The Verge column argues AI detectors are driving a new era of writing distrust

The column and its core argument

The Verge's weekly "The Stepback" newsletter on Aug 9, 2026, 12:00 PM UTC, carries Emma Roth's column "AI detectors are creating a new era of distrust," which states outright that "Tools that promise to detect AI writing can be dubious, but educators and publishers are using them anyway." The framing piece recasts the long-standing anti-plagiarism workflow, which historically compared student or freelance text against a web database, as something that has now morphed into "a war on AI-generated work." Roth names GPTZero, Pangram, and Turnitin's detector as the principal commercial AI-text classifiers and stresses that, unlike plagiarism matching, these tools "rely on their own AI models to guess whether something might not be human-written — a process that's arguably even murkier than matching text on the web."

That structural move from matching to guessing is the column's analytical pivot, and it sets up every downstream case study as a stress test of that guessing. Readers who draft, edit, or process text should treat the column as a status report on how institutions are weaponizing that guesswork against writers today, rather than as a vendor announcement or a confirmed benchmark. RISK PROFILE: the column itself frames the tools as dubious, and the high-risk detection accuracy claims that follow should be read with that posture in mind throughout the rest of the briefing.

Adoption pace and vendor accuracy claims

Adoption metrics are at the heart of the column's urgency. Roth reports a Center for Democracy and Technology survey finding that 43 percent of sixth to 12th grade teachers in the US regularly used AI detectors between 2024 and 2025, a reported adoption figure that converts what was an experimental policy in 2023 into a mainstream classroom practice inside the one-year window the column describes. She also notes that some universities already using Turnitin in their learning management systems "found that the service automatically enabled AI detection when the tool launched in 2023," meaning adoption was partially unilateral and unannounced.

Vendor accuracy statements receive separate treatment as reported, high-risk detection accuracy claims rather than audited benchmarks: Turnitin says its detector "falsely flags less than 1 percent of human-written content as AI," Pangram "claims its false positive rate is just 1 in 10,000," and GPTZero claims a similarly low rate. Roth then immediately lists caveats from the same vendors — Turnitin conceding its tool "may not always be accurate," Grammarly warning users "should never rely on the results of an AI detector alone," and GPTZero saying "no AI detector can ever truly be 100% perfect."

Readers weighing drafting tools against editorial risk should keep both halves of each vendor's statement in view, and can sanity-check extracted passages with Column to Comma Separated List before sharing flagged snippets.

Documented harms to specific writers

The column's body-of-evidence section moves from aggregate statistics to named individuals, which is where the stakes become concrete. Roth reports that "Last month, the publisher Minotaur dropped a $2 million book deal over concerns that its author, Jerry Falade, used AI — something he vehemently denies," casting a major trade imprint's first refusal-to-publish in the AI era as a publisher-supplied report rather than as a confirmed act of fraud. She then profiles Thierry Rignol, a French national who sued Yale after a professor used GPTZero to scan his final exam, gave him a failing grade, and imposed a one-year suspension; the lawsuit argues that "AI surveillance and detection tools are known to unfairly target non-native English speakers."

A third case closes the run: a student at Adelphi University "won a lawsuit against the school" in February after his professor "similarly claimed he used AI to write an essay," with the column flagging that Adelphi holds a licensing agreement with Turnitin. Each case lands differently — a contract collapse, an ongoing suit, a student-side legal win — and together they form a small, diverse record of detector-driven damage that writers should be able to cite when policies threaten their deliverables.

Bias findings and what the science shows

Roth lifts the underlying research that makes those harms legible at population scale, devoting a full segment to a 2023 Stanford study that found "AI detectors falsely flagged essays written by non-native English speakers as AI more often than native speakers," and notes that "many services still argue that their tools are accurate when dealing with text written by non-native speakers." A separate university source, the University of California, Los Angeles, is cited for the mechanism: detectors are trained to flag patterns such as repetitive terms, formality or informality mismatches, nonsensical phrasing, and unpredictability scores like QuillBot's measure of common language choices, which can be generated just as easily by neurodivergent or atypical human writers.

The OpenAI caveat sits in this section too: "OpenAI even shut down its own AI writing detector in 2023 due to low accuracy." These are the column's high-risk detection accuracy findings, presented as a report from the article rather than as confirmed contract terms, and the trust profile should stay elevated as readers build policy positions on them. Writers and editors can flag suspect passages for review with Compare Two Lists to surface vocabulary overlap before letting any detector output drive a deliverable.

What remains uncertain and what to watch

Several facts are explicitly unsettled inside the column itself and should be flagged before any reader builds a workflow on top of them. The age or current shelf state of the named detection products is not pinned down; the column reports vendor self-descriptions as a report, not as audited figures. Minotaur's $2 million decision is described as a publisher's report of a deal dropped, with Falade's denial noted, and it has not been litigated on the record in the column.

The Yale and Adelphi matters are presented as lawsuits at different stages, with the Adelphi outcome recorded specifically as a student-side legal win, not as a final judgment on detector reliability across courts. Even the social-accusation hook the column teases — Jack Osbourne accusing a Verge contributor "in a video broadcast to the more than 3.5 million followers across his social channels" — is cut off in the recovered body text, so its outcome, retraction, or settlement sits squarely in what-to-watch territory.

Watch next: any publisher policy memos on AI authorship clauses, any court rulings on detector admissibility, and any revised Stanford-style bias replications. If drafting teams need a quick counterweight to detector alarms, Bold Text Generator and Cursive Text Generator only reformat copy, so they cannot certify authorship either — that responsibility stays human.