Semafor Sep 18, 2026
With Max Spero, CEO of Pangram
Pangram has a couple hundred thousand users, trains on tens of millions of documents, and measures its own false-positive rate against a few million pieces of writing published before ChatGPT existed.
The researchers Max Spero asked when he started told him the job could not be done, or would become impossible within a year. He trained a model on college admissions essays anyway.
"I talked to my researcher friends, they all said, it's impossible."
Spero trained machine learning models at Google and then at a self-driving car company before starting Pangram in September 2023. The company now has a detection deal with Substack, and the hosts put its most recent funding round at $9 million.
The full episode is covered here so you can skip it.
Here are the 16 insights that matter.
Key Takeaways
The model is trained on pairs: a human document and an AI-written synthetic mirror of it
There is no single tell — the model adds up weak signals across a whole document
The false-positive rate is measured by running the detector over pre-2022 writing that is not in the training set
Instruction tuning and reinforcement learning put the tells there, on his account, not next-token prediction
The detector separates model families but cannot tell Opus 4.5 from Opus 4.6
A llama model fine-tuned on Wall Street Bets comments got past it, so evasion is possible for a determined writer
Pangram catches the laziest uses of AI, which is most of them
Publishers turning against AI text surprised him, and he ties it to the lawsuits over pirated books
He would rather see AI compress five pages into an op-ed than expand five bullets into one
If intelligence becomes abundant, he expects human touch to be the scarce input that carries the value
1. How Pangram Works
Asked to explain the product, Spero described a machine learning model trained on pairs of documents: a human-written one, and an AI-written imitation of it that the company calls a synthetic mirror. The model learns the difference in style, word choice and sentence choice rather than looking up a banned word list.
There is no single red flag
It's really mostly operating on these really subtle features.
Max Spero
He said the detector is aggregating weak signals over the length of a document, so one suspicious word decides nothing. It returns a degree of AI involvement rather than a yes or no: human, AI-assisted or AI-generated. Text and image detection are live, with video coming.
2. The 1 in 10,000 Test
The hosts put it to him that "you guys have a 1 in 10,000 false positive rate" and asked where the figure comes from. His answer is a measurement, not an estimate.
The method is to take writing published before ChatGPT, which is known not to be AI-generated and is also held out of the training set, run a few million of those documents through the detector, and count how many come back as AI. Anything flagged there is almost certainly wrong, so that share is the false-positive rate.
3. The Tells He Notices
Asked what marks writing as machine-made, Spero named specific habits. Two of them are the words honestly and genuinely, and a third is the phrase "I need to push back on that". He reserved his sharpest objection for a setup he called a cataphoric teaser, the line that promises a revelation, as in "here's the part that nobody tells you" or "here's what most people get wrong".
Every sentence is trying to win
But I think overall it's just like AI really wants every sentence to be the most important sentence.
Max Spero
Which is why it is tiring
It's just trying to make every sentence sound impressive, which makes it very tiring to read and it kind of like makes my eyes glaze over.
Max Spero
There is no thinking ahead to the end of the essay, he said. Every line competes with the one before it.
4. Why Models Write Like That
Asked why the tells exist at all, Spero put the blame on the training stages that come after the base model rather than on the base model itself. GPT-2 and GPT-3, he said, had a little mode collapse but mimicked human writing reasonably well and lacked these habits.
Assistant tuning put the vocabulary in
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.
Max Spero
Reinforcement learning did the rest, on his telling: a reward model rates what counts as good or engaging writing, and the constructions that score well get pushed far past where a human writer would leave them.
The reward signal amplifies a normal habit
I think this is a very reasonable construction in normal writing, but then it just gets turned up to 11
Max Spero
5. Who Uses It, Who Pays
The user base is a couple hundred thousand people, and Spero listed students, educators, writers, lawyers and general internet users among them. The heaviest use is not the paste-in dashboard.
The Chrome extension is the sticky product. It labels posts as AI-generated or not directly in social feeds, among them Twitter, LinkedIn, Reddit, Substack and Medium, and he said that proactive labeling is what people love. The dashboard gets used, but less often than people scrolling.
Asked who the real paying customer is, he called it an open question, and said so with his investors listening.
The consumer case is the internet getting worse
I think that there's going to be more internet users who need something like this to navigate the internet as it's just like increasingly filled with bots and AI slop.
Max Spero
The enterprise ambition is everywhere at once
We want Pangram to be integrated in every internet platform every like law firm, every school.
Max Spero
6. It Started With Essays
Spero had been training machine learning models at Google and then at a self-driving car company when ChatGPT was released, and said he could not tell AI text by eye at the time, which is exactly why he thought a machine would have to do it.
The field told him not to bother
I talked to my researcher friends, they all said, it's impossible. Like don't try or if it's not impossible today, it'll be impossible next year.
Max Spero
He started anyway
So we started off in September 2023.
Max Spero
The first models were trained on college admissions essays found online, against AI-written mirrors of them. The results beat his expectations, and working out why taught him what the models prefer — including how they like to open and close an admissions essay. Generalizing that beyond essays to reviews, blog posts, poems, recipes and ordinary web writing took until early 2024.
7. Reviews And Mode Collapse
Reviews were where the company spent its early effort, because Spero thought fake-review detection would be the first large use. What he found was sameness: the restaurant "nestled in the heart of Brooklyn", the moment the writer walked in, the juicy steak, the closing recommendation.
Asked what mode collapse means, he gave the plain version. A model can produce the whole range of reviews, from a one-star complaint to a long rave, but asked simply to write a review it lands in the middle of the distribution every time — descriptive, warm, and close to the center.
One looks fine, ten thousand do not
it looks reasonable when you look at one but when you look at 10,000 you realize they're all the same
Max Spero
He described accounts earning Yelp elite status by publishing an AI-written review a day, with those reviews reaching the front page because every ranking signal read them as detailed and helpful.
The ranking systems could not see it
But I think the algorithms didn't realize that it's fully AI generated.
Max Spero
Pressed on whether this was personal outrage, he said the objection is that it is inauthentic, and pointed at review farms: a business pays for hundreds or thousands of reviews, and the same accounts seed themselves with AI-written reviews elsewhere to look like real users. The old version of that business had to hire writers abroad and the output was poor. Now, he said, the best reviews of a restaurant are the machine-written ones.
8. Telling The Models Apart
Spero said the model families have distinguishable voices, and that the word "loadbearing" is a Claude habit rather than a ChatGPT one. His description of Claude was verbose and prone to hedging; ChatGPT he described as trained toward being curt, with a preference for one-to-three-word sentences that he finds odd in prose.
The company looked inside its own detector to check how much of that it had learned. One of the later layers separates the model family on its own.
Family, yes; version, no
it actually can't really distinguish Opus 4.5 versus Opus 4.6
Max Spero
9. Publishers Took A Stance
Asked what had surprised him as Pangram became widely known, Spero said the scale of public interest, and then the reaction of book publishers.
He expected the industry to give up
we've had a series of literary scandals, which honestly I didn't really expect
Max Spero
Instead it drew a line
But I think I've been very surprised and impressed that publishers are taking a more principled stance against AI text.
Max Spero
He offered his own explanation for it, which is commercial as much as principled: many of those publishers are suing OpenAI and Anthropic over pirated books used without consent, and a company in that position does not want to be printing the output of the same models.
10. What He Asks Of Media
Spero said every media company should be running detection internally, and framed it as disclosure rather than prohibition.
The objection is hidden use, not use
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
Max Spero
Detectability is part of his argument: where a tool can spot it, readers and listeners will eventually notice it too, so a score tells a publisher how the writing will be received. The rest is keeping people honest against the temptation to hand over work when busy.
He put himself in the same position, with investor updates and memos to his own team.
He has tried it on his own memos
So then I have to go back and like rewrite most of it myself anyway.
Max Spero
11. "The point is the point"
One host raised the counter-argument. Paul Gigot, the Wall Street Journal's editorial page editor, defended an op-ed by Stanley Druckenmiller criticizing his protégé Scott Bessent on the grounds that what matters is whether the piece reflects the author's own argument and whether the author has the standing to make it. Nobody doubted the opinion was Druckenmiller's, and the news was that he had taken a shot at Bessent, whoever assembled the sentences. The host's own version was blunter: the point is the point, not the arrangement of the words.
Spero said he understood the argument and does not fully agree, and his reason came from trying to write that way himself.
Drafting with a model shifts what you claim
I find that like it's so easy to let AI put words in your mouth and say things that you maybe like wouldn't have necessarily actually said but seem reasonable.
Max Spero
He then split the practice in two. Writing five pages and asking a model to compress them into an op-ed is defensible, because the thinking is yours and the model is editing. Writing five bullet points and asking for an essay is not.
Expansion adds nothing a reader wants
Just show me the bullet points so I can read it and move on.
Max Spero
12. The Substack Deal
Substack brought Pangram in after taking what Spero called a principled position: Substack's Chris Best posted that he did not want people passing off AI text as their own, while allowing that some uses, such as an aggregation of research or financial news, are legitimate and that readers do not much mind.
The mechanism is a scan available on any Substack article from the three-dot menu, which returns the full score across human, AI-assisted and AI-generated. Writers can also publish a standing disclosure describing how they use AI, in their own words — anything from not at all to talking into a voice notes app for 15 minutes and having a model turn it into a newsletter.
The reception split exactly as you would expect
On the other side, I think we did get a lot of hate from people who use AI to write and compose their writing.
Max Spero
Writers can say how they work
On top of that, Substack writers can also add an AI disclosure for their column.
Max Spero
Pressed on whether readers actually click the three dots, Spero said the signal matters most at the moment someone decides whether to pay for a subscription.
13. Does The Line Dissolve
The hosts put the hardest question to him: if human writing drifts toward machine writing while machine writing improves, the distinction stops meaning anything, and one host wondered aloud whether Spero is a monk complaining about the printing press in 1457.
His first answer is that the comparison fails
AI is really different technology than anything that has come before it.
Max Spero
A calculator, a typewriter, a word processor and a printing press do not produce cognition, he said, and that is the difference.
He accepted both halves of the problem: models getting better at imitation, and human writing getting closer to the machine as people read more of it. He then described a way to validate a prolific writer without inspecting the text.
Bracket a year between two human ones
So something that we were looking at is if you're a prolific writer, say you've published a lot and you have everything you wrote in 2024 was human written and everything you wrote in 2026 is human written then we can probably guess that your 2025 is also human written.
Max Spero
Authorship is getting muddier as keyboards and mail clients finish sentences, and he said he does not dispute that. What he wants to separate is not human writing from lightly assisted writing, which he treats as the same category, but either of those from text with no human input at all — his example being an autonomous system loose on forums and messaging apps, talking to people and trying to get them to act.
The case he actually cares about
I think here like these are fully autonomous AIs that are saying things with no human input at all.
Max Spero
The hosts pushed on the escalation: a system like that would run its own output through humanizers to get past him.
He thinks the defender can hold
I think we can win.
Max Spero
His argument is friction rather than certainty. A student can still edit an AI essay until it reads as human, but it costs more effort than it used to, and the same logic applies to a future adversary.
Cover-ups do not last
And there's always a cat and mouse but there's always at some point we can catch up and like tracks can't be covered forever.
Max Spero
14. Fine-Tuning To Evade It
The hosts raised a colleague who had broken stories about undisclosed AI op-eds and who thinks he could fine-tune a Chinese model on his own writing samples and beat the detector. Spero said he buys it, and gave a precedent from his own lab.
An extreme fine-tune already got past it
And we fine-tuned it on Wall Street Bets comments and it was just kind of undetectable
Max Spero
That model produced belligerent nonsense rather than publishable prose, which is why it worked. He said a writer with thousands of published articles and a careful fine-tune might manage the same thing with usable output, and that he would like to see someone try.
What the tool is actually good at
I think Pangram is best at catching the laziest uses of AI.
Max Spero
He disagreed with the premise underneath the question — that everyone ends up with a personal model trained on themselves.
He expects concentration, not personalization
They're far too expensive for any individual to run and that's what everybody uses because they're the best.
Max Spero
15. The Detection Roadmap
The company is scaling in two directions at once: more kinds of media, and finer resolution on each one. Video is next and AI audio is likely after it, covering both synthetic speech and machine-generated music.
The second axis is granularity. Detection started binary, became three-way with an AI-assisted category, and is heading toward a per-passage verdict.
A verdict per paragraph, not per document
So we could say these sentences are AI assisted. This paragraph is fully AI generated and this part looks mostly human written.
Max Spero
The more complete the account of a document, he said, the more useful it is across industries as authenticity becomes more valuable.
16. Where Value Accrues
His closing argument is an economic one. He expects models to keep improving quickly, and takes the abundance case from Sam Altman and Dario Amodei at face value for the sake of the argument: frontier intelligence for everyone at close to no cost.
What follows, on his reading, is that the money moves to whatever is still scarce. His analogy is component pricing.
The bottleneck captures the price
suddenly RAM prices went up 10 times because RAM is just the scarcest resource
Max Spero
And the same thing happens to writing
So I think same here where human touch is going to be the limiting factor and it's going to be the scarcest resource
Max Spero
Bonus Insights
The hosts pointed out that the company best placed to build an undetectable text generator is the one that built the detector. Spero agreed, and said the tools already exist internally.
He builds humanizers and will not ship them
We've actually built our own humanizers internally
Max Spero
Those models exist to make the detector more robust and will not be released, which he called dangerous and powerful. Outside, he counted a couple dozen humanizers of varying size, most of them selling to students because that is where the money is. Pressed on the industry's standard defense, that if you do not build it someone else will and so you had better build it fast, he said he is not sure he agrees, particularly where the thing being built could kill everyone.
What his friends at the labs say
Spero said he has friends across the major AI labs, and that their assessments of Pangram range from a very pro-social AI company to the only good AI company. One host noted that every technology company says it is trying to make the world better, and that the ones that mean it eventually take the line out of the mission statement.
A skeptic changed his mind on air
The hosts closed by arguing about the tool between themselves. One had been a Pangram skeptic, on the grounds that a piece of journalism passes through an editor who changes enough that nobody can reconstruct where it started, and said he found the people who tweet detector scores at him more annoying than the practice they are complaining about. He came around after the interview.
The other host had the mirror-image position: convinced by the technology, unconvinced by the moral case, and put off by public shaming on Twitter. He said Spero's wider argument moved him off Gigot's side of it, while adding a case that fits neither camp — a Substack writer he reads who drafts in Mandarin and publishes machine translation, which reads as machine-written and is original and valuable anyway.
They also ran one host's own articles through Pangram. The verdict was human.
Spero's bottom line is that the distinction worth defending is not human writing against assisted writing but either of those against text with no human input, and that if intelligence becomes cheap the return goes to whatever human contribution is left.
Products, Companies & Tools Mentioned
Pangram (A couple hundred thousand users, a Chrome extension that labels social feeds, a dashboard, and a false-positive rate measured against pre-2022 writing)
Substack (Brought Pangram in as a scan available on any post, and lets writers publish a standing disclosure of how they use AI)
Claude, ChatGPT, Gemini and Grok (Distinguishable by family inside the detector, though not by version; he described Claude as verbose and hedging and ChatGPT as trained toward short sentences)
Yelp (Where he found accounts reaching elite status by publishing an AI-written review a day, with the reviews ranking well because they read as detailed)
Llama (The open model his team fine-tuned on Reddit comments, producing text the detector could not catch)
Wall Street Bets (The subreddit whose comments made that fine-tune undetectable, because the output was too abrasive to look like assistant writing)
OpenAI and Anthropic (Defendants in the publisher lawsuits over pirated books, which he says is part of why publishers now refuse AI text)
Google (Where he trained machine learning models before starting the company, and whose mail product he named as an example of authorship getting muddier)
Suno (Named as the kind of machine-generated music an audio detector would have to identify)
Watch the full episode:
More Max Spero interviews and podcast appearances
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: