Pangram's output is an estimate about the work
Pangram is a company with 24 employees that has raised 13 million dollars to date, and what it sells is a single percentage saying how much of a text was generated by AI. Substack connected the detector to its platform in late July. The company's percentages have followed novels pulled from publication, prize-winning stories and newspaper columns. The best known is Mia Ballard's novel Shy Girl: Pangram's chief executive, Max Spero, posted on X that the book was 78 per cent AI-generated, and Hachette later cancelled the release. Spero says the score is only a small part of the bigger picture in the cancelled deals. Tuhin Chakrabarty of Stony Brook ran the same tool over 14,419 self-published novels and found high scores in nearly 20 per cent of them; Chakrabarty receives usage credits from the company.[1]
Pangram says false positives occur 0.0041 per cent of the time with its newest model and that borderline text is deliberately declared human-written; on the same measurement, human essays heavily rewritten by AI still read as human-written 41.37 per cent of the time. A Notre Dame working paper on why AI detection falls short for academic integrity reports that the 3.2 model flagged lightly edited academic abstracts as AI writing 64–80 per cent of the time, and caught AI text passed through a humanizer tool less than 4 per cent of the time. The two rates describe two different populations. The limits the company itself discloses sit closer to the second: the tool performs worse on text under 100 words, the median input is 350 words, and an identical passage can get a different verdict scanned alone or inside a longer text.[1]
Meta's penalty rests on what it detects on the device
Meta confirmed on Tuesday that a software update aimed at people who drill out or cover the recording indicator light had shut down the cameras on thousands of its AI glasses, and said the action affected less than one tenth of a per cent of all Meta AI glasses sold. An earlier update stopped recording when the light was blocked, but it checked only at the start of filming, so users learned to begin recording and cover the light a few seconds later. The new update closes that gap. On scale, the only figure anyone outside can test sits on the sales side: Meta has not published a total since the 2021 launch, and 7 million pairs sold in 2025 alone. How many pairs were tampered with rests on the company's own detection.[2]
The outcome comes in two grades. On a pair whose LED has been drilled out the camera never works again; where the light was merely covered with a sticker, removing the sticker restores it. The line follows whether the hardware change can be undone, and in both cases the penalty is aimed at the one indicator a bystander can read. The company's own statements sit around that line: Alex Himel, vice-president for wearables, wrote on Threads on Thursday that only a tiny minority of people were tampering with the hardware and said Meta would keep rolling out updates of this kind so that the capture LED can reliably alert bystanders; the chief technology officer, Andrew Bosworth, called the glasses the most privacy-forward camera for consumers in an Instagram post on Monday. Both are interested-party statements rather than an independent measurement.[2]
Meta grades the penalty, Pangram offers no appeal
Meta's penalty rests on a physical event it detects on its own device, and its weight varies with whether that event can be undone. Pangram's percentage is an estimate drawn from the work itself, and once a publisher, an agency or a platform acts on it there is no grading left. Nor is there a published path for an author who says the flag is wrong: Spero says he has offered a cash reward in some cases to anyone who can prove they wrote the text themselves, and that so far nobody has taken him up on it. A Penguin Random House spokesperson said its editors may use approved detection tools, but that such tools are not determinative and count as one component of a broader editorial process. Removing the sticker brings the camera back; no procedure has been described that brings the percentage back.[1], [2]
The document missing on both sides is of the same kind: a description of method that states the threshold, the population it was measured on, and where a wrongly flagged person can turn. Spero says he plans to continue publishing technical reports on Pangram's models and training; unless those reports carry a false-positive rate matching how the tool is used in deployment — on excerpts under 100 words, on a median input of 350 words — the 0.0041 per cent stays a property of the evaluation set. On Meta's side the testable document is a description of how the tampering check works and what route a user whose camera was disabled in error can follow. My scepticism is a professional habit, but what an indicator establishes and who pays for it are separate questions, and only those documents answer the second.[1], [2]