Inside Pangram, the AI Detector Disrupting Publishing and Triggering Debates Over Accuracy and Bias From canceled book deals to Substack integrations, Pangram claims to flag AI-generated text, but growing scrutiny highlights risks of false positives, bias, and technical limitations. Pangram, a software tool designed to evaluate how much artificial intelligence contributed to any piece of text, has swiftly positioned itself at the center of a growing battle over digital authenticity. By returning a best-guess percentage of AI involvement, the tool has gained widespread attention across publishing, media, and academia. In January, online speculation surfaced regarding author Mia Ballard's self-published novel Shy Girl, which had been acquired for traditional publication by Hachette. Despite Ballard's denials of using AI, Pangram's chief executive officer posted on X that the manuscript was 78 percent AI-generated. Shortly thereafter, Hachette canceled the book's release. High-Profile Scandals and Industry Fallout The cancellation of Shy Girl marked the beginning of a series of high-profile accusations driven by Pangram scores. The New York Times faced public scrutiny over an installment of its Modern Love column, which Pangram flagged as 100 percent AI-generated. Similar flags followed for a winning entry in the Commonwealth Short Story Prize (100 percent), the novel Daggermouth (60 percent), and a high-profile thriller titled Call Me, I’ll Hide the Body that had secured a $2.4 million deal (97 percent). By late July, Substack announced an integration with Pangram, enabling readers on the platform to check whether published newsletters or articles contained AI-generated text. The integration met mixed reactions from creators. Jane Friedman, an author and publishing industry expert, noted widespread frustration among writers, stating that many view AI detection tools with the same level of hostility as the underlying AI generation tools themselves. While Pangram was initially unknown outside specialized circles, increasing demand to separate human writing from large language model outputs has propelled the company into mainstream discussion. The Origins of Pangram and Founders' Background Pangram co-founder Max Spero developed an early interest in computer science while growing up in the Los Angeles suburb of La Crescenta, participating in competitive robotics before attending Stanford University. At Stanford, Spero met Bradley Emi, who later became Pangram's co-founder. Following graduation, Spero joined Google, working on the Federated Learning of Cohorts (FLoC) initiative aimed at replacing third-party cookies by categorizing Chrome users by interest. Google ultimately phased out FLoC in 2022 following privacy concerns. Spero subsequently worked at autonomous vehicle developer Nuro, while Emi held roles at Tesla and AI biotechnology firm Absci. Following the public release of ChatGPT in late 2022, Spero and Emi recognized a commercial need for tools capable of detecting machine-generated content. In 2023, they established Checkfor.ai, renaming the venture Pangram a year later. Despite entering a crowded market alongside established players such as Originality.ai, GPTZero, and Turnitin, Pangram gained traction following early performance tests. Synthetic Mirroring and Training Mechanics Pangram detects AI content using a technique referred to as synthetic mirroring. The system takes human-authored text and instructs large language models to produce close approximations, training the detection model on the structural patterns characteristic of AI output. In addition, Pangram employs hard negative mining, identifying false positives within datasets and synthetically mirroring them to refine model accuracy. Spero states that Pangram relies exclusively on properly licensed datasets, noting that the tool requires significantly less data than generative models like ChatGPT or Claude. While Pangram serves clients across education, legal compliance, and recruitment, creative writing constitutes the largest segment of its training corpus. Controversies Surrounding Detection and Pirated Files The controversy around Shy Girl brought Pangram into the public spotlight after Spero received a PDF of Ballard's manuscript via Reddit. Spero processed the file through Pangram and published the results online, leading to media inquiries. Critics, including investigative outlet The Drey Dossier, later revealed that the manuscript copy used for the scan had been sourced from an unauthorized file-sharing site. Spero acknowledged that he had not examined the origin of the file before running the scan. Pangram subsequently expanded its analysis to other literary works. Following the flag on the Commonwealth Short Story Prize winner, the company analyzed past winning entries dating back to 2012, alleging potential AI involvement in three additional cases. Spero has maintained that Pangram scores represent only one factor in publisher decisions, rather than the sole cause of canceled contracts. Substack Integration and Publisher Reaction Substack's decision to integrate Pangram reflected a policy focused on disclosure rather than restricting AI use. Representatives from Substack emphasized that the goal was to provide transparency to readers regarding content origin. However, several authors expressed concern that automated scans could inflict lasting damage on professional reputations based on unverified scores. Responses among major publishing houses remain varied. Major publishers including Simon & Schuster and HarperCollins declined to comment on their use of detection tools, while Hachette and Macmillan provided no public response. A spokesperson for Penguin Random House confirmed that editors may utilize approved AI detection tools as one supplementary component within a broader editorial review process, emphasizing that detector scores are not considered definitive. Meanwhile, literary agents, including Todd Shuster of Aevitas, have adopted Pangram to evaluate manuscript submissions, using flagged scores ranging from 50 to 95 percent to initiate discussions with authors. Academic Criticism, Bias, and Technical Limitations Researchers and academics have raised concerns regarding the reliability and potential bias of AI detection software. Sam Illingworth, a professor of critical AI literacy at Edinburgh Napier University, pointed out that detection models frequently exhibit higher error rates when evaluating text written by non-native English speakers or neurodiverse writers. Concerns have also been raised regarding demographic disparities, as several high-profile public accusations involved authors of color. Regina Brooks, president of the Association of American Literary Agents, stressed the need to monitor how detection tools are deployed within the industry. Technical evaluations further highlight operational limits. A working paper from the University of Notre Dame, titled "Why AI Detection Fails for Academic Integrity," found that Pangram's 3.2 model flagged academic abstracts undergoing minor AI editing as AI-written in 64 to 80 percent of tests. Conversely, when AI-generated text was processed through text-humanizing software, Pangram's detection rate dropped below 4 percent. Pangram acknowledges that its software is less reliable when analyzing short passages, particularly those under 100 words. The platform's free tier imposes a limit of approximately 2,000 words per day, with a median submission length of 350 words. The company also notes that analyzing excerpts in isolation can produce different percentage scores compared to analyzing the same passage within a full-length text. False Positive Rates and Systemic Bias Pangram claims a false-positive rate of 0.0041 percent for its current model, down from a previously reported 0.01 percent. Spero states that the system is calibrated conservatively, favoring human attribution in ambiguous cases. As a result of this conservative threshold, human essays heavily edited by AI models are categorized as human-written in 41.37 percent of test cases. Looking ahead, Pangram is adjusting its development strategy. Former contractor Rod Breslau noted an initial emphasis on active social media engagement and public flagging of suspected AI text. Pangram has since shifted toward developing tools capable of identifying subtle AI editing assistance, aiming to provide detailed breakdowns for educational institutions, publishers, and enterprise users. What this means for you The growing adoption of Pangram and AI detectors directly impacts digital content creation, academic integrity, and publishing workflows. • Across India: Students and professionals using AI for light editing risk false-positive flags on academic papers and professional submissions. • For Global Writers: Non-native English writers face disproportionate risk of their authentic work being wrongly flagged as AI-generated. • For Authors & Freelancers: Integrations with platforms like Substack put publishing deals and professional reputations at risk based on automated scores. • For Readers: Increased transparency allows consumers to verify whether articles and books relied on AI generation. Questions & Answers 1. What is Pangram and how does it detect AI text? Pangram is an AI detection software that uses synthetic mirroring to identify structural patterns of large language models and output a percentage score. 2. Why was the book deal for Shy Girl canceled? Hachette canceled the release of Shy Girl after Pangram posted that the manuscript was 78 percent AI-generated. 3. Is Pangram fully accurate in detecting AI writing? No, academic studies show Pangram can produce false positives on lightly edited text and miss AI content processed through humanizer tools. 4. Has Substack integrated Pangram into its platform? Yes, Substack integrated Pangram in late July to allow readers to check for potential AI usage in published posts. https://trendkia.com/en/ai/sahityika-duniya-men-tahalaka-machane-vala-ai-ditektara-pangram-kitana-satika-hai-janie-isaki-puri-sachchai-26431 TrendKia — Har trend, sabse pehle.