Inside Pangram, the AI Detector Publishers Are Turning To — and Questioning

Pangram has become one of the most influential AI-detection tools in publishing, education, and recruiting, even as authors, researchers, and industry insiders warn that its judgments can be wrong, biased, or overstated. The Brooklyn startup’s rise matters because its scores are increasingly shaping whether books are published, prizes stand, and writers are publicly accused of using artificial intelligence.

What began as a little-known company with a small team and modest funding is now being used by publishers and platforms as a gatekeeper for originality in an era when AI-generated text is everywhere. But the deeper Pangram spreads, the more urgent the question becomes: how reliable can any detector really be when careers, contracts, and reputations are on the line?

How Pangram became a publishing power broker

Pangram did not arrive with the scale, brand recognition, or funding of the major AI companies it now helps police. Headquartered above a Popeyes in Brooklyn, the startup has roughly two dozen employees and has raised about $13 million so far. By Silicon Valley standards, that is tiny. Yet in the niche but consequential market for AI-text detection, Pangram has managed to position itself as a tool that editors, agents, platforms, and researchers increasingly look to for answers.

The company’s rise accelerated in the first half of the year after a series of public accusations involving novels and short fiction that were suspected of being generated or heavily assisted by AI. Those incidents turned Pangram from a specialist product into a name circulated across publishing circles, on social media, and in debates about authorship and disclosure.

Its pitch is simple: upload text and receive a percentage estimate of how much of it was likely generated by an AI model. In practice, that number has become a kind of informal verdict in disputes over originality. A high score can spark scrutiny. In some cases, it can help unravel a book deal or force an author into public defense mode.

Why the startup’s role is so controversial

The controversy comes from the fact that AI detectors are not neutral instruments in the way a ruler measures length. They make probabilistic judgments about language, style, and pattern — and those judgments can shape real-world outcomes. That makes them powerful, but also dangerous when used as if they were definitive proof.

Jane Friedman, an author and publishing expert, says many writers view AI-detection software with deep suspicion. In her view, the resentment can be as intense as the anger directed at the AI tools that produce the text being flagged.

“There is such distaste and anger at the AI detection software,” Friedman said, adding that many writers see these tools as nearly as harmful as the generative systems they are designed to catch.

That backlash reflects a broader tension across the creative industries. Writers want to know whether a manuscript was created honestly. But they also fear a false accusation could damage their reputation before they ever have a chance to respond.

Who built Pangram and why did they start it?

Pangram was founded by Max Spero and Bradley Emi, two Stanford-educated technologists who saw a business opportunity after ChatGPT made AI-generated text a mainstream concern in late 2022. They first launched the company in 2023 under the name Checkfor.ai and later rebranded to Pangram.

Spero, now 30, grew up in La Crescenta, California, where he became interested in programming and robotics. He later studied at Stanford, where he met Emi. After college, Spero worked at Google on FLoC, the ad-targeting system the company eventually abandoned amid privacy criticism, and then at autonomous vehicle company Nuro. Emi’s background includes time at Tesla and at the AI-biotech firm Absci.

After the explosive public adoption of generative AI, the pair saw a different kind of opportunity: not to create text, but to decide whether text was human-made. They joined a crowded market that already included companies such as Originality.ai, GPTZero, and Turnitin. But Pangram’s early test results and quick uptake in publishing and research helped it rise above the pack.

What does Pangram actually do?

Pangram is designed to estimate whether a passage of text was written by a human, an AI system, or some mix of both. The output is not a yes-or-no ruling but a percentage estimate that reflects the company’s best guess.

That estimate is produced using machine learning methods that the company says are built specifically for detection rather than generation. Spero says the model is trained using a process the company calls synthetic mirroring, in which human text is paired with AI-generated text that closely matches it. Pangram also says it uses hard negative mining, a technique that looks for false positives and incorporates them into training so the system can improve over time.

Spero argues that Pangram’s datasets are licensed and carefully assembled, in contrast to the aggressive data scraping often associated with large AI model training. He also says the company’s approach is less data-hungry than the major generative AI labs, making it easier to maintain more controlled training sets.

How Pangram says it differs from other detectors

Pangram’s central claim is that it focuses on trust, transparency, and specificity. The company says it plans to publish more technical information about how its models work, while many rivals remain comparatively opaque.

That pitch is important because detectors are often criticized for acting like black boxes: they issue confident-looking scores without enough explanation for users to judge whether the result is fair. Pangram says it wants to reduce that uncertainty by giving users more detail and finer-grained results over time.

• It scans text and returns an AI-likelihood estimate.
• It is used in publishing, education, recruiting, and legal work.
• Its team says it is building toward more granular detection of AI-assisted editing.
• It emphasizes licensed data and public technical reporting.

How the Shy Girl controversy changed everything

Pangram’s public profile changed sharply after speculation swirled around the self-published novel Shy Girl, which had been picked up for a traditional release by Hachette. The online conversation began on Reddit and YouTube, where users questioned whether the book had been written with the help of AI.

Ballard denied using AI, but Pangram’s CEO later posted that the manuscript tested as heavily AI-generated. That claim fed a widening dispute that eventually helped derail the book’s path to publication. The case became one of the first widely discussed examples of how AI detection could affect a commercial book deal.

From there, more public accusations followed. Pangram was used to assess a Modern Love installment in The New York Times, a Commonwealth Short Story Prize winner, the novel Daggermouth, and a thriller titled Call Me, I’ll Hide the Body. The numbers attached to those texts — from 60 percent to 100 percent AI-generated, according to Pangram — were widely repeated online and became part of the controversy itself.

Spero has said that Pangram’s role in these disputes is often overstated, arguing that any decision to cancel or challenge a book typically involves broader concerns than a single detection score.

Still, the optics matter. Once a detector is associated with a scandal, its output can quickly become a shorthand for guilt in the public imagination.

Why publishers and platforms are paying attention

Pangram’s momentum has been reinforced by the fact that publishers and writing platforms are under pressure to know what kind of content they are handling. Substack announced in late July that it was integrating Pangram so readers could see possible AI use more easily, reflecting a growing appetite for disclosure tools.

That move is especially notable because Substack has tried to frame itself as not anti-AI, but rather pro-transparency. In other words, it is less interested in banning AI than in making its use visible.

Major publishers, meanwhile, appear to be experimenting cautiously. Penguin Random House confirmed that editors may use approved detection tools as one factor among many when evaluating content, but said those tools are not decisive on their own. Simon & Schuster and HarperCollins declined to comment, while Hachette and Macmillan did not respond to inquiries.

The silence is telling. Publishing has long moved slowly on technology, and AI content is forcing the industry to confront questions it has not yet standardized: When does AI assistance become disqualifying? Who gets to decide? And what kind of evidence is enough?

What the data says about AI use in writing

There is no simple answer to how common AI use is among writers, but the numbers cited by researchers suggest it is not rare. A survey by Gotham Ghostwriters found that 61 percent of 1,481 working writers said they use AI tools in some capacity, while 7 percent said they had published AI-generated text.

Separately, Stony Brook computer science professor Tuhin Chakrabarty analyzed 14,419 self-published novels with Pangram earlier this year and found that nearly 20 percent returned substantial AI-detection scores. His work has been widely cited in the publishing conversation and played a major role in some of the accusations that followed.

Those figures do not prove widespread fraud, but they do show why detection tools have become more attractive to publishers, agents, and editors trying to separate human work from machine output.

Is Pangram trustworthy?

Pangram says its system is highly accurate, but the company’s own disclosures and outside research suggest caution is necessary. The reason is simple: detection is easier to market than to validate, especially when the target is language itself.

The company says its newer model produces false positives only 0.0041 percent of the time, lower than an earlier rate it had reported. Spero argues that Pangram is deliberately conservative: when a passage falls near the line, it tends to be labeled human rather than AI. That reduces false accusations, but it also means some AI-written text may go undetected.

That tradeoff is central to the debate. If a detector is too aggressive, it risks accusing innocent writers. If it is too cautious, it misses AI manipulation. Pangram says it has chosen to minimize false positives, but critics argue that no threshold can eliminate the structural weaknesses of the category.

Where Pangram struggles most

Even Pangram acknowledges limitations. It performs worse on shorter passages, especially anything under 100 words. That matters because the company’s free product provides only a limited number of credits per day, amounting to about 2,000 words, with a median input length around 350 words.

The company also says the same text may produce different results depending on whether it is scanned alone or inside a larger block. That is a major issue for book publishing, where suspicious lines or paragraphs are often evaluated out of context.

Researchers have also shown that AI detectors can be fooled or confused by light editing, “humanizer” tools, or stylistic quirks. A Notre Dame working paper found that Pangram’s 3.2 model often flagged lightly edited AI text as fully AI-written, but missed most AI text that had been run through a humanizing tool.

In short, Pangram may be useful as a signal, but not as a final arbiter.

Why false positives worry writers

The biggest fear among authors is not that AI detection will occasionally make mistakes. It is that a mistake can be public, permanent, and career-altering.

Publishing professionals say a suspicious score can trigger private conversations, stalled deals, and reputational fallout long before any formal defense is possible. Writers from marginalized groups worry that those risks are not evenly distributed.

Some critics argue that detection systems are more likely to misread the writing of non-native English speakers, neurodiverse authors, and others whose style does not resemble the model’s idea of “normal” prose. Researchers and industry observers say this bias concern deserves more attention because the people most likely to be flagged may also be the least powerful in the publishing chain.

Regina Brooks, president of the Association of American Literary Agents, said the industry needs to think carefully about fairness, especially for African American writers and other authors who may already face unequal scrutiny.

That concern is not hypothetical. The public scandals tied to Pangram have involved writers of color, adding another layer of sensitivity to a debate already charged with accusations of deception and gatekeeping.

How the publishing industry is responding

Reactions inside the industry are mixed. Some agents now use detection software as a screening tool. Others avoid discussing it publicly, even when they rely on it privately. Friedman said that many agents choose silence after rejecting a manuscript, which can leave authors guessing about the reason a deal fell apart.

One major literary agent who has worked with Pangram, Todd Shuster of Aevitas, said the tool became useful quickly after he was introduced to it. He said the software helped initiate difficult conversations with authors whose manuscripts registered as substantially AI-assisted. He also said he has asked some writers to rewrite in a more authentic voice.

Chakrabarty, the Stony Brook professor, now speaks publicly in defense of Pangram and says the company has supported his research with API credits. He argues that the detector should not be treated as the final word, but believes it can be a helpful input when combined with human judgment.

“Pangram should not be the de facto judgment,” Chakrabarty said. “But I think your own discretion coupled with Pangram’s judgment cannot be wrong.”

That view may reflect the middle ground many publishers are trying to reach: using detection to inform decisions without letting it dictate outcomes.

What is the company’s next move?

Pangram appears to be shifting from pure detection toward broader platform influence. Spero says the company wants more detailed outputs and finer-grained analysis, including the ability to spot light AI editing rather than just heavily generated passages.

That change would make the tool more useful for editors and publishers trying to determine whether a manuscript was machine-assisted in subtle ways. It would also make Pangram more central to the gatekeeping role it already occupies.

The company’s internal posture seems to be evolving as well. Former contractor Rod Breslau, who described himself as an online “attack dog,” said Pangram has moved away from simply chasing down every suspected AI user. The company now appears to expect AI use to become more accepted over time, even as it remains vocal about cases it views as deceptive.

The strategy suggests a long-term bet: not that AI will disappear from writing, but that readers and buyers will demand more disclosure about when and how it was used.

Timeline of Pangram’s rise

Can AI detectors really protect human writing?

AI detectors can help identify patterns, but they cannot settle authorship disputes on their own. Pangram’s growth shows how badly the publishing world wants a fast answer to a hard problem, yet the tool’s limitations show why no machine can fully adjudicate literary truth.

That tension explains the company’s appeal. It gives publishers a number they can act on. It also gives writers a number they may have to fight. In a culture increasingly shaped by machine-generated language, the desire for certainty is powerful — but certainty may be the one thing detectors cannot honestly provide.

For Pangram, that is both the business opportunity and the existential risk. The more the company becomes the industry standard, the more its mistakes will matter.

And as AI-generated prose becomes easier to produce, the pressure to detect it will only intensify. Pangram is betting that publishers, platforms, and readers will accept its judgments as an essential part of the new literary landscape. Critics are betting that the cost of that trust could be too high.

Key issues at a glance

• Pangram is being used to assess whether books and articles were written with AI help.
• Its scores have affected high-profile publishing disputes and book deals.
• Experts warn that false positives and bias can harm writers unfairly.
• Publishers are increasingly using detection tools, but rarely as the sole criterion.
• The company plans to expand into more granular AI-use detection.

In the end, Pangram’s story is not just about one startup. It is about an entire industry trying to decide how to police authenticity in a world where machine-generated writing is no longer a fringe concern, but a daily reality.