An AI detector scored a passage from Mary Shelley's Frankenstein as 99% AI generated. Six other detectors tested the same text and called it human, which tells you more about the state of AI detection than it does about Shelley.
Mary Shelley wrote Frankenstein in 1818. Two hundred and eight years later, a piece of software decided she never wrote a word of it.
In March, a screenshot started circulating on X and Substack Notes. It showed an AI detector scoring a passage from Frankenstein's fifth chapter as 99% AI generated. The excerpt was Victor Frankenstein waking to the sight of his creature by candlelight, pulled straight from Project Gutenberg's public-domain text. No paraphrasing. No tricks. Just Shelley's own prose. The post drew more than 6,500 likes before anyone thought to check the detector's homework. You don't need a literature degree to find the joke here. A novel about reanimating dead flesh from spare parts got mistaken for a machine.
But the funny part is also the part that should worry you, because this wasn't just a toy sitting harmlessly on the internet. It's one of dozens of detection tools now hovering around university plagiarism checks, publishing workflows, and hiring decisions, and this wasn't the first time one of them flagged something absurd. When a Substack writer ran the same 203-word Frankenstein passage through six other detectors, the results didn't just disagree with the first tool, they demolished it. One called it 100% human. Another found no AI content at all. A third was fully confident it was original. A fourth scored it entirely human-written. A fifth put the odds of AI authorship at zero. A sixth landed at 98% human. One tool, out of seven, thought a Romantic-era novel was a chatbot's output.
The writer noticed something else worth keeping in the frame: the flagging tool's results page included a humanize button that sent flagged users toward a paid rewriting service. Draw your own conclusions about the incentives at play when a free detector's business depends on finding something to flag.
AI detectors are turning ordinary student writing into evidence. AI detectors are increasingly flagging careful academic writing as suspicious, especially in structured assignments like literature reviews. The real issue is not whether students use AI, but whether universities are treating uncertain software scores as evidence. - how AI detectors flag legitimate student writing - why careful essay writing triggers academic AI detection
How the machine actually decides #
Here's the mechanism, as best anyone outside these companies can see it. AI detectors don't read for meaning. They measure statistical patterns such as perplexity, roughly how predictable a piece of text is to a language model, and burstiness, how much sentence length and structure vary from line to line. Large language models tend to produce smooth, evenly paced text. So, unfortunately, does a lot of older literary prose written in long, formal, grammatically regular sentences. Shelley's writing can look steady to a detector for the same reason a student essay can look steady: the tool is measuring pattern, not authorship.
That distinction matters. A detector isn't catching a person in the act. It is assigning a probability score based on patterns it has learned from training data, and the weak point is obvious once you put Frankenstein beside a real classroom complaint. The number looks exact. The judgment underneath it isn't.
The students paying for it #
This isn't a hypothetical problem confined to novelists who are safely dead and can't be expelled. In a 2023 Patterns paper, researchers including Stanford's James Zou reported that seven AI detectors misclassified 61.3% of TOEFL essays by non-native English writers as AI-generated, compared with 5.1% of essays by U.S. eighth-grade students. That isn't a rounding error. It's a warning label.
Students have already taken this fight to court. Ars Technica reported in July 2026 that Thierry Rignol, a Yale School of Management executive MBA student, sued Yale after a cheating dispute involving an AI detector and a final exam. In Michigan, a University of Michigan student sued in February 2026 after being accused of using AI in coursework, with the complaint saying her formal writing style and AI comparison outputs were treated as evidence against her. Those cases are not the same as the Shelley screenshot, but they live in the same world: institutions are giving software-shaped suspicion too much weight.
Some schools have worked that out. A major Australian university said it would disable its plagiarism platform's AI writing detection feature starting January 1, 2026, across its campuses and study periods. A leading U.S. university turned off the same kind of AI detector back in 2023. The plagiarism platform's own 2026 guide tells instructors its model may misidentify human and AI-written text and should not be used as the sole basis for adverse action against a student.
Vendors keep insisting the technology has improved. One detector maker pointed to a 2026 Chicago Booth benchmark where it reported 99.3% recall and a 0.05% false positive rate on that dataset. Another vendor has said its document-level false positive rate is under 1% for documents with 20% or more AI writing, while its sentence-level false positive rate is around 4%. Fine. Controlled benchmarks have their place.
They didn't save Frankenstein.
Superhuman acquired GPTZero on June 23, the AI detection platform built by Princeton grad Edward Tian that grew to 19 million users and $30M ARR. The deal folds GPTZero's hallucination detection, plagiarism analysis, and authorship verification into Superhuman Go, a cross-app AI assistant, creating what the company calls an authenticity layer for... - Superhuman acquires AI detection startup - AI detection features for productivity tools
Here's the thing nobody wants to say out loud in a faculty meeting or an HR review: a tool that can't tell the difference between a large language model and Mary Shelley has no business making accusations about anyone's authorship on its own. Universities and employers keep treating a percentage score as evidence. It's really just a guess dressed up in a number. If a detector can't clear the lowest possible bar, canonical English literature that predates computers by more than a century, it shouldn't be clearing a much higher one. Not someone's grade. Not someone's job. Not someone's reputation.
The most honest response to the Frankenstein incident isn't laughing at one detector and moving on. It's asking why so many institutions are still outsourcing judgment calls about honesty to software that just failed a test written by a 20-year-old in 1816.
Also read: Stanford Payroll Data Shows AI Has Cut Young Workers' Job Gap to 19% • Sam Altman Fears a Handful of Companies Will Control AI • SEC Subpoenas Wall Street Banks Over AI Hedge Fund Situational Awareness