The Pangram Word Police: Substack’s AI Detector

On July 21, Substack announced an unholy partnership with a word police contractor called Pangram, whose business consists of examining prose for evidence that it may have been assembled by an artificial intelligence rather than by the traditional method of disappointing one’s parents and becoming a writer.

Pangram’s mission in life is to determine whether a piece of writing originated with a human being or with a machine directed by a human being. It is, in effect, a private investigator hired to follow every sentence home.

Substack has framed its AI detector as a service to readers. It is, in truth, a lie detector test administered by an algorithm and marketed by a company that built its reputation on the promise that no one would ever look over a writer’s shoulder again.

The following offenses are believed to attract Pangram’s professional interest.

Tier-1 Offender. Proper Semicolon Usage.Subjects display unusual confidence in independent clauses. Their punctuation suggests either an expensive education or an unhealthy attachment to The Chicago Manual of Style. Recommended action: continued surveillance and mandatory insertion of one casual “lol.”

Tier-2 Offender. Contraction Deficiency.The subject habitually writes it is instead of it’s, as though negotiating a hostage release rather than composing a newsletter. According to a 2025 internal review conducted by Pangram’s research division, and cited with evident satisfaction in Pangram’s own promotional material, writers who avoid contractions are 340 percent more likely to be flagged than writers who do. This statistic says considerably more about the review than about the writers.

Tier-3 Offender. Excessive Structural Integrity. A writer who avoids contractions and uses semicolons correctly is no longer merely suspicious. He is considered “structurally consistent with non-human output,” which is a polite way of saying he writes like he means it. Paragraphs begin where expected, proceed logically, and conclude without emotional collapse. Human probability is downgraded accordingly.

Tier-4 Offender. Folks Avoidance.The subject never uses the word folks, which has become, over the past eighteen months, the internet’s preferred theatrical gesture of sincerity. Executives use it. Influencers use it. Customer service representatives use it immediately before disappointing you.

A writer who does not say folks is not merely withholding a word. He is declining to perform the approved ritual of internet warmth, and Pangram has been built, at considerable expense, to notice such things.

Consider the case of Warren Deshler, 61, a retired actuary from Dayton, Ohio, who once calculated, purely for his own amusement, the exact odds of being struck by lightning while filing a tax extension.

Deshler has written a weekly Substack newsletter about municipal bond markets since 2022. He does not use contractions. He does not say folks. He uses semicolons the way other men use handshakes, firmly and without apology. Under the new Substack-Pangram taxonomy, Deshler qualifies as a Tier-4 Offender on prose style alone, despite being, by every available account, entirely human.

“I have been accused of many things,” Deshler said. “Artificial intelligence was not on my bingo card.”

Or consider Priya Raghunathan, 34, a labor organizer in Queens who keeps a small collection of vintage subway tokens on her desk. Raghunathan writes in short, declarative sentences. She trusts her premise. She exits quickly.

Pangram’s detection model, trained in part on the assumption that brevity itself may indicate synthetic authorship, flagged three consecutive posts during its first week of deployment.

“Apparently,” Raghunathan observed, “I am too concise to be alive.”

Substack described the incident, in a statement notable for its complete absence of the word sorry, as “expected variance.”

It should also be noted that Pangram’s own founding paper, published in 2024 under the magnificently confident title Toward a Unified Theory of Synthetic Text Detection, acknowledges an error rate the authors describe as “statistically nontrivial.”

The paper further recommends continued refinement of detection methodologies, which is academic language for, “We know it misses people. We intend to miss them more accurately.”

Substack did not build Pangram. They merely invited it inside, handed it a clipboard, and asked it to determine which writers appeared sufficiently biological.

There is an irony here. A platform that spent years encouraging writers to abandon institutions, ignore gatekeepers, and cultivate a direct relationship with readers has now installed a gatekeeper whose principal qualification is statistical suspicion.

The classification system remains preliminary. It will undoubtedly expand.

There is already quiet speculation about a Tier-5 designation for writers who begin paragraphs with “Consider.” Tier-6 may be reserved for anyone who voluntarily revises a first draft. Tier-7 could encompass writers who know the difference between less and fewer. Tier-8 may include anyone capable of sustaining an original thought for longer than one hundred words without mentioning dopamine, productivity, or “building in public.” Tier-9 will likely be reserved for prose that remains internally consistent across multiple paragraphs. Tier-10, according to documents that almost certainly do not exist, identifies writers whose humanity has become statistically improbable and recommends immediate replacement with someone who says folks every fourth sentence.

If the ramblings of our fearless editor in briefs fancy your tickle, wander over to his Satirical Commentary page and try a line or two.

Editor’s Note: The author of this piece writes economical prose, uses semicolons without shame, and has never once said “folks” in earnest. Draw your own damn conclusions.