Pangram Wants to Decide Who Gets Published

Imagine a polygraph company paying to sponsor a police chiefs’ convention, then putting its own executive on stage to tell the chiefs the machine is as reliable as DNA. Nobody from the defense bar is on the panel. Nobody takes questions. A couple of chiefs walk out and sign up on the spot.

That’s close to what happened in New York this fall.

The product was Pangram, an AI detection company, and the room was full of literary agents and editors. The people who decide whether your book gets published heard a sales pitch dressed up as a panel, and some of them bought it.

I’ve written about this before. AI detection can’t be evidence, and a detector score on your manuscript is worth a lot less than people think. What’s new is the sales push. Pangram doesn’t need the tech industry to believe it. It needs publishing to believe it, and publishing is starting to.

This video lays out the whole mess, and it’s where several of the numbers below come from.

What is Pangram AI detection?

Pangram is a startup that sells software to estimate whether a piece of text was written by a person or by an AI model. You paste in text, and it hands back a percentage. Its chief executive is Max Spero. It raised nine million dollars this summer, and in July Substack built it into its platform, so any reader can now scan a post, a note or a comment of more than 100 words and get a private verdict on whether a human wrote it.

It’s also become the tool people reach for when they want to accuse a book. The Atlantic used it to publicize a reported 60% AI score for Dagger Mouth by H.M. Wolf.

How does it decide? According to Pangram’s own technical writing, as summarized by the novelist Andrea Bartz, the company took a large set of human-written text and had AI models produce synthetic doubles of it, same length, same topic, same tone. Then it trained its model to tell the two apart. Bartz points out that the human side of that training set leaned heavily on old public-domain books and product reviews, with very little modern fiction in the mix.

The tool learned what “human” looks like mostly from writing that isn’t a contemporary novel, and now it’s being sold as the test for contemporary novels.

What did Pangram claim at the AALA conference?

In late September, the Association of American Literary Agents held its People of Publishing conference in New York. Pangram was a sponsor. Its chief technology officer, Bradley Emi, sat on an AI panel with a literary agent, a law professor and the founder of an AI licensing platform. Bartz was in the audience, and her account is the reason any of this is public.

Emi told the room that Pangram’s current model has a false-positive rate of less than 1 in 24,000. He said they want it to be as reliable as DNA testing, and he put DNA’s rate at 1 in 100,000. The agent on the panel disclosed that he had a financial stake in Pangram and encouraged everyone in the industry to use it. Afterward, Bartz overheard two attendees saying they had just signed up for accounts and were relieved a reliable tool finally existed.

There was no author on that panel. Nobody stood up for the person whose manuscript gets scanned. In a follow-up, Bartz reported that the connections ran deeper than one agent’s investment. A computer science professor whose research uses Pangram is a close friend of its CEO and gets free credits from the company. He’s also in a relationship with the co-CEO of a top New York agency.

Jane Friedman, who runs the most widely read newsletter in publishing, came out of the same panel with a headline that says it all: an agent compared using AI detection to using spellcheck.

Spellcheck. A tool that can end a career, and somebody in the business of representing authors compared it to the red squiggle under a typo.

How accurate is Pangram really?

Nobody outside the company knows.

That’s the problem.

Pangram’s own numbers keep moving. When it released Pangram 4 in July, it claimed a false-negative rate below 1% and a false-positive rate below 0.1%. Emi said less than 1 in 24,000 on stage.

Spero told reporters the latest model runs at 1 in 10,000 and admitted it’s “definitely not zero, especially when we’re scanning millions of documents a day.” Bartz found a published paper that put the false-positive rate at 1 in 800. She also found no independent study that tests Pangram on fiction published in the last five years, and no outside benchmark anybody agrees on. The company built its own test set and then graded itself on it.

Real-world results look worse than any of those figures. The science writer Janelle Shane ran her own text and got 98.7% AI. In another test, swapping a single em dash for a colon flipped a passage from one verdict to the other. Plenty of writers report pasting in the same pages twice and getting two different answers.

Even the friendly studies don’t support the sales pitch. A June 2026 paper in the International Journal for Educational Integrity tested four detectors on 160 papers and found Pangram the most consistent of the group. Most consistent of four flawed tools is a long way from DNA.

Is Pangram better than other AI detectors?

Not by much. As far as I can tell, Pangram is pretty much the same as every other detector. It’s maybe slightly better in some areas and slightly worse in others. The tool isn’t what sets it apart.

The marketing is. Pangram obviously got money, and it spent it on changing the game. The other detectors sell themselves on whatever they sell themselves on. Pangram sells itself as perfect, the best there is, as reliable as DNA. And it goes after the big contracts and gets itself built into the places people already work, like Substack. Nobody else is doing that, to my knowledge.

Once a tool is built into the platform and sponsoring the conference, nobody has to go looking for it. It’s just there, carrying the platform’s credibility, and people stop asking whether it works.

What happened to Jerry Falade’s $2 million book deal?

Jerry Falade is a debut novelist, educated in Nigeria and studying at Southern Methodist University in Dallas. His thriller, Call Me, I’ll Hide the Body, set off a bidding war among more than fourteen publishers. Minotaur, an imprint of Macmillan, won the American rights in a two-book deal reported at about $2 million, and HarperCollins took the U.K. rights. Film interest was circling.

Then it was gone.

According to Falade, a journalist ran his manuscript through AI detection software and sent the report to the people he was working with. The video puts the score at 97% AI. In her introduction to his essay, the publishing consultant Kathleen Schmidt says the book was uploaded to Pangram without his consent. After one phone call, his agents withdrew the book, telling publishers they were “no longer able to authenticate how the manuscript fully evolved from origin to completion.” The deals stopped. Reports say he’ll likely have to return the advance.

Falade says he didn’t use AI to write the novel. He did use it for research. His main character is a chemistry student, so he ran a small local model to answer questions he didn’t want sitting in a search history, like how to clean up blood and conceal a body, and to convert Nigerian units of measurement into American ones. He kept those records and tells other writers to keep theirs.

That’s research. Every crime writer alive has looked up worse.

Why do AI detectors flag writers whose first language isn’t English?

This part made me angry. Falade grew up with Yoruba, a language built on proverbs, parallelism and wordplay. In his essay he quotes his mother’s proverb: “It is not the fastest runner who arrives first, but the one who knows the way.” That’s a not-X-but-Y sentence. It’s also the exact construction the internet now calls an AI tell.

After the accusation, he went through his own 85,000-word manuscript and counted. Eighteen not-X-but-Y constructions and 46 em dashes. That’s how he writes, because that’s how the language he grew up in tells a story. Every publisher who bid on the book praised the voice.

The same voice is what a detector reads as a machine.

I’ve said it before: “It’s not X, it’s Y” is ordinary human writing. Shakespeare used it. Preachers and lawyers have used it for centuries. AI overuses it because it learned it from us. Punish the construction and you punish every writer whose culture leans on it, and you teach the rest to write like a form letter.

Can a handwritten book fail an AI detector?

Yes. Mia Ballard is the author of Shy Girl, the horror novel Hachette pulled in March 2026 after Pangram scored it at about 78% AI. Ballard says a freelance editor added the AI material without her knowledge. Whatever happened with that book, look at what happened next.

For her next novel, Poly, Ballard wrote the first chapter by hand, on paper. According to Bloomberg’s reporting, Pangram labeled that chapter entirely AI-generated. An earlier book of hers, Sugar, came back at 83%.

A woman sat down with a pen to prove she’s human, and the machine said she wasn’t.

If that doesn’t end the argument about reliability, nothing will.

Are writers changing how they write to beat AI detectors?

They are, and it’s the worst outcome of the lot. Bloomberg found writers dropping em dashes, cutting rhetorical repetition and parallel structure, and planting deliberate spelling mistakes so their work looks human. Some have abandoned styles they used for years.

Falade saw it coming. He predicts a day when a debut writer pays a detector twenty dollars to prove they wrote the book they spent years on, and gets a 95% AI score back. Then they spend months rewriting it, killing the voice in the process, pay again, and hope the machine is in a better mood.

The best sentences in a book, the ones with rhythm and balance and a little music in them, are the ones most likely to get flagged. So writers will sand them down. Pangram will end up teaching people how to write, and it’ll teach them to write worse.

This is AI Derangement Syndrome with a business model. Once AI becomes the only question anyone asks about a book, somebody will sell you the answer.

How can authors protect themselves from false AI accusations?

Keep everything.

Drafts, outlines, research notes, the files with dates on them, the emails with your editor, the voice memos. A score is a guess, and process is something you can show. Falade kept his research logs, and they’re the strongest thing he has.

Don’t run your book through a detector and rewrite it to please the machine. If you do, you’re letting a startup edit your novel.

If an agent or publisher tells you a detector flagged your work, ask what tool they used, what the score was, and what else they have. A percentage by itself isn’t evidence of anything. Then put your drafts on the table. If you use AI as an assistant, as I do in my writing, say so plainly. I explain how in AI disclosure on your book. Hidden AI use is what turns a non-issue into a scandal.

And if you’re an agent or an editor reading this, ask yourself who paid for the conference.

Judge the book

Publishing already has a test for whether a book is any good.

Somebody reads it. Agents read it, editors read it, and if it’s good, readers buy it. Falade’s book survived that test with fourteen publishers.

It didn’t survive a percentage.

Pangram will get better, and the models it’s chasing will get better faster. The arms race doesn’t end with a reliable verdict. It ends with writers twisting their own sentences to stay out of trouble. Keep your drafts, keep your voice, and don’t let a sales pitch decide who gets to be an author.

From the author

The Birth of the Augmented Human
My case for using AI to sharpen your own judgment, and why the human directing the tool still matters most.

Frequently Asked Questions