Substack hands its readers a button that scans your post for AI text

A screenshot lands in an editor's inbox. Someone has run a scan on a contributor's piece and got back one word, AI or human or something in between ^[1]. The tool has finished its work at that point, and everything that follows is unpriced.

Fast Company reports that Pangram's technology routinely beats other AI detectors in independent testing ^[3], and that comparison is carrying more weight than it can hold. The coverage names neither the benchmark nor an error rate ^[4]. The field being beaten is the one in which an early iteration of GPTZero flagged the U.S. Constitution as AI-generated ^[16], because the first generation of detectors read predictable human prose as machine output ^[15]. Better than that is a genuine improvement, and it is still not a figure a managing editor can write a policy against.

Max Spero, Pangram's cofounder and CEO, frames the product around voice: "If you don't have an accurate AI detector, then there's no way to express displeasure that somebody else is sending you AI content" ^[11]. That describes the extension, which is where most people will meet the company, along with its flags on a prize-winning short story and on Pope Leo XIV's posts on X ^[13]. The volume sits elsewhere. Substack's button lives on the article page ^[7], NewsGuard is hunting AI content farms and the exam platform Inspera is passing over student essays ^[9], and the company raised $9 million in July to keep going ^[5].

Here is what teams tell themselves: detection is a procurement question, so the policy conversation starts when someone signs the invoice. Here is what actually happens: Spero points at viral posts and high-profile accounts himself and calls the role "slop janitor" ^[12], and the check now sits next to the writer's work where any reader can press it ^[7]. The first internal argument about a flag gets started by someone outside the building.

The subscription is $20 a month, which is $240 a year per seat ^[17], and that is the cheap end of this. Patrick Traynor, an engineering professor at the University of Florida who has studied AI detectors, states the expensive end as a question: "So you accuse somebody. Now what? What do you do with that information? What if you're wrong?" ^[14]. The seat buys a label. It does not buy the half hour with the freelancer or the student's appeal.

Two axes settle it for a specific team. The first is reversibility: does a flag feed a conversation you can walk back, or a grade and a kill fee. The second is corroboration: is there a second signal, such as revision history or a draft trail. Low stakes with corroboration is the only cell where a detector is useful on its own. Consequential with a single signal is where exam-platform use sits ^[9], and there the defensible move is to treat the flag as grounds to ask rather than a finding. The document worth having ready is short: what one flag entitles a manager to do, and who reviews the case when the flag is wrong.