A $2 million book deal collapsed on an AI accusation

The publisher Minotaur dropped a $2 million book deal last month over concerns that its author, Jerry Falade, used AI to write the manuscript. Falade says he didn't. The deal is gone either way. That sequence is the cleanest summary of where AI suspicion now sits: an accusation gets made, there is no reliable way to verify or refute it, and the accused carries the cost.

Detection tools have become a fixture in classrooms and newsrooms. A survey from the Center for Democracy and Technology, reported by The Verge, found that 43 percent of sixth to 12th grade teachers in the US regularly used them between 2024 and 2025. When Turnitin introduced AI detection in 2023, some universities found the feature switched on for their accounts automatically. And the companies behind these tools keep publishing disclaimers that undermine their own marketing.

How a probability became a verdict

Plagiarism checkers earned their keep by matching text against a database of published work. AI detectors do something murkier. GPTZero, Pangram, and the detector created by Turnitin all rely on their own AI models to estimate whether text might be machine-written. GPTZero describes its method as an algorithm that analyzes wording, rhythm, and structure, hunting for patterns of length and tone more common in AI prose. That is a judgment call, not a match against anything. Researchers are only beginning to open that kind of black box, tracing model decisions to specific training cases, and none of that transparency has reached the detectors flagging student essays. It stumbles over writers whose first language is not English and flags people who never vary their sentence structure.

The numbers the companies publish sound reassuring. Turnitin says its detector falsely flags less than 1 percent of human-written content. Pangram claims one false positive in 10,000. GPTZero says its rate is similarly low. Then the disclaimers follow. Turnitin warns the tool "may not always be accurate" and should not be used to take action against a student. Grammarly says users "should never rely on the results of an AI detector alone." GPTZero concedes "no AI detector can ever truly be 100% perfect." OpenAI shut down its own detector in 2023 over low accuracy. The pattern extends beyond detection: vendors keep asking the public to trust their word about their own models. That gap is why Microsoft funds 18 university labs across six continents to independently red team its AI systems, opening AI safety testing to outside researchers.

The claims, next to the caveats, read like this:

The false positives land on people

The public cases follow the same shape. Thierry Rignol, a French national, sued Yale after a professor ran his final exam through GPTZero and accused him of AI use. The failed grade and one-year suspension followed. The lawsuit argues that AI detection tools are known to unfairly target non-native English speakers. In February, a student won his lawsuit against Adelphi University after a professor made a similar accusation. Adelphi has a licensing agreement with Turnitin.

The bias is documented. A 2023 Stanford study found detectors flagged essays by non-native English speakers as AI more often than those by native speakers. The English bias is not unique to detectors: half of Asia's enterprises don't use English for AI, and Alibaba Cloud's NielsenIQ survey of 1,000 Asian IT decision makers calls that the real bottleneck, the language gap in AI adoption. UCLA explains that the tools are trained on signals like repetitive terms, text that sounds too formal or too informal, and nonsensical phrasing. QuillBot measures the "unpredictability" of text on the theory that AI reaches for the most obvious word. None of these signals proves anything on its own. Some people just have that style. The same concern now extends to neurodivergent writers.

Institutions are backing away

The uncertainty has pushed some schools to the exits. Yale, Johns Hopkins, Vanderbilt, and Georgetown have disabled or restricted AI detection tools. MIT states its position plainly: "AI detectors don't work." The University of Chicago instead recommends slowing down reading, breaking up long assignments, and asking students to reflect on their work. Stanford suggests professors consider in-class assessment. MIT wants room for students to disclose AI use without penalty. The pattern across all of them: redesign the assignment instead of policing the output.

The accusations are outrunning the tools

Online, the volume is higher and the stakes are still real. Substack has built Pangram into its app so readers can scan newsletters for suspected AI writing. LinkedIn added a "seems like AI slop" button to posts. In a video broadcast to more than 3.5 million followers across his social channels, Ozzy Osbourne's son Jack accused Verge contributor Kat Tenbarge of using AI on a Rolling Stone article, waving an AI detector called Getsolved around as proof. Tenbarge has refuted the claim in a video and a post on her own site. Osbourne has not retracted the accusation, leaving Tenbarge to deal with the trolls.

Real writers now edit themselves to sound less like machines, the strangest tax this era has invented, one that fits a wider pattern: a majority of workers report AI tools add to their workload rather than cut it, the productivity paradox. Wikipedia has banned AI-generated articles and published a guide to spotting AI writing, including the habit of "puffing up" a topic's importance. The Authors Guild is handing out "Human Authored" certifications. In visual art, the same fight has a price tag: Pippa pays artists $0.005 per image while its models still train on scraped art, the economics of that payout. Badges from Not by AI and Written by Human let people label their work as made by a person.

The asymmetry is the story. An accusation costs a screenshot. Clearing your name costs a lawsuit or a public reputation. Until the detectors stop overstating what they can do, the era of distrust runs one way, and the people paying for it are the ones who never used AI at all.