
Late 2023, we had just started incorporating reviews into the AI detection model. I was looking at a review for a water bottle and noticed that it seemed AI-generated. I looked at the profile and they just had hundreds of fully AI-generated reviews. I just had the thought, “This is going to be so bad if we’re unable to tell these apart from the real reviews.”
BUILDING THE INTERNET’S HUMANITY SCORE
Walk me through what Pangram is actually doing when it reads a piece of text. What is it looking for, and why does it work when other detectors don’t?
Other detectors use a metric called perplexity, which is a measurement of how confusing a piece of text is to an LLM. On average AI-generated text tends to be lower perplexity than human written text, but well-known documents like the Declaration of Independence are also low perplexity, because the LLM has seen the document many times during training. This is why so many other AI detectors have high false positive rates.
Pangram is a deep learning classifier trained on millions of pairs of human written text and AI-generated mirrors of that text. If you think about writing as a decision tree, there are thousands of decisions made on how exactly to turn sentiment to words. Pangram is able to extract the patterns in text that AI makes much more consistently than humans, and uses this information to build confidence on whether a text was AI-generated or not.
You claim a false positive rate of one in ten thousand. How do you actually know that, and what happens when someone says you got them wrong?
This is what we spend a lot of our research cycles on. We measure our false positive rate by looking at documents that we know to be human-written and making sure the rate stays below one in ten thousand, which is very, very low.
With that said, it is not zero. Pangram does occasionally make mistakes. The longer the document is, the less likely the mistake is. The other thing I want to emphasize is that the future we believe in, or at least the direction the world seems to be going, is one where a lot of writing becomes AI-assisted. So a huge amount of our research effort is going toward understanding the spectrum and degree of AI assistance.
When somebody brings a false positive to us, we take it seriously. Those cases go into an evaluation set. Every time we retrain our model from scratch, which we do every three to six weeks, we look at whether we’ve improved on past mistakes.
YOUR FEED IS MORE SYNTHETIC THAN YOU THINK
Tell me about the Chrome extension, because this is where Pangram moves from a B2B tool into something culturally significant. What does it actually do, and what behavior are you hoping it changes in how people consume content online?
The Chrome extension was an idea we had after launching our Twitter bot where people could tag @pangramlabs is this ai? And get an answer? And we had been seeing this big rise in AI slop on our personal social feeds. So we built this extension that automatically scans all of your social feeds, including Twitter, Reddit, LinkedIn, Substack, and tells you if any posts come back as AI. It gives you a feed health score which tells you how much AI content is in your feed. I think it’s great, it’s really helped me think about where I’m getting my information diet from. I’ve unfollowed a few big accounts that post a bunch of AI content, and overall I feel like I’m spending my time more wisely by reading things that were actually written by humans.
You mentioned that before people turn on Pangram they don’t know what they’re consuming, but once they do and realize 10 to 20 percent of their feed may already be AI-generated, it lands differently than expected. Why do you think people are so surprised?
I think the value of filtering out unknown AI-generated content from your information diet is supremely valuable in a way that a lot of people don’t fully understand until they actually do it.
Before people turn on Pangram, they don’t know what they’re consuming or how much of it is AI. When you turn it on and it’s between 80 to 90 percent human, that means up to 20 percent of what you read is AI-generated. That’s a lot. And the fact that people don’t know this means they could be filling their feeds with AI content without even realizing it. Especially for people who are reluctant to use AI in the first place.
The trend line is the thing that worries me most. Even if it’s still mostly human right now, that 10 to 15 percent AI-generated content didn’t used to be there. If we look back a year from now and people aren’t actively curating their feeds, that number could be dramatically higher.
Do platforms actually want less AI-generated content, or do they just want better AI-generated content?
Most platforms exist to connect people, and by definition AI-generated content does not do that. If you see a Reddit comment saying they love a specific clothing brand, but the comment is AI-generated, I’m not sure it matters that the comment is “better” than a human written one. It’s just inauthentic.
Over the last couple of years I’ve spoken to a lot of people working at these platforms, and the attitude has slowly shifted from “wait and see” to “this is actively bad for our platform.” Even LinkedIn, which is filled with “Write with AI” buttons, is now testing removing the AI buttons to encourage people to write in their own voice.
Do you think “human-made” eventually becomes a status signal online the way organic or handmade did offline?
It’s already high-status to be a writer or artist and actually make your own stuff. I know a bunch of people who refuse to consume AI content and look down on others who produce slop. I think of it more like fast food, where if all you eat is fast food you’re going to have a really unhealthy diet. You can eat a bunch of non-organic veggies and be fine. But if all you consume is AI content then that’s going to have negative effects on your brain and your information diet.
WHERE DOES EDITING END AND GHOSTWRITING BEGIN?
The Shy Girl case put Pangram in front of a mainstream audience. You ran the manuscript, declared it 78% AI-generated, and the publisher pulled the book the same day the Times piece ran. The author said she didn’t personally use AI, that an editor inserted it without her knowledge. How do you think about your role in what happened?
For the Shy Girl case, I think it is very clear that the book is largely AI-generated. As far as I know, this is the first major published book to reach that level. We haven’t seen a legitimate false accusation at that scale before. But the fear people have around that possibility is real, and it’s something we think about constantly.
The future we believe in is not one where everything is binary. We’re trying to understand degrees of AI assistance. To be able to say this text was translated by AI, or this had a light polish pass, versus this was generated wholesale from a prompt. Those distinctions matter enormously.
The Wall Street Journal’s James Taranto reran the op-eds your research flagged through Pangram’s own tool and got significantly different scores. One came back 100 percent human. He argued the inconsistency alone is disqualifying. What’s your honest response?
The Pangram model has evolved significantly since mid-2025, when the study was run. At the time, Pangram was trained to detect any significant AI use and predict either “Yes, there was AI use” or “No, this is fully human-written.” In December, we launched a model that was able to differentiate between fully human written, AI-assisted, and AI-generated. It made the product a lot more useful, but by definition it meant we were catching less AI than before. We had to ensure that our new model still had a 1 in 10,000 false positive rate, which meant that it was less sensitive to text that had some AI use but wasn’t fully AI-generated.
I still stand by the results. They were the most accurate numbers we had in 2025, but since then we reran all data with the latest Pangram model to get the latest numbers for the machine learning conference it was accepted to. In my opinion, today’s data is much more interesting because of Pangram’s additional capabilities in differentiating between AI assistance and generation.
You’ve said the future is AI-assisted, not AI-generated, and that Pangram needs to move toward a spectrum rather than a binary verdict. What does that actually look like in practice?
I want to be able to say this text was translated by AI versus this looks like it just had a light polish pass applied to it. Right now, if somebody takes a piece of writing and pastes it into ChatGPT and asks it to rewrite it, we might call that AI-assisted or AI-generated depending on how much changed. We still don’t fully understand how to precisely measure how much AI altered the original text. That’s the next frontier for us.
The degree of AI assistance is genuinely the harder problem. But it’s also the more important one because that’s where most real-world usage actually lives.
I run a newsletter and I use AI to clean up my tenses, sharpen transitions, polish grammar. But the thinking, the voice, the argument is mine. Pangram might flag that as assisted. Should it? Where does editing end and ghostwriting begin?
I think you need to be really careful when you’re using AI. By default it’s going to prefer it’s own voice and ways of saying things. There was a really cool paper by Google DeepMind called How LLMs Distort Our Written Language where they showed that AI models not only alter voice and tone, but also intended meaning. They tend to make responses more neutral, less opinionated, and less creative overall. I have no issue with the grammar and tense aid, but I think if you’re having an AI model take a full pass at rewriting your text, you should understand that the AI model will inject enough of its own voice that other people will be able to notice the AI assistance.
Casey Newton, who runs Platformer, recently wrote that AI has already commoditized link aggregation and is starting to eat into news analysis too. Soon, the only thing worth paying for will be original reporting, scoops, things only a human with real sources can produce. as seen on primarily curates and contextualizes the news. Reading that piece was uncomfortable. You’re building a tool that can tell what’s human and what isn’t. Curious to know what future do you see for writers like me?
One of the things that AI will never be able to replace is real human connection. I think we’ve seen a shift in that direction with increasing influence to streamers and Substack writers. There’s always going to be value in original reporting, getting quotes and perspectives from real people. But I also think that, for a writer like you, the value is in the fact that your readers trust you and want to hear your personal perspective. You’ve built that deep trust with your readers over time, and it would totally betray that trust if you, for example, turned your Substack into an AI workflow. Of course, I’m glad you wouldn’t do that. But I think imagining the reaction from readers is an indication of the value that you provide.
WHAT HAPPENS WHEN EVERYBODY STARTS SOUNDING THE SAME?
You said something interesting during our conversation. That it would be disappointing if we ended up in a world where someone writes bullet points, asks AI to fluff them into a long email, and then somebody else’s AI condenses it back down into bullet points. Why does that bother you so much?
I think AI is going to bring a lot of good into the world. We’re already seeing things like autonomous vehicles saving lives. AI is going to help cure diseases. It’s going to automate rote tasks and give people back leisure time.
But there’s this risk where people fail to draw any line at all around AI and end up letting it dictate their personality, beliefs, opinions, and style. At a certain point it stops feeling like you’re talking to a real person.
People should absolutely use AI to make their own lives better. A personal query to an AI assistant is probably more useful than a Google query at this point. But going out into the world and presenting the outputs of a chatbot as your own, that’s the dangerous part. That’s the thing that erodes the trust we have in society. And once that trust is gone, it’s very hard to get back.
“Going out into the world and presenting chatbot outputs as your own is the thing that erodes trust. Once that trust is gone, it’s very hard to get back.”
Do you worry AI language itself becomes the default corporate voice? That eventually every deck, email, and presentation starts sounding vaguely optimized in the same way?
We’re seeing this growing backlash to AI content as more people are learning to recognize it. Companies are trying to adopt it because it saves time and effort, but if an email or press release sounds AI-generated, people are going to dismiss it as low-effort. They didn’t even take the time to write it themselves!
So we’re already starting to see this push where companies doing outreach are writing things themselves, they’re keeping it a little unpolished, maybe missing some punctuation to show that they’re human. I don’t necessarily condone that either, but I think it’s starting to become clear that, in business, you are seen in a different light if you don’t put effort into your stuff.
What happens if platforms themselves don’t actually want transparency?
Sure, it’s an uncomfortable truth to confront a quality problem on your platform. I got into a bit of a spat with the Medium CEO over it, when we found that 50% of new Medium posts were AI-generated. Ultimately, the consumer chooses, and I don’t think you can brush a quality problem under the rug. It’s the core value that a platform brings to consumers, so they’re either going to have to care, or hemorrhage users until they die.
THE NEXT VERSION OF THE INTERNET
What’s the biggest thing Pangram is working on right now?
The biggest thing is granularity. We want to get much better at understanding degrees of AI assistance. I want Pangram to be able to distinguish between text that was translated by AI versus something that just had a light polish pass applied. I think that’s very possible.
On the other side, we’re expanding the Chrome extension to more platforms. One thing I’m particularly interested in is AI-generated YouTube transcripts. We want Pangram to work there too.
News is another huge one. I want Pangram to automatically classify every news article you read. Because if you’re reading a news article and it’s AI-generated, there should be some alarm bells going off in your head. Is this real? Is there an actual journalist behind this? Is this some kind of disinformation campaign?
You’ve said the internet has operated as a relatively high-trust society for decades and we’re at risk of losing that. What does the internet look like if that trust collapses?
I think we see a collapse of the incentive structures that have historically kept the Internet open and free. If people can’t trust what they read online, they’re going to have to retreat into echo chambers filled with voices they trust. We’re probably going to see the continued rise of influencers for this reason, but I think there is also a lot of value in the anonymous side of the internet. The people on Twitter with 400 followers, not 40,000. If these voices get drowned out in a sea of AI spam, I think we are all going to be a bit worse off for it.
Last question. What’s your Industry Hot Take?
Hot take, the LLMs are never going to “get so good” that their output is indistinguishable from human writing. Today’s AI models are much closer to a real human being with personality and preferences, which is interesting from a capabilities standpoint but I don’t think they’re really multifaceted in a way that they could, for example, properly model the full distribution of humans. If you ask any major LLM what religion it would choose, it says Buddhism. This is just one example of the preference collapse that we see in assistant LLMs. We may be close to AGI by many standards, but I think we are only close to building a handful of distinct AGI technologies.
Pangram’s real bet is not that AI-generated content disappears. It’s that human intention becomes newly valuable.
In an internet flooded with infinite content, actually meaning what you say may become the rarest signal left.
Facts