What happened
An AI text detector called Pangram is now deciding whether published writing counts as human-made, and the fallout has already reached major publishers. The Brooklyn-based startup, founded in 2023 with just 24 employees and $13 million in funding, built a model that scores any piece of text with a best-guess percentage of AI involvement. In January, Pangram's CEO posted on X that Mia Ballard's self-published novel Shy Girl was 78 percent AI-generated, shortly after Reddit and YouTube users raised the same suspicion. Hachette, which had acquired the book for traditional release, canceled publication. The company denied it used AI.
That wasn't an isolated case. Pangram scored a New York Times Modern Love column at 100 percent AI-generated, a Commonwealth Short Story Prize winner at 100 percent, the novel Daggermouth at 60 percent, and a thriller that sold for $2.4 million, Call Me, I'll Hide the Body, at 97 percent. In late July, Substack announced it was integrating Pangram directly into its platform so readers can check any post for AI involvement in a click. The company also just raised $9 million and released a new model, Pangram 4, and it's hiring six new roles that would grow its headcount by 25 percent.
### Why the timing matters
Pangram's rise tracks a broader shift: AI-generated text is now common enough in fiction, journalism, and student writing that publishers, editors, and platforms want a fast way to flag it before it reaches readers.
Why it matters
For publishers and editors, an AI detector functioning as a de facto gatekeeper changes the economics of trust. A single score from a 24-person startup can cancel a book deal, as it did with Shy Girl, even when the accused author denies the claim. Jane Friedman, an author and publishing expert, describes the backlash bluntly: "There is such distaste and anger at the AI detection software... there's this feeling like they are just as evil, if not more evil, than the AI companies themselves."
The stakes are commercial as well as reputational. Call Me, I'll Hide the Body sold for $2.4 million before Pangram flagged it at 97 percent AI-generated — a score that, if it sticks, could unwind a seven-figure deal. And Pangram isn't operating alone: at least a dozen competitors, including Originality.ai, GPTZero, and Turnitin, are chasing the same market, which means multiple tools could soon be scoring the same manuscript with different results.
### The trust problem
No AI detector, Pangram included, publishes a peer-reviewed false-positive rate that publishers can rely on before making a cancellation decision. That gap is the real story: a percentage score is being treated as verdict, not as evidence.
How to use it today
If you publish long-form content, treat an AI-detection score as a signal to investigate, not a final ruling. Before you submit a manuscript, article, or client deliverable, run your own draft through a detector so you know what a reviewer or platform will see. Substack authors can now do this natively; everyone else can use a standalone checker.
Writers and marketers who use AI as part of their process — for outlines, first drafts, or editing — should keep records of their drafting stages (notes, revision history, prompts used) so they have evidence to counter a disputed score. This matters most in traditional publishing and academic settings, where a single flagged score can trigger a cancellation or an integrity review before anyone reads the actual text.
For teams producing AI-assisted content at scale — blog posts, product copy, SEO pages — the more practical move is to build a workflow where a human editor substantially rewrites and fact-checks AI output before publication, rather than trying to game a detector after the fact. Tools like the free AI utilities at [mykreatool.com](https://mykreatool.com) can help with drafting and editing steps, but the final published version should always carry a human editorial pass, both for quality and to reduce the odds of a high AI-detection score.
### Checking your own content
Run a sample paragraph, not just a title or summary, since detectors score based on sentence-level patterns. A short excerpt can return a misleadingly high or low score compared to the full piece.
Who benefits
Publishers and literary agents get a fast pre-screening step that used to rely on reader tips and gut instinct. Substack's integration means every writer on the platform can self-check before publishing, which could reduce disputes after the fact. Academic institutions, which already used tools like Turnitin for plagiarism, now have a parallel option for AI-generated submissions.
Pangram itself benefits from being first to dominate headlines: its $9 million July raise and 25 percent headcount increase suggest investors see detection as a durable business, not a fad tied to one news cycle.
Risks
The central risk is treating a probability score as proof. Ballard denied using AI, yet the accusation alone was enough for Hachette to cancel the release — a decision made on a percentage, not a forensic audit. False positives carry real professional and financial consequences: a wrongly flagged author can lose a book deal, a student can face an integrity charge, and a freelance writer can lose a client, all based on one company's model output.
There's also a concentration risk. With Substack now baking Pangram into its platform and competitors like Originality.ai and GPTZero fighting for the same niche, a handful of largely unaudited startups are gaining outsized influence over what counts as "real" writing, without independent, published accuracy benchmarks that publishers or writers can check against.
Conclusion
Pangram's growth — from an unknown 24-person startup to a tool now built into Substack and cited in Hachette's decision to cancel a novel — shows how quickly AI detection has moved from a curiosity to a gatekeeping function in publishing and media. The technology is useful for flagging content worth a closer look, but the Shy Girl case shows the danger of treating a percentage score as a verdict. Writers, editors, and platforms adopting these tools should use them to start a conversation about a piece of writing, not to end one.
Comments 0