Pangram’s CEO wants to expose everyone’s AI slop habit

Pangram Labs, a Brooklyn-based startup whose browser extension and website detect AI-generated text, has identified AI-written content in an award-winning short story, a popular novel, a billionaire's Wall Street Journal column, and Pope Leo XIV's posts on X, according to CEO Max Spero. The company, which raised $9 million in July in a round led by Menlo Ventures, is expanding into AI image and video detection and partnering with Substack and NewsGuard to filter AI content at scale.

Installing Pangram in your web browser feels a bit like having X-ray vision. Suddenly you can see how much artificial intelligence https://www.fastcompany.com/section/artificial-intelligence slop is quietly coursing through the modern internet. Pangram knows which of your LinkedIn connections have been posting with Claude or ChatGPT, and how many viral Reddit posts are betraying the site’s claims of being authentically human. Maybe you had suspicions—too many em dashes here, overused constructions like it’s not x, it’s y there—but Pangram delivers the closest thing to actual proof, marking every post as “AI,” “Human,” or something in between. While AI writing detectors have been around for as long as ChatGPT itself, Pangram’s technology routinely beats other AI detectors https://www.pangram.com/blog/third-party-pangram-evals in independent testing, lending credibility to its role in various recent writing scandals. An award-winning short story https://lithub.com/a-prize-winning-story-published-in-granta-was-very-likely-written-by-ai/ , a popular novel https://x.com/max spero /status/2014791270137249995 , a billionaire’s Wall Street Journal column https://www.notus.org/media/stanley-druckenmillers-wsj-op-ed-bessent-ai , and even Pope Leo XIV’s posts on X https://www.wired.com/story/pope-tweets-ai-generated-pangram-chrome-extension/ are among Pangram’s recent marks of shame. The Brooklyn-based startup, whose full name is Pangram Labs, raised $9 million in July https://techcrunch.com/2026/07/29/as-ai-content-floods-the-internet-pangram-raises-9m-to-detect-it/ in a fundraising round led by Menlo Ventures and is expanding into AI image and video detection https://www.pangram.com/blog/introducing-pangram-image-detection as well. It has a couple dozen employees and is hiring https://www.fastcompany.com/section/hiring more. Max Spero, Pangram’s cofounder and CEO, says Pangram’s goal is partly to arm people for a world of shifting social norms. Is it okay for AI to write your newsletter? How about drawing your holiday greeting cards? Should it be drafting your personal emails and text messages, too? Before we can answer those questions, we need to be certain about what’s AI in the first place. “If you don’t have an accurate AI detector, then there’s no way to express displeasure that somebody else is sending you AI content,” Spero says. “Having Pangram as a tool helps people say, ‘Don’t let AI replace your voice.'” Spero sometimes serves as the instigator of that discontent, pointing out when viral posts https://x.com/max spero /status/2056836817173786708 or high-profile accounts https://x.com/max spero /status/2069863726623101303 might be entirely relying on AI. He has taken to calling himself a “slop janitor,” arriving at the job by way of Google where he worked on machine learning “to figure out which ads someone is most likely click on” and the autonomous vehicle platform Nuro. As he and the other cofounder, Bradley Emi, pondered AI’s potential ramifications a few years ago, they agreed on AI detection as both an important need and an interesting problem to solve. “This is once-in-a-generation technology that’s going to change society, so we need to be building for the world when that happens,” Spero says. But Pangram isn’t just about outing individual AI users through its $20 per month website and browser extension. It’s also trying to automate AI at scale by weaving itself into the fabric of the web. The newsletters platform Substack, which recently added a Pangram-powered “ Scan for AI text https://support.substack.com/hc/en-us/articles/50891130623508-How-can-I-detect-AI-on-Substack ” button to its articles, is considering ways to filter AI writing from users’ recommendations. NewsGuard, which helps advertisers and news consumers identify reliable content sources, is tapping Pangram to detect AI content farms https://www.newsguardtech.com/press/newsguard-launches-real-time-ai-content-farm-detection-datastream-to-counter-onslaught-of-ai-slop-in-news/ . The online exam platform Inspera is using Pangram to sniff out AI in student essays https://inspera.com/press-releases/partnership-with-pangram/ . Pangram’s real goal isn’t merely to facilitate some high-profile gotchas, but to be an arbiter of what’s human online. But as tools like ChatGPT start to sound more human, and as humans themselves adopt the AI-isms they’re so often exposed to, those boundaries may only get harder to define. Being responsible for those decisions is a big job for a small startup. “So you accuse somebody. Now what?” says Patrick Traynor, an engineering professor at University of Florida who’s studied AI detectors. “What do you do with that information? What if you’re wrong?” Pangram’s approach to AI detection is different than those from few years ago, including OpenAI’s own failed endeavor https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/ . To identify human writing, those early tools looked for signs of messiness and unpredictability. As Spero explains, AI might write “The boy ate a bowl of soup,” but a human is more likely to go off-script with something like “The boy ate a bowl of spiders.” String enough surprises together, and you can get a rough measure of a text’s humanity. On the downside, those early detectors were too eager to dismiss predictable human writing as AI, especially when it resembled the data that large language models trained on. An early iteration of GPTZero, for instance, famously flagged the U.S. Constitution as AI-generated https://arstechnica.com/information-technology/2023/07/why-ai-detectors-think-the-us-constitution-was-written-by-ai/ , and Spero notes that nonnative English speakers were often victims of false positives. These days, most AI detectors focus on spotting the tells in AI writing. They accomplish this by feeding large amounts of human- and AI-generated text into machine learning models, then training them to understand the differences. This works better overall, but it also leads to a different kind of false positive, in which the model becomes suspicious of writing styles that fall outside its training data. Pangram’s innovation, and for now the key to its superior accuracy, comes from teaching its models to recognize how AI detection usually fails. If its models flag a document from before 2022 as AI-generated, Pangram creates a “synthetic mirror” of AI text on the same subject matter. Pangram’s model trains on the differences to understand where humans and AI diverge. “Anything that the model says is uncertain, or says it’s AI, we know it’s wrong because it couldn’t have been written by AI if ChatGPT didn’t exist yet,” Spero says. “That’s the kind of text we can feed back to our model to correct it.” For any document that users scan using Pangram’s website or browser extension, Pangram provides a percentage rating, estimating how much was human- or AI-written with a segment-by-segment breakdown. It can also recognize AI-assisted text that a human might’ve written and subsequently revised with AI, and it’ll call out cases where it suspects that someone is paraphrasing the words that AI came up with. Installing the Pangram extension allows it to judge the posts in your social media feeds automatically. Pangram’s results have gotten more accurate over time. While last year’s Pangram 3 model falsely flagged roughly one out of 10,000 human-written documents as the work of AI, the company claims that its latest Pangram 4 model has a false positive rate of just one out of 24,000 and catches roughly 99.7% of AI-generated text. The company also points to https://www.pangram.com/blog/third-party-pangram-evals peer-reviewed research from University of Chicago https://bfi.uchicago.edu/wp-content/uploads/2025/09/BFI WP 2025-116.pdf , University of Maryland https://arxiv.org/pdf/2501.15654 , and Vrije Universiteit Brussel https://link.springer.com/article/10.1007/s40979-026-00226-w showing that it outperforms other detectors such as GPTZero and Originality.ai. In regards to false positives in particular, University of Chicago’s study concluded that Pangram “dominates the other detectors across all thresholds.” “We’re just working on continuing to execute, continuing to push the frontier,” Spero says. “Other people can catch up, but we’re going to be the ones forging this path forward and being state of the art.” One problem with Pangram’s success is that it’s become a bigger target. Earlier this year, The Atlantic ‘s Matteo Wong noted that he could trick Pangram https://www.theatlantic.com/technology/2026/05/pangram-ai-detection-accuracy/687381 with a tool called Walter Writes AI https://walterwrites.ai/ , which promises to rewrite AI text into something more human. And on X last month, an AI researcher by the handle Rosmine touted a new AI writing tool called Deft https://deftwriting.com/ by posting a screenshot of a clean Pangram score for an AI-generated essay. Pangram has responded to these “humanizer” tools by adding an extra layer of training to its models, designed to target how each tool operates. In my own testing, it seems to have neutralized Walter Writes AI already, and Spero recently implied on X that Deft would soon suffer the same fate https://x.com/max spero /status/2089905639644418127 . Still, he acknowledges that a cat-and-mouse game is underway. In cases where Pangram hasn’t trained on a specific humanizer tool, Spero says it’s likely to flag the text only 50% to 70% of the time. As the humanizers adapt with new methods, Pangram will have to come up with new responses. “That’s just life,” Spero says. “We’re settling into this dynamic where there is this back‑and‑forth.” It’s also possible that Pangram could unintentionally become a tool to evade AI detection. One of Pangram’s partners is GradPilot https://gradpilot.com/ , which coaches prospective university students on the quality of their admissions essays. Last fall, GradPilot tapped Pangram for an AI detection feature that discourages students from submitting AI-generated content. The feature is optional, but Nirmal Thacker, GradPilot’s founder, says that most students leave it on by default. “There is a new anxiety that they don’t want to come across as AI,” Thacker says. Even so, those students are typically using AI-generated content as the starting point for their essays, putting GradPilot in the position of making the prose seem more human again. Students still have revise the text on their own, as Gradpilot doesn’t generate any of its own content, but it does provide a rubric for building more authentic essays that also perform better on Pangram’s AI checks. Thacker acknowledges that there’s no way to tell whether students are using GradPilot to evade AI detection specifically. “I’m hoping that students align with their ethical views, and they’re using the tool in the right way,” he says. Traynor, the University of Florida engineering professor, is skeptical that tools like Pangram will ever be foolproof. Traynor cowrote a 2026 study https://news.ufl.edu/2026/05/traynor-ai-detector-study/ finding that commercial AI detectors were easily to trick, for instance by asking LLM tools to sound more human or use fancier words, though Pangram wasn’t among the detectors tested. “It’s important to remember that the space is adversarial,” Traynor says. “A new mousetrap will be invented, people will get around thatmouse trap, and it’s just not clear to me that a good enough mousetrap can be invented here.” Even if Pangram is correct in identifying human writing nearly all of the time, the cases where it’s wrong can get messy. In July, the author Freddie deBoer had to defend himself from an accusation that he’d used AI in a 2025 essay https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-is-broken-but . Although Pangram deemed the entire essay to be 100% human written, it marked a 12-sentence passage as 100% AI when the accuser scanned it in isolation. The finding sent deBoer down a rabbit hole. When he broke those 12 sentences into smaller chunks, Pangram deemed them human again. When he buried a block of human text inside reams of AI slop, Pangram said the whole thing was AI. The opposite experiment, with an AI passage tucked into a larger human-written document, earned a 100% human rating. All of this happened before Pangram released its latest model, whose analysis is more granular, but deBoer was able to coax a false positive from the new version by adopting as many AI habits as possible https://freddiedeboer.substack.com/p/how-to-produce-a-pangram-4-false . Speaking with Fast Company , deBoer says he’s concerned about fallout from false accusations, especially when someone can’t pay $20 per month for a Pangram subscription to show that their full body of work is human. While he credits Pangram for being transparent and forthright about its technology’s limitations, the people who use it aren’t always as level-headed. “If you put ‘Pangram’ into a search bar on Twitter, you’re going to find a ton of people who have kind of a religious zeal about the Pangram results,” he says. “The company is much more sober and responsible about the claims of its technology than some of its boosters are online.” Spero isn’t too concerned about the potential for witch hunts https://www.404media.co/substackers-say-new-ai-detection-tool-is-a-witch-hunt/?ref=daily-stories-newsletter , partly because the people complaining about such things are often the ones secretly using AI, and partly because he doesn’t believe the consequences are too severe. “People aren’t calling for someone to get burned at the stake,” Spero says. “What they’re doing is, they’re unsubscribing from a newsletter. They’re saying, ‘I don’t want to read this.'” Of course, the impact can be more significant than just some lost newsletter subscribers. When readers accused the author Mia Ballard of using AI to write large portions of the horror novel Shy Girl earlier this year, Hachette conducted its own review, canceled its plans to publish the book in the United States, and discontinued it in the United Kingdom. Ballard has denied using AI to write the book, blaming an acquaintance who used AI tools for editing. Ballard told The New York Times https://www.nytimes.com/2026/03/19/books/shy-girl-book-ai.html that “my name is ruined for something I didn’t even personally do.” Pangram’s latest technical paper also hints at a longer-term concern https://arxiv.org/pdf/2607.27183 with false positives: The more time people spend conversing with AI, the more likely they are to adopt chatbots’ writing styles as their own. Pangram calls this “data drift” and says it’s a become a major research focus, but acknowledges the current model doesn’t account for it. In the meantime, writers are learning to dodge some of the habits that AI has taken from them—the em dash is a prominent example—lest they land on the wrong side of an accusation themselves. deBoer worries about a future where writers will have to further “censor themselves and condition their writing to satisfy the meta-goal of not triggering Pangram,” while Traynor fears that task may eventually become impossible. “It’s not clear to me that text is sufficiently multidimensional to say that the machines can’t, or won’t, within a year or so, write in a completely indistinguishable fashion,” Traynor says. A few weeks ago, I was scrolling through LinkedIn when Pangram’s extension flagged a post from a fellow writer as “AI-Assisted.” My initial reaction was one of smug superiority— haha, imagine needing AI for this —but then the ruling struck me as odd given the post’s subject matter, a celebration of two decades in the field. If not for the sake of this story, I probably would have just moved on, carrying a seed of doubt about anything I saw from this writer in the future. But then I decided to reach out. In a private message, the writer assured me that no AI was involved, and that they’d in fact spent way too much time formulating the right words on their own. Then I noticed, a few clicks deep into Pangram’s report, a rating of “low confidence” about the passage it flagged. In this case the X-ray vision probably didn’t work quite right. Yet this is a taste of what a future Pangram-adjudicated web might look like. As Spero points out, AI content is already creating real-world harms though fake news websites, foreign influence campaigns, and all other manners of inauthentic behavior. An AI slop filter feels increasingly necessary, but it’ll also lead to more second-guessing of everything we see and read, even from the people we know and trust. Once you start applying something like Pangram at the individual level, you might not like what you see.