A student who wants nothing to do with a chatbot now opens Pangram on her own time, because the TA might, and a highlight is cheaper than a conversation. An agent asks for a “clean” scan with the manuscript. A midlist novelist pastes the third chapter into a $20-a-month box the way her mother once ran spellcheck. If the box is yellow she pays again: an editor, a humanizer, a weekend she does not have, to make the sentences look less like 2026.
That is a tax on people who type. It is collected by the same capital stack that funded the models that made the fear, then funded the referee that sells the alibi. Call it a money waste if you want to be polite. It is a toll both ways: the institution pays for suspicion, the writer pays for innocence, and neither purchase produces a better paragraph.
The product is not “is this AI.” The product is “can I discard this.”
Pangram’s own 2024 paper said the quiet part: We strongly discourage the use of our classifier as a sole arbiter of academic integrity. All detectors have a nonzero false-positive rate. Use other evidence. Two years later The Atlantic could write that basically every high-profile accusation starts in Pangram: a horror novel pulled days before publication, prize stories, a newspaper, a papal encyclical. WIRED asked if you should trust the gold standard. The CEO had been posting named-journalist “receipts.” The ethics paragraph did not travel with the screenshot.
Spero told TechCrunch the API customers include schools, universities, publishers, agents, recruiters. Substack wired it in so readers can see which newsletter is a model. The Chrome extension grades the feed. The consumer plan is twenty dollars a month. Dashboard credits are one per hundred words. A novel is a bill. A dissertation is a bill. A week of student essays, at institutional seat licenses, is a line item next to the LMS.
Turnitin, which already sat on every paper, added its own AI number and still tells schools not to treat it as the sole basis for discipline. Some universities disabled it. UT Austin warned faculty who buy a detector on a personal card they may be personally liable. The market did not shrink. It moved to the tool that claims one false positive in 10,000, or 24,000 on Pangram 4’s own holdout blog post. One in 24,000 is a rounding error until it is your kid, or the only draft an editor will open this month.
Why living writers lose even when they are telling the truth
The referee was not trained on “did a human sit here.” It was trained on pre-2021 book / student / review prose versus 75 models rewriting that prose. Clean, even, professional English after ChatGPT lives next to the rewrite. A 1925 sentence lives in the human bin. We scored leftover-free Gutenberg fills at a thousand words: 14 types, entire-text Human. We scored Llama 3.3 70B, Qwen 2.5 72B, and Grok 4.5 on the same original story: 100% AI, celebrity voice or not. The box is not a lie detector. It is a costume contest for dead prose.
Students are already in the published evals — PERSUADE, ELLIPSE, PELIC, ICNALE — 23,000 student documents in the 2024 pool against 7 million book examples. The people with the least power are both the training data and the customer. ESL writers have been the canary since the first detector papers. Pangram boasts it beat that bias on TOEFL and ELLIPSE. The boast assumes the job is to sort students. The job, if you are a person, is to write without a second job as your own defendant.
Then you pay the other window
If the scan is dirty you do not get a conversation. You get a market. Humanizers. Editors who now advertise “Pangram-safe.” A weekend of wrecking your own voice so the windows look like 1925. We sell one of those tools. I will not pretend otherwise. The existence of the second window is the indictment of the first. A healthy literary economy does not require a teenager to buy an alibi and then buy a disguise.
Add the model subscription she was told would “just help with outline.” Add Claude’s new invisible watermark if she did, which she cannot inspect, which travels with copy-paste, which a school or a publisher may one day buy access to. Three payments. One paragraph. No one involved is more sure the paragraph is good.
That is the waste. Not “AI is bad.” The waste is a stack that first socialized the cost of the libraries onto writers, then privatized the panic, then sold both the panic and the cure. Menlo can be right that Anthropic is a historic company and still be the firm on both sides of a shakedown. The writer is not a customer in that story. The writer is the surface the tax is collected from.
Part 4 is the watermark they added to the sentences without putting a refusal in the contract. If you only send one URL, send part 1. If you are a writer, this one is the one that is about you.
Sources: arXiv 2402.14873 ethics note and Table 5; Pangram 4, 1-in-24,000 holdout; TechCrunch, 29 Jul 2026; The Atlantic; WIRED; The Verge on detector distrust; Requarth on sole-arbiter use.