The Curious Case of Pangram [guest post]

• The Case For: Chris Best, Sarah Fay

• The Case Against: Dr Sam Illingworth, Kelly Courtez, Lisa Kostova, Sara da Encarnação, JHong, Freddie deBoer, Lipgloss and LLMs, The Strategic Linguist, Neela 🌶️, The Raven Scholar

• The Legal Case: Luis Campos AI Engineer, Carey Lening

• The Nuance: Emma Klint 🦆, Fran Davis, Rebecca Mbaya

• Bonus Links: more than 20 other authors, tagged below

A natural place to begin is the post that started it all - Against Claudefishing by Substack’s co-founder and CEO Chris Best.

The backbone of the article, and the stated reason behind the partnership, is a term Best coined himself — “Claudefishing”:

“a mismatch between a reader’s expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end.”

To give Chris credit, he makes a solid case for safeguarding the platform from LinkedIn-style ensloppification, while acknowledging that many writers use AI to help them produce thoughtful, high-quality work. His core point is transparency of provenance. Pangram isn’t about blanket-banning AI, he argues; rather, it’s about letting readers know what to expect.

Best closes by saying that the Substack team is open to feedback. They kept the promise, updating the feature to let writers disable AI scans without running a scan first, shortly after many writers pointed out that the “disable” option was hidden behind the “scan for AI” button. Chris’s last line — “when I want Claude’s opinion, I’ll ask Claude” — re-emphasizes his key message: what makes writing on Substack valuable is the human judgement behind it.

While Best provides the reasoning behind the integration, Sarah Fay takes a more proactive approach in The Truth About Substack’s New AI Detection Tool. She shares what readers and writers need to know about the tool, and explains how to create a “How I Make This” statement. She also made her post on writing an AI policy accessible to free subscribers. Echoing Chris’s reasoning, she calls for more transparency regarding the use of AI by Substack writers:

“It’s all about transparency and using AI in a way that is integrity for you. If you don’t want people to know you’re using it, then it’s time to check in with yourself.”

Perhaps my feed is biased, but I didn’t come across many posts supporting Pangram integration. If you’ve written one, or know someone who has, link it below; hearing all sides of the debate is all this is about.

Since the discussion referenced lots of witch-related imagery, I’ll let Dr Sam Illingworth open with a sharp analogy: The Crucible. In Does AI Detection Work?, he compares Pangram’s scan to Salem witch trials, where the very accusation is the evidence. The person accused, in this case a writer assigned a high AI score by Pangram, is the one the burden of proof lies on.

Although Dr. Sam acknowledges Substack’s careful approach to AI detection, he still argues that the tool relies on the assumption that there is “hidden guilt” to be discovered:

“Chris is careful. He says the tool shows results only to those who ask, that Pangram is not perfect, that people should make their own judgements. Salem was careful too, in its way. It had procedures. Underneath them sat a prior belief that hidden guilt was everywhere, and a will to build the device that would find it. The care rode on top of the suspicion and gave it a clean face.”

He then cites research that once again confirms what we’ve known for a long time now: AI detection doesn’t work. It labels human edits as AI, misses heavy AI use, and can be easily fooled by humanizers. Dr. Sam also reemphasizes how strongly biased detectors are against non-native English speakers and neurodivergent writers. He ends with a thought that is hard to disagree with: if you want to know if an author used AI in any way, you can ask them. Let the human speak for themselves instead of relying on an unreliable percentage.

Extending Dr. Sam’s witch analogy, Lisa-Marie Cabrelli, Ph.D.’s post Nobody Made a Detector for the Rest of It compares Pangram to the witch pricker’s bodkin — a rigged tool used by men to tell if a woman was a witch or not. Plot twist: those men were paid per conviction, meaning the more witches they exposed, the more they earned.

“A fee for every mark found, every conviction secured. So the instrument that “objectively” detected guilt was wielded by a man whose income depended on detecting it.”

If this rings a bell, that’s because Pangram gets paid per API call, i.e., every time someone runs a scan. It doesn’t protect us from AI slop, Lisa-Marie argues: those who produce actual slop simply aren’t going to care. It only punishes those who do genuine work by making them question their authenticity, something women have been forced to do since the dawn of time. Her advice? Don’t submit to the judgement.

Kelly Courtez’s approach in I Don’t Care About Pangram is to ignore it entirely. Just as you would ignore internet comments hating on a song you love, she argues, you should ignore a score assigned by a machine to a piece that moved you. The more we compare our own opinions to external verdicts, the more we lose our ability to think for ourselves - one of the most crucial skills in this day and age:

“One person can read a paragraph and feel nothing. Someone else can still be thinking about it days later because it brought back something the writer never considered. Check that reaction against everyone else’s verdict often enough and, before long, recognising the approved opinion comes more easily than trusting your own taste.”

Kelly’s message is: trust yourself. A big tech company shouldn’t dictate what you are and aren’t allowed to like.

Lisa Kostova makes a similar point right away in the title: Substack’s Scarily Accurate AI Detection Will Lead To Our Collective DUMB-ification. Her core worry is atrophy: just as GPS has been shown to weaken spatial memory, outsourcing judgment to a detector will dull our own critical thinking over time. She calls out the illusion of precision as well:

“The results are presented as percentages: “8% of this text is AI,” “37% likely AI,” “92% human.” These numbers create what researchers call false precision — the presentation of data with a degree of specificity that exceeds the reliability of the underlying evidence.”

Readers searching for some authority to help filter out low-quality noise will be more inclined to trust the verdict, even though the percentage is, to some extent, a guess.

In the last section of her article, Lisa offers practical advice on how to protect your agency, both as a writer and a reader. Spoiler: running Pangram scans isn’t part of it.

Sara da Encarnação’s Human 0% draws a parallel between Pangram and the “blue pencil” of Estado Novo-era censorship. The blue pencil forces writers to preemptively mutilate their writing to avoid being flagged: remove em-dashes, get rid of antithesis. Sara also makes a distinction between resemblance and provenance. A machine can only spot statistical patterns; it can’t witness the actual thought, struggle, and revision behind a piece of writing. Handing a machine the authority to judge a person’s honesty is indecent, she believes, regardless of how accurate the detector is. The consequences of such labeling, Sara argues, are destructive too. Readers, seeing ‘proof’ of a work being AI-generated, may simply stop rearing without an official ban — a kind of “dispersed censorship” with no actual censor to hold accountable:

JHong uses a framing from Blade Runner 2049, comparing Pangram to a “baseline test” - a test that decides whether you’re still allowed to be counted as human. She breaks down the out-of-context false-positive rates, the publisher agreement, the financial stakes - pretty much everything Pangram probably doesn’t want you to pay attention to. Another point worth highlighting is the detector-humanizer tail chasing. Jennifer references Goodhart’s law — “when a measure becomes a target, it ceases to be a good measure” — to explain the eternal cycle of detectors built to detect AI writing, humanizers built to hide AI writing, detectors built to recognize humanizers, and so on. Writing quality deteriorates, she argues, while tech companies profit.

She also echoes Sara’s point about writers’ reputation being silently eroded when their writing is flagged as AI:

“The “Scarlet AI” travels. The research on this: once writing is attributed to AI, readers rate the author lower on competence, morality, and future success. The flag doesn’t stay on the Post. It classifies how the writer gets perceived as a person. And once someone is badged, trust deteriorates, subscriptions are canceled, and Substack has now worked against its stated objective to increase paid subscriptions to writers.”

An interesting entry comes from Freddie deBoer who, ironically or not, is on record hating AI writing. In I Wouldn’t Say Pangram Is Broken, But I Would Say It’s Brittle, he documents a “Russian doll of contradictory results”: a 300-word passage from one of his own essays got flagged 100% AI, while the full 5,000-word essay containing that exact passage scored 100% human. His “Frankensteining” test — stitching human and AI text together — showed that the tool doesn’t return mixed percentages so much as flip between two extremes. My favourite part from his post is:

“The right analogy isn’t a stress test that a bridge passes at loads no truck will ever carry; it’s discovering that the bridge fails at a load some trucks routinely carry, simply by virtue of being those trucks, and being told it doesn’t count because you drove the truck onto the bridge with the express purpose of overloading it.”

Like most other writers featured in this section, Freddie is blunt about the social risk. He notes that writer culture already runs on petty grievances and jealousy, and handing that culture a tool promising to expose ‘frauds’ all but guarantees it gets used as a one-shot gotcha machine.

Lipgloss and LLMs makes a similar case from the other direction. Running a personal, metaphor-heavy essay through the scanner, the writer expected nothing given how deeply theirs it felt, and got back 100% AI, while a university thesis riddled with errors scored 100% human. She observes how Pangram goes against everything Substack stands for: since AI models are trained on the internet’s best prose, the more distinctive your writing is, the more likely you are to be mistaken for the very thing that copied you. She concludes with two questions:

Rebecca from The Strategic Linguist analyzes AI detection through the lens of linguistics in Written As If Watched. Her main objection is a category error: an Agent acts with intent, while an Instrument — a pen, a dictionary — is just a tool. AI-assisted writing, according to this framework, is an agent using an instrument, but Pangram scores it as if AI is a co-author with its own agency. Like Lisa Kostova, Rebecca doesn’t buy the credibility promise. The tool has face validity, since percentages and confident visuals make it look rigorous, but not construct validity, since “AI-assisted” and “AI-written” are two different relationships that shouldn’t be collapsed onto one scale.

Where Sara da Encarnação used the analogy of a “blue pencil”, Rebecca uses Foucault’s and Bentham’s concept of a Panopticon:

“a prison design with a single guard tower at the centre of a ring of cells. Every cell is visible from the tower. No prisoner can tell whether the tower is occupied. So the prisoner behaves as though it always is.”

This analogy illustrates how the mere possibility of being scanned pushes writers into “docile prose”. She calls it “verbal hygiene,” once again reminding us that it’s harshest on writers whose voice is already different from a narrow, dominant standard of English.

Neela 🌶️’s take, The Ministry of Other People’s Business, might be the wittiest one in this section. Drawing on her childhood in Trinidad, she portrays Pangram as an API reincarnation of Miss Millie, that one village lady who minds everyone’s business from her porch. She traces the same phenomenon through history, from “Grundyism” to Anthony Comstock’s career as a self-appointed vice enforcer, later dubbed as “Comstockery.”

The internet, Neela explains, removed the one thing that used to keep gossips in check: the proximity of the village:

Needless to say, this asymmetry is not in favour of the writer.

And in the spirit of Miss Millie’s image, I’ll conclude this section with a note by The Raven Scholar, which somehow summarizes what many of us feel like in a single picture:

(Image posted by The Raven Scholar showing the character of Dolores Umbridge from the Harry Potter book series with the stylized text “If AI Detectors had a face”. Riff on the “If AI was a person” meme showing the same character.)

Some authors took a different approach. Rather than examining the pros, cons and ethics of Pangram or AI detection at large, they looked into whether Substack’s integration of the tool is even legal.

In Major Injuries with Pangram, AI engineer Luis Campos AI Engineer argues that the integration violates multiple international laws, such as the EU AI Act and Brazil’s LGPD. These frameworks guarantee a right to human review of decisions made by automated processing - something Substack and Pangram don’t seem to provide. Luis states that even though scans are only available on demand, it doesn’t exempt the developers from accountability. He points out that some legislations require companies to conduct impact assessments, evaluating how users can be affected or harmed by automated decision-making (in this case, AI scans). In China and the Canadian province of Quebec, the assessment must be done prior to the release. In the US state of Colorado, such assessments have to be conducted annually.

Carey Lening focuses specifically on GDPR in her post One Weird Trick for Opting Out of Pangram on Substack. She explains that the tool constitutes profiling, because readers naturally make judgements about a writer’s behavior, reliability and worthiness based on a percentage determined by an AI. For some writers, especially those who earn a living via paid subscriptions, this can have measurable economic consequences. Based on this, under Article 21 of GDPR, creators have a right to object to this processing.

Carey suspects that Substack relied on ‘legitimate business interests’ to have the feature enabled by default, but she argues that such interests don’t override a creator’s right to not be automatically profiled. In the second half of her post, she shares a template of an email that EU- and UK-based writers can send to Substack to demand more transparency about Pangram’s processing of their data.

If Substack acts as an actual business or personal brand for you, rather than just a hobby, these two articles are worth taking a closer look at.

Last but not least, there are some articles that didn’t fit neatly into the “pro/against” binary. In fact, I’ve left them for the very end because some of them raise questions bigger than Pangram itself.

Emma Klint 🦆 doesn’t dwell too much on whether Pangram is good or bad. Instead, her essay I use AI to write everything I publish here. Substack now grades that. explores how conversations with AI helped her regain the playfulness of creating something out of nothing, just for the sake of it. As a kid, she used to document local soccer games in a pink notebook. Now, she describes AI as the notebook that can talk back:

“AI can keep going for as long as I need. It holds the thread while I orbit what I am really after and gives me a version of the thought to push against. Sometimes it finds the word I couldn’t reach. Other times it gives me five paragraphs that are technically correct and make me feel absolutely nothing, which is useful too. Now I know what I don’t mean.”

Emma joins other writers in saying that a Pangram-generated percentage can’t reflect the whole story behind an article’s making. She doesn’t call for the readers to not use Pangram, nor for writers to disable it. Her whole publication, she says, is a testament to what happens when someone follows a thought to wherever it leads. So the article leaves us with an implicit (and perhaps rhetorical) question: is it fair to judge someone for using that pink notebook?

Fran Davis takes a similar approach in her post The Thinking Is Mine, except she connects it to a different kind of work: family history research. Her method — Capture, Connect, Reflect, Refine, Create — got developed throughout years of searching, hitting walls, and getting things wrong. When AI arrived, it merely followed this process with her. She echoes Emma’s take about stories that would’ve never been told otherwise:

“Family history is full of ordinary people whose stories were never told well, not because they weren’t worth telling, but because nobody in the family had the craft, or the time, or the confidence to tell them. Those stories have been waiting generations for someone with the tools. If the thing that finally gets them written is also the thing that gets them flagged, we lose them twice."

From this perspective, AI is the tool that finally makes it possible for the overlooked stories to get written. So let them be written, Fran says. Her piece invites you to tell a family story you didn’t have the words for before, now that you have the tool to help you find them.

Rebecca Mbaya states plainly that she’s not in either “for” or “against” camp. Her piece To Pangram or Not to Pangram? (100% Human, T&Cs Apply) is neither a critique nor a defense of Pangram. It’s a sharp observation on the priorities of a platform that calls itself the “economic engine for culture”. Despite existing solutions, Substack still hasn’t made monetization accessible for African writers. An AI detector got shipped sooner.

Some Substackers have said that as paid subscribers, they’d like to know whether the work they’re paying for was made with AI. Rebecca puts this argument into perspective:

“A tool gets built to solve a problem someone already has the infrastructure to have, verifying authorship assumes you’re already getting paid enough for authorship to be worth verifying. Meanwhile the more basic problem, actually being able to participate in the economy the tool sits inside of, stays unsolved, quietly, indefinitely, because it’s harder, slower, and less exciting to fix than shipping a new badge next to someone’s byline."

For many African creators, Rebecca explains, readers’ attention is the one thing they still own. Now Pangram has come to take it away, too.

She invites Pangrammers to scan her work for AI if they want. But she also invites them to ask themselves: can she even get paid for the piece they’re scanning?

It’s hard, but necessary, to draw a stopping point for folding in more articles; there was an initial flood, and new ones come out every day. Here are some other excellent articles we read and enjoyed but didn’t have time to include in depth:

• Alyssa Fu Ward, PhD plus 10 others, “How we human in the face of AI detection: A constellation of voices”

• Amberhawk, “Trust Isn’t a Number”

• Anna • bubble boss 🫧, “I Don’t Work for the Scanner”

• Apostolos Stamenos, MS, “A New Polarized Economic Engine for Abuse Culture”

• Daniel P. Hirschi, “There Is a Human Behind These Words. Your Tool Can’t See Him.”

• Her Majesty of Ink and Exits, “Different Chair, Same Tool”

• Jade The Hooman, “What Happens to Writing When Everyone Is Being Scored?”

• Jennie O'Connor, “I Have Opinions, and They Are That This is Some Bullshit”

• Jordan Harrod, “re: Against Claudefishing”

• Karo (Product with Attitude), “AI-assisted Craft: A Manifesto”

• Kim Doyal, “Substack Wants to Know How Much AI I Used. I’d Rather Show You What I Built”

• Limited Edition Jonathan, “Nothing Was In The Water”

• Linda Caroll, “The Pangram Gong Show”

• Melanie Goodman, “How LinkedIn and Substack detect AI content in 2026”

• Monica Hebert, “Nobody Asks a Painter What Brush She Used”

• MrComputerScience, “AI Got Judgy, Got Caught Hacking, And Almost Got Someone Killed”

• Niki Bowen, “The Irony of “Scan for AI Text””

• Noemi - Legal genie, “Substack failed to notify of Publisher Agreement updates”

• rosie, “thoughts on ai coming to substack”

• Ruv Draba, “Pangram isn’t perfect — and wouldn’t work if it were”

• Sam Rees, Reg.Psych, “What Substack’s AI Detector Gets Wrong About Behaviour”

• The Raven Scholar, “The Humanity AI Detectors Keep Erasing”

• Tim Requarth, “What a Better AI Detector Doesn’t Fix”

• Valentina, “Against Chromatography”

• William R. Crichton, “The Pangram Witch Hunt”

• Yana G.Y., “Substack’s new AI detector Pangram said my AI-assisted post is 100% human”

We also wrote our own articles, not reflected in the analysis above:

• Daria's Tech Musings, “Let the Pangram Games Begin”

• Karen Smiley, “Why I won’t be using Pangram AI scoring”

Karen here. Thanks to Daria for the collaboration, and thank you for reading Everyday Ethical AI! If you’ve written a piece on Pangram that didn’t get featured or listed in this article, feel free to share a link in the comments, or just let us know what you think in the comments or in this poll:

For the record, the policy “statement” of this newsletter is:

Leave a comment