The 24-Person Brooklyn Startup That Can End a Book Deal
Pangram says it can tell whether writing is human or machine-generated. Publishers are listening. Writers are furious. The question is whether the tool is actually reliable.
Key points
- Pangram, a Brooklyn-based AI-detection startup with 24 employees, has raised $13 million total, including a $9 million round in July 2025.
- The company scored novelist Mia Ballard's debut book at 78% AI-generated in January 2025; publisher Hachette subsequently cancelled the release.
- Substack announced in late July 2025 that it would integrate Pangram into its platform so readers can check posts for AI involvement.
- A survey of 1,481 working writers by Gotham Ghostwriters found that 61% already use AI tools in their work.
- Critics note that Pangram's CEO posted his assessment of the Ballard book using a copy obtained from a piracy website.
Above a Popeyes in Brooklyn, in a office staffed by two dozen people, sits the company that may have more influence over writers' careers right now than any literary agent in New York.
Pangram is an AI-detection startup. Feed it a piece of writing and it returns a percentage score estimating how much of that text was generated by an AI system, the technology behind tools like ChatGPT and Claude. The company has raised $13 million since it was founded in 2023, a tiny sum by the standards of an industry where rival OpenAI has raised roughly 14,000 times that amount.
For most of its short life, Pangram was unknown outside specialist circles. That changed fast.
How did Pangram become so influential?
A single book put it on the map. In January 2025, rumours spread online that Mia Ballard, author of the self-published novel Shy Girl (since picked up by major publisher Hachette), had used AI to write it. Ballard denied the claim. Pangram's CEO and co-founder Max Spero posted publicly that the manuscript scored 78% AI-generated on his tool. Hachette cancelled the release.
More accusations followed in quick succession.
| Work accused | Pangram score | Outcome |
|---|---|---|
| Shy Girl (Mia Ballard, Hachette) | 78% | Deal cancelled |
| New York Times Modern Love column | 100% | Public controversy |
| Commonwealth Short Story Prize winner | 100% | Public controversy |
| Daggermouth (novel) | 60% | Under scrutiny |
| Call Me, I'll Hide the Body (thriller, sold for $2.4 million) | 97% | Under scrutiny |
In late July, Substack, the newsletter platform used by millions of writers and readers, announced it was weaving Pangram directly into its product so subscribers could check any post at a click.
Should writers and readers trust the scores?
That depends on who you ask, and the honest answer is: not unconditionally.
Spero, who is 30, explained to Wired AI that the tool works through two main techniques. The first, which Pangram calls "synthetic mirroring", feeds human writing into large language models (AI systems trained on vast amounts of text) and asks them to produce something similar, teaching the detector what AI output looks like. The second technique, "hard negative mining", hunts for cases where the tool gives a wrong answer, then uses those mistakes to improve future accuracy.
Critics point to a more basic problem. Spero's assessment of the Shy Girl manuscript used a copy downloaded from a piracy website. When asked about this, Spero said: "Yeah. And, like, yeah. It is what it is. I hadn't looked too closely. I didn't go read the whole PDF. I just put it straight into Pangram."
Spero himself plays down Pangram's role in the book cancellations. "Basically, with every book deal, to my knowledge, it hasn't really been about the Pangram score," he says. "That is a part of it, but if you talk to anyone involved, it is only a small, small part of the bigger picture."
Not everyone is reassured. Publishing consultant Jane Friedman says writers feel that AI-detection companies are "just as evil, if not more evil, than the AI companies themselves." One author told Wired AI there is a growing sense that an entire career can be destroyed with a single click.
What does this mean for ordinary readers and writers?
If you read on Substack, you will soon see Pangram scores next to posts. Treat them as one signal, not a verdict. The tool gives a best-guess percentage, not a definitive finding, and false positives (incorrectly flagging human writing as AI) remain a real risk across all detection products in this category.
If you are a writer, the practical reality is that publishers and platforms are already using these scores as part of broader editorial conversations. Spero's own framing is worth keeping: if the value of your work depends on readers believing it is human when it is not, that is the core problem, not the detection score itself.
Pangram is currently hiring six new roles, which would grow its team by 25%. A small company with a lot of influence is about to get bigger.
Common questions
Can AI-detection tools be trusted to give a correct answer?
No detection tool is perfectly accurate. Pangram and its rivals, including GPTZero and Turnitin, all produce probability estimates rather than confirmed facts, and all carry a risk of wrongly flagging human-written text.
Does a high Pangram score mean a book was definitely written by AI?
No. Spero himself says the score is one factor among many and has stated that book-deal cancellations have not been driven by the score alone. A high score should prompt further investigation, not immediate action.
Why is Substack adding AI-detection to its platform?
Substack says it wants readers to know what they are consuming. The company describes its position as "not anti-AI" but pro-transparency, meaning it is not banning AI-assisted writing, just labelling possible uses of it.
