The Problem with AI Detector Companies

Be wary when the same companies co-author the research, rile up the mob, and sell the solution.

Third Hemisphere art by noted human Matteo Farinella

Welcome to The Third Hemisphere, where I try to make sense of how AI is reshaping work, thinking, and creativity, often by watching my own assumptions get upended.

I’m publishing the rare back-to-back post simply because I’m watching an AI-detection-fueled train wreck in media and publishing, and I fear people aren’t taking into account the limitations of AI detection software (or more precisely, AI “inference” software), nor the conflict of interest of the companies riling up the mob. If you want to catch up, I wrote about the self-defeating dynamics of detecting and punishing individual authors for AI use; this article is about the self-serving and morally questionably behaviors of people associated with the AI detection company Pangram.

If you were forwarded this and want to subscribe, click below. If you want to support a real human writing about AI, upgrade to paid.

Example Pangram result

Update, April 17, 2026: I got this partly wrong and want to say so at the top rather than bury it in a footnote. I also updated the title of this article to reflect. The backstory: After the piece ran, Pangram’s CEO and sparred a few rounds on social media, and rather than dig in I went back to the technical literature and called independent computer scientists who work on AI detection. This is what I got wrong: The state of AI detection has changed, and Pangram turns out to be quite accurate on fully AI-generated text from major models, even on short bits of text. I stand corrected that detection is only good for population-level work. The concerns I still hold are thus narrower. I think AI detection still suffers from real-world failure modes that the companies discount, so company-reported false positive and false negative rates should be taken with a grain of salt when applied in the real world. But there are many nuances to this. Technical specs aside, I still take issue with the CEO leaning into call-out culture to the advantage of his product, as I argued in another Substack post, Detect and Punish. I laid out my full reporting and updated thinking in a piece for Slate.

Over the past month, a company called Pangram has been at the center of several high-profile AI-authorship controversies. Pangram’s detector flagged the Shy Girl novel that Hachette canceled. Pangram’s detector flagged the Modern Love column that went viral on X. And a University of Maryland preprint, co-authored by four Pangram employees, used Pangram’s detector to scan 250,000 newspaper articles and claim that 9 percent of U.S. news contains AI-generated text, including op-eds by very fancy people in the New York Times, Wall Street Journal, and Washington Post. That study powered a credulous piece in the Atlantic and generated national coverage. On LinkedIn, Pangram’s CEO, Max Spero, self-servingly urged publishers to “strictly moderate AI-generated content” and “draft and enforce robust AI-use policy.” On Twitter/X, where he calls himself a “slop janitor,” he has been using his own product to publicly accuse individual journalists by name:

This strikes me as a morally questionable way to police AI slop.

There's a basic problem with this: Pangram is probably a decent tool for studying population trends, but Pangram's CEO is wielding it as a weapon against individuals. These are fundamentally different applications. The term AI “detector” is a bit of a misnomer, as tools like Pangram don't scan a database of known AI text for a match; they make a guess about likely authorship based on differences in patterns between AI-generated and human-generated text. Rather than a detector, thinks of it as an “AI inferencer” or more colloquially, an “AI best guesser.”

As an inference tool, Pangram is vulnerable to everything that makes inference so difficult: small sample sizes (a single article rather than a corpus), small effect sizes (as the statistical distance between human and AI writing shrinks), and confounds (a human who reads and mimicks AI-generated text before writing her own). Once you understand that these tools are inferencers, not detectors, you start to see why universities, researchers, and even Pangram's competitors all warn against doing exactly what Pangram's CEO is doing on social media.

This behavior is why I was glad to see the Wall Street Journal‘s James Taranto take Pangram to task last Friday in a wonderful dispatch of boots-to-the-ground skeptical journalism. Taranto targeted Pangram’s scientific paper, which outed three WSJ op-eds as AI-generated:

“My first step was to run the articles through Pangram, whose website allows four free scans a day. To my surprise, the same tool applied to the same articles produced different results. Mr. Blum’s piece came up 100% human; Mr. Spencer’s, 44% AI and 56% human, which would make it “mixed,” not “AI-generated.” Only Dr. Saphier’s article was still labeled 100% AI-generated. Complicating matters further, when I checked the researchers’ database, it labeled the Blum and Saphier articles “mixed” and only the Spencer one “AI.” To my mind, these wildly inconsistent results are enough to discredit any accusation based on a Pangram analysis.”

Taranto contacted the accused writers. Spencer said he doesn’t use AI at all, though he circulates drafts among friends for feedback. Blum said he uses ChatGPT to check grammar but doesn’t paste output back into his drafts. Saphier said she sometimes pastes AI-edited text back in but has since stopped. I’m not here to judge the practices of individual writers but the practices of Pangram. As the paper’s lead author, a University of Maryland doctoral student named Jenna Russell, told Taranto: “We do not accuse anyone of using AI, rather we report trends at an aggregate level,” she wrote. “We do not in any way say that using AI is inherently good or bad!”

But as Taranto points out, actions speak louder than words:

A look at Ms. Russell’s and Mr. Spero’s Twitter feeds belies that nonjudgmental pose even more glaringly than their study does. On Feb. 16, Mr. Spero taunted a staffer for a British newspaper, whose name I will withhold: “We fetched 871 articles published in the Guardian by [the journalist] over the last six years. It’s clear that he is increasingly relying on AI. In two weeks in February he churned out nine articles classified by Pangram as fully AI-generated. Receipts below.”

On Feb. 18, Ms. Russell replied: “Reminder that you can search over 250k news articles for AI slop at . . .” followed by the URL of the Maryland study’s database.

I think it’s a distraction to litigate whether these particular accusations are true or not. The important thing is that the intellectual sleight-of-hand by Pangram’s CEO and the lead author is laughable. You can say you’re only interested in reporting aggregate trends. But it’s hard to take that at face value if you also publish a table of named individuals with “AI-generated” next to their names, plaster screenshots of bright red “AI-generated” graphics next to real authors’ names, if you call yourself a “slop janitor” on social media and respond to approvals of your public shaming posts with a salute emoji. Come on.

Taranto seems to dismiss Pangram completely, but AI detection tools aren’t snake oil, and Pangram specifically has done genuinely useful work at the population level. Its newspaper and peer review studies are credible. When the American Association for Cancer Research used Pangram to screen tens of thousands of abstracts, methods sections, and peer reviews submitted to its journals, what they found was genuinely useful: a sharp rise in suspected AI-generated text after ChatGPT’s release, higher rates in peer reviews than in manuscripts, and the finding that authors at non-English-speaking institutions were more than twice as likely to use LLMs. That’s fantastic information for setting policies and thinking about evolving technology use. But even Mohammad Hosseini, a research ethicist at Northwestern who reviewed the findings for Nature, cautioned that “there is no way to verify that flagged manuscripts actually used AI.” When you point the same tool at a single writer and stake their reputation on the output, you are asking it to perform a task that may be beyond its abilities.

Example of an excellent use of the Pangram tool, which estimates that 21% of peer reviews submitted to the academic conference ICLR are fully AI-generated (source).

To understand why the individual accusations are so reckless, consider a calculation by Arvind Narayanan, the Princeton computer scientist who wrote AI Snake Oil. P Pangram claims it has a false positive rate of 1 in 10,000. Narayanan took that number at face value and applied it to a concrete scenario: If every instructor used an AI detector on all student submissions, and students submit roughly 500 to 1,000 written works over four years of college, 5 to 10 percent of the student body would be falsely accused of cheating at some point during their undergraduate career. The statistical lesson is if you run any test with even an impressively small error rate enough times, you'll still generate a lot of false positives. This is why Pangram's aggregate findings are credible, but any individual accusation built on the same tool requires far more caution than a screenshot and a salute emoji.

I understand this trade-off firsthand because, as a professor, we have to make ongoing decisions about using AI detectors to adjudicate cases of academic misconduct. So far, no technology has proven reliable enough for universities to adopt for this purpose. NYU, where I teach, has disabled Turnitin’s AI detection. UT Austin prohibited all third-party AI detection software for evaluating student work and warned that faculty who purchase such tools on a personal credit card may be “personally liable for paying any damages or legal costs.” Berkeley, Yale, Georgetown, Vanderbilt, Northwestern, MIT, and many more reached the same conclusion.

The reason is simple: No reputable study I’m aware of published between 2024 and 2026 supports using AI text detectors as the sole basis for accusing an individual of AI authorship. Findings are particularly damning when it comes to individual accusations. Detectors are systematically biased against non-native English speakers, and the same text submitted to the same detector on different days sometimes gets different classifications—something Taranto very publicly demonstrated in his WSJ story. Turnitin’s own documentation states that its tool “should not be used as the sole basis for adverse actions against a student.” Originality.ai’s terms of service go further: “You must not use any Output relating to a person for any purpose that could have a legal or material impact on that person,” including educational and employment decisions, or “rely on Output from our Services as a sole source of truth.”

Pangram’s CEO argues that his tool detects “signals much deeper than surface-level style” and that “only an LLM can effortlessly make choices consistent with an LLM.” To be clear, Pangram is the best commercially available detector available, and has overcome many of the problems that plagues the first generation of AI detection software.¹ No serious person can argue that Pangram is snake oil; its performance been demonstrated in multiple independent studies at this point. However, soon you run into problems that no AI “detection” software can overcome because they are problems of inference.²

One is the theoretical ceiling. An interesting study showed that the accuracy of even the best possible detector is mathematically bounded by the statistical distance between human and AI text distributions—and as that distance shrinks, the ceiling on detection approaches random chance. This theoretical finding is important because of several forces. The first is that as models increase in capability and personalization, their outputs will be less distinguishable form human outputs. The second—and Marshall McLuhan would love this one—is that humans are generally absorbing the linguistic patterns of LLMs, so human writing is becoming less distinguishable from AI output. A tool trained to guess whether it’s human or machine text will probably find that guessing game harder to play each year

Then there’s the arms race. AI use is rampant in education, so are “humanizer” tools designed to evade detectors. A writer who deliberately wants to evade detection can rewrite AI-generated text until it passes, which means the people most likely to get flagged are the ones using AI casually—like some of Taranto’s writers—while more savvy users avoid the public stoning. Pangram says it checks against humanizer tools, but as Pangram gets more popular, arms race dynamics suggest evasion will get more sophisticated.

Take into account the convergence of human and AI writing, along with active attempts to evade detection, not to mention the whole spectrum of evolving hybrid uses and proliferating tools, and Pangram’s “deep signals” will get shallower and shallower.

I sometimes turn to Claude for stages of research, for editorial feedback, for copy-editing, factchecking, and occasionally for a sentence. I don’t think these decisions are without moral, cognitive, or literary costs, and I’ve chronicled my experimentation with AI extensively because I’m trying to understand the ways this technology helps and harms the scientific community. As I argued in my last Substack post, AI use deserves realistic editorial policies and empirical study. Pangram’s population-level work is genuinely valuable in that it contains informative aggregate data about AI adoption across the American press and scientific community. But the individual call-outs are more suspect.

The charitable read is that the CEO is on a genuine moral crusade against AI slop, with unwarranted confidence in his product. The other read is Pangram is running a savvy PR campaign: co-author an academic paper, get it covered in major newspapers and magazines, toss red meat to the AI-anxiety crowd on social media, then charge worried publishers a subscription to clean up the mess, all while authors sign up for Pangram to screen their writing for false positives, lest they become the next Mia Ballard. I can’t say which best explains the situation, but either way, the effect is the same: individual authors get sacrificed for a tech company’s bottom line.

In February 2024, Spero co-authored Pangram’s technical report, which I read. It includes an ethics section. It reads: “We strongly discourage the use of our classifier as a sole arbiter of academic integrity and plagiarism checking. All AI detection tools have a nonzero false positive rate, and should be used in conjunction with other evidence to prove or disprove plagiarism.” Two years later, the same CEO was posting Pangram analyses of named journalists on X and calling it “receipts.”

Not good enough to give a student a slap on the wrist, but apparently good enough for ruining careers on social media.

1. Here is the state of the field of first-generation AI detectors: The best detectors achieve high accuracy on unmodified AI text from known models, but accuracy collapses by 25 to 50 percentage points on paraphrased, hybrid, or lightly edited text, which is the kind of text many people actually produce when they use AI as part of their writing process. What’s worse, detectors are systematically biased against non-native English speakers, a finding confirmed in multiple 2025 studies. Research also backs up the unreliability Taranto demonstrated in his WSJ story. One study tested 14 detection tools and found that the same text, submitted to the same detector on different days, sometimes got different classifications. Their conclusion: “The available detection tools are neither accurate nor reliable.” The RAID benchmark, the most rigorous independent evaluation to date, tested detectors across six million text generations and found that when you require a low false positive rate—which you’d need if you’re going to accuse a specific person—most detectors fall apart. Frustratingly, if AI companies would just start watermarking their text, which they are not incentivized to do, then the most egregious AI use would be easily flaggable.

2. Update, April 6, 2026: This paragraph was revised to better reflect the literature on Pangram’s performance.