Pangram, a lean AI startup with a team of 24 employees, operates from a rather unconventional headquarters situated above a Popeyes in Brooklyn. Despite having raised $13 million to date—a figure that pales in comparison to the vast resources of industry giants like OpenAI—Pangram has rapidly emerged from relative obscurity to become a central figure in the escalating debate surrounding artificial intelligence and the integrity of written content. The company’s bold claim is that it stands as a crucial bulwark against the unchecked proliferation of AI-generated text, particularly within the creative realm.
The core of Pangram’s offering is its ability to estimate the percentage of AI involvement in any given piece of text. This sophisticated analysis, ironically, is powered by AI itself, presenting its findings as a probabilistic best guess. The company’s ascent to prominence was significantly amplified by a high-profile controversy in January. Speculation, fueled by discussions on platforms like Reddit and YouTube, suggested that author Mia Ballard may have employed AI to write her self-published novel, Shy Girl. While Ballard publicly denied these allegations, Pangram’s CEO, Max Spero, took to X (formerly Twitter) to assert that the book was 78 percent AI-generated. This declaration had immediate repercussions, leading Hachette, the publisher that had picked up the novel for traditional release, to cancel its publication.
This incident marked the beginning of a series of high-profile accusations of AI authorship. Soon after, The New York Times faced scrutiny for an installment in its esteemed Modern Love column, which Pangram’s analysis scored at a full 100 percent AI-generated. The controversy continued to ripple through the literary world with the Commonwealth Short Story Prize winner being flagged at 100 percent AI, the novel Daggermouth receiving a 60 percent score, and a thriller, Call Me, I’ll Hide the Body, which had commanded a $2.4 million advance, being assessed at 97 percent AI. In a significant development in late July, Substack, a popular platform for independent writers and newsletters, announced its integration of Pangram’s technology, aiming to provide readers with a tool to quickly assess potential AI usage in the content they consume.
However, this rapid rise and the aggressive application of AI detection have not been met with universal acclaim. Jane Friedman, a seasoned author and publishing expert, observes a palpable "distaste and anger at the AI detection software" among writers. She articulates a sentiment that "they are just as evil, if not more evil, than the AI companies themselves." While Pangram may not yet be a household name outside of academic and publishing circles, its increasing visibility suggests a future where distinguishing authentic human expression from the output of large language models (LLMs) becomes a critical concern. The fundamental question that now looms is the extent to which Pangram and its pronouncements can be reliably trusted.
The Architect of Detection: Max Spero and Pangram’s Origins
Max Spero, the 30-year-old co-founder and CEO of Pangram, exudes an energetic, albeit sometimes harried, demeanor. During an interview conducted via Google Meet, he was initially dialing in from his phone, juggling a takeout box and the backdrop of Brooklyn’s cityscape. His attire—a simple beige T-shirt—and his boyish expression masked a sharp business acumen. He candidly admitted to rushing home with lunch before reconnecting for a more focused discussion. In the ensuing conversation, Spero revealed the relentless pace of his company, detailing ongoing hiring calls that were poised to increase Pangram’s headcount by a significant 25 percent, following a $9 million funding round and the launch of their latest model, Pangram 4, in July.
While Spero proved more comfortable discussing the intricacies of Pangram’s technology, personal inquiries, particularly those touching on his upbringing, elicited a degree of discomfort. He eventually shared that he was raised in La Crescenta, a suburb of Los Angeles, where his early fascination with programming led him to join his high school’s robotics team. His undergraduate years at Stanford provided the crucial connection with his future co-founder, Bradley Emi. The company declined multiple interview requests for Emi. Post-graduation, Spero joined Google, contributing to the development of FLoC (Federated Learning of Cohorts), a controversial initiative aimed at replacing third-party cookies by grouping Chrome users based on their interests—a project ultimately shelved in 2022 due to privacy concerns. Emi’s career path included stints at Tesla and the AI biotech firm Absci. The widespread impact of ChatGPT’s launch in late 2022 presented the duo with a clear business opportunity: to address the burgeoning challenge of AI-generated content. They initially founded Checkfor.ai in 2023, rebranding to Pangram a year later. By this time, the AI detection landscape was already becoming crowded, with over a dozen competitors, including Originality.ai, GPTZero, and Turnitin. Pangram, however, distinguished itself through promising early performance in independent testing, quickly positioning itself as a front-runner.
Decoding AI: Pangram’s Methodologies and the Ethics of Detection
Spero’s interview style, characterized by pauses and shifts in focus, occasionally mirrored the very AI generation challenges his company aims to identify. When questioned about his hiring philosophy, he offered a brief "hmm" before stepping away to reheat his meal, returning after a significant silence to state, "The average person is at Pangram because they care about the mission." This emphasis on mission over mere technical proficiency suggests a deliberate effort to cultivate a team aligned with the company’s purported objective of safeguarding content integrity.
Pangram employs a sophisticated technique known as "synthetic mirroring." This process involves feeding human-written text into LLMs to generate closely analogous AI-generated versions. By analyzing these generated texts, Pangram’s models learn to recognize the distinct patterns and characteristics of AI writing. Complementing this is "hard negative mining," a strategy that involves actively searching datasets for instances where the AI detector might incorrectly flag human text as AI-generated. These false positives are then synthetically mirrored and incorporated into the training set, essentially using mistakes to refine the machine’s accuracy. Spero highlights that Pangram’s datasets are "properly licensed," a claim he acknowledges is not universally true in the AI industry, attributing this to the fact that his product requires significantly less data than more broadly focused LLMs like ChatGPT or Claude. He expresses a nuanced view on other AI companies, stating, "I don’t want to completely throw the AI companies under the bus, but I think they’ve lost a lot of trust, especially in the world of creatives." While Pangram serves various sectors including education and legal services, creative writing remains the largest segment for its training data.
The Shy Girl Spark and Subsequent Controversies
The Shy Girl incident served as a pivotal moment for Pangram, propelling it into the public consciousness. Spero, who had already been vocal on social media about suspected AI-authored content, was alerted to the Shy Girl manuscript through a Reddit post and subsequently received a PDF copy. He recounted, "I put it in [Pangram], I posted it, and then later The New York Times asked me for comment." However, the narrative of Pangram’s direct involvement in bringing the story to the Times’ attention is more complex. A Pangram account executive had reportedly discussed the case with a publishing industry analyst, who then approached The New York Times. This chain of communication underscores how a company’s influence can be amplified through indirect channels.
Further complicating matters, critics, including the investigative project The Drey Dossier, pointed out that Spero obtained his copy of the manuscript from a pirated website. When confronted with this fact, Spero responded with a degree of nonchalance, stating, "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." This admission raises questions about due diligence and the ethical sourcing of material used for public pronouncements.
Pangram has continued to be associated with controversy. Following the Commonwealth Short Story Prize winner’s high AI score, the company conducted an analysis of all winners since 2012, identifying three additional potential instances of AI authorship. Despite these high-profile cases, Spero often downplays Pangram’s direct role in the cancellation of book deals. He asserts, "Basically, with every book deal, to my knowledge, it hasn’t really been about the Pangram score. That is a part of it, but if you talk to anyone involved, it is only a small, small part of the bigger picture."
The integration with Substack represents a strategic move to embed Pangram’s detection capabilities directly into the reader’s experience. Spero’s perspective on this partnership is that "People shouldn’t be afraid about disclosing this because your work should stand on its own as quality and something that people want to read, regardless of… Well, how do I want to put that? I think I… oh, I know what I said the other day. If the value of your work is dependent on deceiving the end user into thinking it wasn’t written by AI, then… that’s going to be a problem." This statement, delivered with a noticeable shift in intonation, echoed a post he had made on X. Substack, while confirming the partnership, remained tight-lipped about the financial specifics, emphasizing their philosophy: "Substack’s philosophy is not anti-AI. We simply believe you should know what you’re consuming." Nevertheless, some authors express deep apprehension, feeling that an entire career could be jeopardized by a single algorithmic assessment.
The Publishing Industry’s AI Reckoning
The book publishing industry, historically characterized by its deliberate pace, has been particularly slow to grapple with the implications of AI. Yet, the reality is that AI tools are being adopted by a growing number of authors, publishers, and literary agents. A survey conducted by Gotham Ghostwriters last year, polling 1,481 professional writers, revealed that 61 percent use AI tools, with 7 percent admitting to having published AI-generated text. Tuhin Chakrabarty, an assistant professor of computer science at Stony Brook University, whose research has been influential in the AI detection discourse, analyzed 14,419 self-published novels using Pangram. His findings indicated that nearly 20 percent exhibited substantial AI-detection scores. This data served as the original basis for the accusations leveled against Daggermouth.
Inquiries to the "Big Five" publishers regarding their use of AI detection yielded mixed responses. Simon & Schuster and HarperCollins declined to comment, while Hachette and Macmillan did not respond. Penguin Random House confirmed that its editors may utilize approved AI-detection tools "as one additional means of identifying potential AI-generated content," but stressed that such tools are "not determinative and are only one component of a broader editorial process." Literary agents are also increasingly employing these tools, with many book deals being quietly terminated behind the scenes, as opposed to public declarations. As Friedman notes, "Normally, agents say nothing."
Allies and Critics: Navigating the AI Detection Landscape
Tuhin Chakrabarty’s perspective on Pangram is noteworthy. He first learned of the company through Spero’s social media posts in late 2024 and subsequently received API credits from Pangram to support his research. Chakrabarty describes himself and Spero as "close friends" and Pangram continues to provide him with resources. He actively shares Pangram’s results on social media and has become a vocal defender of the company, even meeting with publishers to discuss AI detection in the wake of recent scandals. He cautions, however, that "Pangram should not be the de facto judgment. But I think your own discretion coupled with Pangram’s judgment cannot be wrong."
Adding another layer of complexity, Chakrabarty is in a relationship with Todd Shuster, the co-CEO of Aevitas, a prominent New York literary agency. Chakrabarty encouraged Shuster to engage with Pangram, describing it as "instantly a very helpful tool." Shuster confirmed that his agency began "having to have difficult conversations [with authors] because Pangram was showing their works to be anything from 50 percent AI written to 95 percent." Aevitas now uses Pangram for manuscript and proposal analysis and Shuster also consults for Pangram, facilitating introductions to publishers. He observes that authors react to AI accusations with a range of responses, from defensiveness to honesty, and has encouraged some to rewrite their work to better reflect their own voice.
The Perils of False Positives and Algorithmic Bias
A significant concern raised by critics of AI detection, including Pangram, centers on the potential for false positives and inherent biases within machine learning models. Sam Illingworth, a professor of critical AI literacy at Edinburgh Napier University, asserts, "I don’t think [AI detectors] work. I think that detectors are prejudiced against certain people." Research has indicated that AI detectors may be more prone to flagging the work of non-native English speakers as AI-generated, and neurodiverse writers have also reported disproportionate flagging of their writing patterns. While the study Illingworth references predates Pangram, Spero points to internal research that he claims counters these findings.
The three novels previously mentioned—Shy Girl, Daggermouth, and Call Me—which represent major AI-detection scandals in publishing, were all authored by writers of color. Attempts to reach these authors for comment were met with non-response from two, while the third declined, citing legal concerns. Regina Brooks, president of the Association of American Literary Agents, highlights the critical need to address "inequities within the industry," emphasizing that the selection of whose work is scanned and who faces public accusation is an inherently human process, thus susceptible to bias.
In instances where writers have claimed wrongful AI flagging, Spero has reportedly offered cash bounties for proof of original authorship. He states, "I’ve offered the bounty in cases where I’m almost certain the person is lying. Or, if they’re not lying, then it’s very valuable for us to know." To date, no one has accepted this offer.
The Evolving Frontier of AI Detection
The efficacy of AI detection tools remains a subject of ongoing debate and research. A recent working paper from Notre Dame, titled "Why AI Detection Fails for Academic Integrity," suggests that Pangram’s 3.2 model flagged heavily edited academic abstracts as AI-generated with a high degree of accuracy (64 to 80 percent). However, when AI-generated text was passed through a "humanizer"—a tool designed to make AI output sound more natural—Pangram’s detection rate dropped significantly, catching AI writing less than 4 percent of the time.
Pangram acknowledges that its tool performs less effectively on shorter text passages, particularly those under 100 words. This limitation is amplified by the fact that Pangram’s free service offers a limited number of daily credits, equating to approximately 2,000 words, with a median input of 350 words. The company also concedes that the same passage can yield different results depending on whether it’s scanned in isolation or as part of a larger document, a scenario that could occur if an excerpt is copied and pasted.
Despite these challenges, Pangram claims its latest model achieves a false positive rate of just 0.0041 percent, an improvement from a previous report of 0.01 percent. Spero maintains that Pangram errs on the side of caution, classifying borderline texts as human-written. "This is an intentional choice that we made," he states, contributing to the low false-positive rate, but also potentially allowing more AI-generated work to go undetected. Even with heavily AI-rewritten human essays, Pangram’s own metrics suggest it still identifies them as human-written 41.37 percent of the time.
Transparency as a Strategy
During the interview, Spero demonstrated a rapid-fire communication style, frequently sharing social media posts to illustrate his points. Rod Breslau, a journalist and former Pangram contractor, described Spero as "personally online all day. He is going to reply to your tweet, he is going to reply to your post on Reddit. He’ll probably reply to your post on LinkedIn." Breslau, who initially offered his services as an "attack dog" to combat the perceived flood of AI content, worked with Pangram on social media strategy and securing interviews, while also publicly calling out suspected AI users. His motivation was to "take a harsher approach because I thought that people were getting off too easy."
Pangram ultimately parted ways with Breslau as its strategy evolved. The company now anticipates that AI use will become more normalized. Spero appears to be shifting focus from aggressively pursuing all AI users to targeting instances of "deception." He indicates that Pangram aims to flag even light AI editing in the future, stating, "That’s the road map. Higher granularity, better detail, provide more information to the end users." The overarching ambition, it seems, is for Pangram to become the definitive arbiter of originality across all sectors.
Pangram’s small team includes former educators and publishing professionals, with a spokesperson noting that one researcher holds a degree in English literature. While there have been hints of potential defamation suits from accused authors, Spero states that "I don’t believe anything has come to pass" regarding lawsuits. He plans to continue publishing technical reports on Pangram’s models and training processes. "Everybody else is sort of this black box, so I think our strategy here is really just build trust through transparency, be as open as possible, and engage with the community online as much as possible."
As the conversation concluded, Spero’s engagement level appeared inconsistent, marked by a peripatetic movement around his apartment and a sometimes disjointed presentation. These observations raise questions about the demands of media coverage, the inherent challenges of conducting interviews, and the potential for human ambiguity to confound even the most sophisticated detection systems. Pangram’s objective is to distill complex communication into a quantifiable number, yet human expression, with its inherent nuances and variables, remains far more elusive.
