An editor acting on one of these scores has a two-digit number in front of them and an author on the phone denying it [3]. So what the number is made of matters. Wired reports that Pangram trains by synthetic mirroring: take human writing, have LLMs produce a close match, and teach the model what machine prose looks like next to the human original [14]. It also runs hard negative mining, combing datasets for false positives it can mirror and fold back into training [15]. That second technique is the most useful disclosure in the pitch, because nobody spends compute hunting their own false positives unless there are enough of them to be worth hunting.
A score of 78 is therefore a classifier's confidence, printed on a scale a reader will take as the proportion of words a machine wrote. In Wired's account of the Shy Girl cancellation there is a post on X and a denial, and no audit of drafts or version history in between [4][5].
The funding tells you how young this judgment is. Pangram has raised $13 million in total, of which $9 million arrived in July [2][12], leaving about $4 million to fund everything before that [1], including the run of scores that ended a Hachette release [5] and put a Commonwealth Short Story Prize winner's authorship in doubt [7]. At least a dozen firms compete in detection, among them Originality.ai, GPTZero and Turnitin, and Pangram performed well in some early independent testing and emerged as the front-runner [16]. Front-runner in early independent testing is a real claim, and it is a considerably smaller one than fit to void a contract on.
Jane Friedman, an author and publishing expert, told Wired that writers feel the detectors are "just as evil, if not more evil, than the AI companies themselves" [11]. Read that as a fight about standing rather than accuracy. The score acquired the power to start a process, and the people it lands on had no part in granting it.
Two axes are enough to sort this for your own desk. First, how reversible is the action the score triggers, running from a private query to the author at one end to cancelling a release at the other. Second, do you hold anything besides the score, such as a revision history or an admission. Score-only plus reversible is what a flag is for, and it is the only cell where a percentage is doing work it can support. Score-only plus irreversible is where every one of these stories has happened, and the publisher's name, not the vendor's, goes on the correction. Substack's reader-facing version leaves the ordinary reader at score-only, with no drafts to consult and nothing available but belief or disbelief [10].
The forcing question: what you do with a 60 percent, the figure Pangram gave the novel Daggermouth [8], against a writer who says no. If the answer is cancel, you have promoted a classifier to acquisitions.
