Pangram’s Emergence in the AI Landscape
Pangram, an innovative startup specialising in artificial intelligence, operates from an unexpected location above a Popeyes in Brooklyn, employing a modest team of 24 individuals. To date, the company has successfully raised $13 million, which accounts for approximately 0.0072% of OpenAI’s resources. Despite being relatively unknown until early this year, Pangram has established itself as a significant player in the realm of AI and written communication.
The company’s key proposition is its ability to estimate the percentage of AI involvement in any given text. This capability came to public attention earlier this year when speculations arose on platforms like Reddit and YouTube regarding the authorship of “Shy Girl,” a self-published novel acquired by Hachette. While the author, Mia Ballard, refuted these claims, Pangram’s CEO publicly asserted that a staggering 78% of the book had been generated by AI, prompting Hachette to cancel its publication.
This incident triggered a wave of similar allegations. Notably, the New York Times faced backlash when Pangram flagged a piece in its Modern Love column as having a 100% likelihood of being AI-generated. Further scrutiny unveiled that a Commonwealth Short Story Prize winner also achieved a 100% rating, while the novel “Daggermouth” was rated at 60%, and the thriller “Call Me, I’ll Hide the Body,” which sold for $2.4 million, received a staggering 97% score.
Integration of Pangram with Publishing Platforms
In late July, Substack announced its intention to integrate Pangram into its platform, allowing readers to swiftly identify potential AI usage in texts. However, this move has not been universally welcomed among writers. Jane Friedman, an author and publishing expert, notes a prevailing sentiment of resistance and frustration towards AI detection software, with some perceiving it as more malevolent than the AI corporations themselves.
Although Pangram remains relatively obscure outside the editorial and higher education circles, its profile is set to rise significantly, driven by the increasing demand to differentiate between the outputs of large language models (LLMs) and human writing. A pressing question now looms: to what extent can one place trust in this budding enterprise?
Understanding the Technology Behind Pangram
Max Spero, the 30-year-old co-founder and CEO of Pangram, joins our Google Meet discussion from his phone, a takeaway box in hand against a backdrop of skyscrapers. Dressed in a beige t-shirt, with short dark hair and a youthful demeanour, he appears momentarily distracted by his hectic schedule of job interviews. Recently, Pangram secured $9 million in funding and unveiled its latest model, Pangram 4, leading to a planned 25% increase in staff.
When asked about his upbringing in California, Spero seems less comfortable discussing personal details than the intricacies of Pangram’s technology or its involvement in literary controversies. He shares that he grew up in La Crescenta, a suburban area of Los Angeles, where he developed a passion for programming and joined his school’s robotics team. His path led him to Stanford, where he partnered with Bradley Emi, co-founder of Pangram.
Post-graduation, Spero joined Google, contributing to FLoC, a technology aimed at replacing third-party cookies by clustering Chrome users based on interests. Later, he worked at Nuro, a self-driving vehicle company, while Emi gained experience at Tesla and the AI biotechnology firm Absci. The launch of ChatGPT in 2022 prompted the duo to seize an opportunity to address a future dominated by AI-generated content.
Pangram’s Response to Literary Controversies
In 2023, they founded Checkfor.ai, which would eventually be rebranded as Pangram. At that time, numerous companies were vying for a share of the plagiarism detection market, including Originality.ai, GPTZero, and Turnitin. Pangram distinguished itself in early independent tests, establishing itself as a formidable competitor.
As our conversation progresses, Spero’s responses appear hesitant, punctuated by pauses not solely due to his meal. When I inquire about his hiring philosophy, a long silence ensues before he states, “Those who work at Pangram care deeply about the mission.” To detect AI-generated content, Pangram employs a method called “synthetic reflection,” where it generates text closely resembling human writing using LLMs. This process instructs its model on how AI composes text.
The company further utilises “hard negative mining,” identifying false positives within datasets to expand its training set. “All our datasets are properly licensed, which is relatively rare in today’s AI landscape,” Spero explains, clarifying that Pangram’s products require significantly less data than ChatGPT or Claude. “I don’t want to criticise AI companies, but I believe they’ve lost a lot of trust, especially in creative circles,” he adds. While Pangram serves various sectors, including education, law, and recruitment, creative writing remains its primary training text segment.
The Impact of ‘The Shy Girl’ on Pangram’s Reputation
Spero had previously flagged suspicious AI-generated writing on social media, so it was no surprise when he was tagged in a Reddit post and later received a PDF of Ballard’s manuscript. “I input it into Pangram, published the results, and later the New York Times sought my comments. I think people exaggerate Pangram’s role in the Shy Girl narrative.” A game of broken telephone ensued, with a Pangram account executive discussing the situation with a publishing industry analyst, which eventually reached the Times.
However, critics pointed out that Spero’s manuscript copy originated from a piracy website. When queried about this, he admitted, “Yes. I didn’t review it thoroughly. I didn’t read the entire PDF. I just input it directly into Pangram.” Since then, the company has found itself amidst ongoing controversy. After attributing a high probability of AI usage to the Commonwealth Short Story Prize winner, Pangram also analysed all winners since 2012, identifying three additional potential cases. Nonetheless, Spero downplays Pangram’s influence, particularly regarding cancelled publishing contracts. “To my knowledge, Pangram’s scores have never been the central issue in any editorial contracts. They are part of the discussion, but if you speak to anyone involved, you’ll see it’s just a small piece of a much larger puzzle.”
Pangram’s Role in the Evolving Landscape of Publishing
Our discussion shifts to Pangram’s integration with Substack. Spero argues that authors should not fear disclosing AI usage if their work is of high quality and resonates with readers. “If the value of your work hinges on deceiving the reader into believing it wasn’t written by AI, then that’s going to be a problem.” Substack, however, has refrained from confirming specific details about its collaboration with Pangram or whether it compensates the company for each analysis. “Substack’s philosophy is not anti-AI. We simply believe you should know what you are consuming,” a representative states. Some authors express distrust, fearing that their careers could be jeopardised with just a click of a button.
Nevertheless, the reality remains that an increasing number of authors, publishers, and agents are utilising AI. A survey conducted last year by Gotham Ghostwriters among 1,481 active writers revealed that 61% employed AI tools, while 7% reported having published AI-generated texts. Professor Tuhin Chakrabarty from Stony Brook University utilised Pangram earlier this year to analyse 14,419 self-published novels, discovering that nearly 20% exhibited significant scores in AI detection. His working paper has been widely referenced, serving as the original source for allegations against Daggermouth.
The Reaction of Major Publishers to AI Detection
In an attempt to understand the stance of major publishing houses, I reached out to the five largest publishers regarding their use of AI detection tools. Simon & Schuster and HarperCollins declined to comment; Hachette and Macmillan did not respond. A spokesperson for Penguin Random House confirmed that their editors may use approved AI detection tools “as an additional means to identify potential AI-generated content,” but emphasised that these tools are not decisive and represent merely one element of a broader editorial process. Some agents also employ detection tools, and for every public publishing deal, several are discreetly discarded behind the scenes.
I was keen to learn more about Chakrabarty’s perspective. He first heard of Pangram through one of Spero’s posts in late 2024. After registering, Spero provided him with API credits. Since then, they have become close friends, with Chakrabarty noting that Pangram continues to support his research. He shares Pangram’s findings on social media and frequently advocates for the company. He mentions having met with publishers to discuss detection in light of the AI-related scandals. “Pangram shouldn’t be the definitive criterion.”
Challenges and Limitations of AI Detection
Chakrabarty also maintains a relationship with Todd Shuster, co-executive director of Aevitas, a prominent literary agency in New York. Chakrab