The Prologue
The Substack Post
Arielle Swedback
Jul 24, 2025
A full year before Substack adopted Pangram, the platform surveyed its own writers about AI adoption. Nearly half, 45.4 percent, were using it, though mostly for research, assistance, and ideation rather than wholesale generation. Adoption ran lowest in Literature, Music, and Art, and the primary concern, amongst both AI users and refusers, was around how AI erodes the author's voice. The report's prediction is worth taking a beat: human-authored work is treated as a premium product, as machine text piles up.
Read →
Pangram
Max Spero
Jul 9, 2026
Using their own AI detection tech, Pangram scanned over a million social media posts. The results weren't shocking: one in four long-form posts (those over 250 words) are generated by AI, and LinkedIn accounts for nearly two-thirds of everything flagged despite being only a third of the sample. Pangram's study is an advertisement for their service (in the blog post, they also floated the claim that their software carries a 0.01 percent false-positive rate) but the results certainly feel right.
Read →
Stat Significant
Daniel Parris
Jun 24, 2026
AI mimics us, and we, in turn, begin to mimic AI. Researchers at Berlin's Institute for Human Development found LLM-favored terms like "delve" and "bolster" spiking in human prose, and Daniel Parris's Stat Significant piece tracks the subtler rewiring beneath the obvious slop: semantic similarity across web pages rising in lockstep with AI prevalence, college essays gaining measurable polish the year ChatGPT arrived. The machine's patterns are inside human writing, and the internet is quickly converting into a single, overconfident voice. Fatigue is beginning to set in.
Read →
Where We're At
The Substack Post
Chris Best
Jul 21, 2026
"It's getting harder to tell what's real on the internet. This is less true on Substack, and we aim to keep it that way." So opens the policy, signed by CEO Chris Best. His solution was to allow users to scan any text over 100 words published after the feature's launch day. Best frames the problem as a matter of transparency, a mismatch between a reader's expectation and reality (in his terminology: Claudefishing), and he offered writers the option to craft a "How I make this" statement.
Read →
TechCrunch
Sarah Perez
Jul 30, 2026
Nine days after Substack's announcement, LinkedIn joined the party, offering readers the option to report writing that "seems like AI slop," and that signal helps guide the algorithm that determines post distribution. In the same breath, the company retired their "enhance your post" feature that allowed users to rewrite their post with AI, replacing it with a feature that proofreads without changing a voice. Hari Srinivasan, the product chief making the case, says the platform already blocks hundreds of thousands of automated comment attempts every day.
Read →
The Backlash
404 Media
Emanuel Maiberg
Jul 28, 2026
The reception for both announcements was mixed, and, for Substack, the backlash arrived on two fronts that pulled in different directions. Writers who use AI were concerned about the stigma now attached to their work ("I'm not going to apologize for using AI in the creation process," writes one Substacker), while writers who never touch it fear a single false flag could wreck a reputation. The latter camp doubted the veracity of Pangram's accuracy claims. The phrase that stuck, "witch hunt," came from a professor's post title, and nobody, the platform included, has defined "AI-assisted."
Read →
Inc.
Chris Morris
Aug 3, 2026
If Pangram was interrogated on a more technical level, the LinkedIn solution was structurally flawed. By not using a wholly deterministic tool, and letting their users make the call, they may have been trying to tease out the difference between the slop churned out by bot farms, and the more intentionally crafted prompts used by the executive class, but their solution relied on the subjectivity of human judgment, and humans are likely to use the feature as a big "dislike" button. MetaCert founder Paul Walsh warned of "an avalanche of false positives," and LinkedIn's own product chief concedes that slop is hard to define and the definition keeps changing. Meanwhile, Originality.ai estimates 81 percent of the site's longform posts in July came from AI, up from roughly half in late 2024.
Read →
Natural Intelligence
JHong
Jul 27, 2026
Of course, humans aren't the only ones with biases. JHong underlines the fact that Substack is using AI software to combat the use of AI software, and the cultural biases that researchers have flagged in other AI programs are present in Pangram. A false positive result isn't a bug report, it's a person, and disproportionately those people are second-language and neurodivergent writers. One study flagged 61.3 percent of real TOEFL essays as AI. As she concludes, a machine that scores methodology isn't defending writing, it's shipping a product.
Read →
The Third Hemisphere
Tim Requarth
Jul 31, 2026
After initially questioning the veracity of Pangram's accuracy claims, Requarth grants that they're true…in testing, but lab efficacy is not real-world effectiveness, and a tolerable rate of false-positives is a values question. Requarth puts this in the context of forensic science, and argues that none of the same checks-and-balances that apply to that field are available here. Requarth also echoes the concerns around second-language and neurodivergent writers (providing a real-world example in Emma Klint), and makes the case that the problem with Substack's solution isn't so much the precision of the tool, but the lack of appeals and oversight processes that provide checkpoints on forensic science. All of this leads to a dynamic that feels less like a nuanced conversation and more like a…witch hunt.
Read →
Rejecting AI Writing
The Atlantic
Deb Roy
Feb 15, 2026
If the conversation lacks nuance, it certainly isn't because we've failed to consider (and articulate) the differences between AI and human writing, and the next few articles illustrate the intellectual boundaries of the resistance. The first piece considers how the moral and emotional stakes differ between man and machine. Roy spent years building robots that learned words by grounding them in the senses, and came away convinced the grounding missed what makes language bind: the vulnerability of a speaker who can be hurt, shamed, and held accountable. A chatbot apologizes flawlessly with nothing at stake; its words sound responsible and even convincing, but are empty. "To mean what we say is to risk something," Roy writes. Speech without consequence, Roy warns, trains us to accept meaning without accountability, and this erodes the moral structure of language faster than our norms can adjust.
Read →
Unpublishable Papers
Eli Stark-Elster
Jul 17, 2026
Eli Stark-Elster wanted to love a New York Times essay on Messi's retirement, bylined to the poet Rowan Ricardo Phillips, but the tells stacked up quickly: the tripartite structures and metaphors that are nearly legible at skimming speed. What ultimately convinced him was the piece's flatness, nothing on the Argentines who burned Messi's jersey, or the stare during the announcement. AI is great at conducting information, but "we write to share our souls; we read to receive the souls of others."
Read →
One Useful Thing
Ethan Mollick
May 26, 2026
Mollick certainly isn't an anti-AI "luddite," but he thinks of AI prose as a "meaning-shaped attention vampire" that reads like thought and parrots meaning (i.e., some deeper synthesis or unique understanding of its subject) but ultimately delivers neither. And the cost runs in both directions. Grasping for words and wrestling with language is the process by which we understand something, and a writer who outsources large swaths of their process to an AI will never grasp their subject.
Read →
404 Media
Jason Koebler
May 11, 2026
Jason Koebler covers AI slop for a living, and it's increasingly difficult to leave his work at the office. His brain, he writes, has become "the AI police," scanning everything for the uncanny: a podcast intro that feels stilted, impersonal apology texts from friends, even wedding speeches that sound mechanical and raise suspicion. None of this requires him to open a chatbot, and his immersion in generated prose has changed how he reads everything, including the human stuff.
Read →
Is It Supposed To Look Like That?
T.M. Brown
Jul 6, 2026
If the other pieces in this section provide the philosophical, spiritual, or emotional grounding for the resistance's objections, this one looks at one writer's pocketbook. The likely economic fallout for writers, he writes, is probably not on literary fiction, but on the commercial tier: white papers, SEO copy, and product prose that kept working writers solvent. A client's first question is whether you use AI, and those gigs now fetch a fraction of their old rates, leading to "the specter of there being no way to make a living as a writer at all."
Read →
The Adaptation Case
The Atlantic
Jasmine Sun
Mar 17, 2026
There's a growing number of writers and public intellectuals (along with publications and agencies) looking for a middle ground between an outright rejection and blanket adoption of AI. The most interesting of these don't dispute many of the claims made above, but make the case that AI does have unique capabilities that, if harnessed, can be effectively incorporated into a creative workflow. The prodigiously talented reporter Jasmine Sun adopts this stance here, initially rejecting an LLM's capacity to write anything truly interesting (they're effectively teacher's pets whose strangeness has been sanded away in post-training) and embracing their capacity to serve as an editor, after being trained on her own archive. This is one of the few pieces built from reporting, rather than assertion, and is a disciplined take on the subject.
Read →
The Atlantic
Stephen Marche
May 9, 2023
Stephen Marche actually did the thing that the discourse keeps imagining. Commissioned in 2023 to produce a novel that was 95 percent AI-generated, he found asking for Raymond Chandler yields cheesy Chandler. Instead, he conjured Chandler through a prompting process that worked its way through Chinese nature poetry and Ernest Hemingway. Marche found the process more akin to curation or sampling than traditional writing, and he draws comparisons to how sample-based hip-hop builds art from the shards and scraps of other people's records. Ultimately, this hybridized approach demanded more understanding of literary style and not less. His verdict was that creative AI "is going to change everything. It's also going to change nothing."
Read →
The Believer
Vauhini Vara
2021
From the Archives
For years, Vauhini Vara struggled to write about her sister's death, so she let GPT-3 begin. "Ghosts" is the experiment: nine versions of the essay, Vara's sentences in bold, with the machine completing each one. Early versions produce a tidy recovery story ("She's doing great now"), a fluent lie. Version by version, her own share of the story grows, and the details turn real, until, in the end, the story is hers. This isn't so much an argument as it is an artifact, and this classic of the genre (it's from 2021) provides the strongest possible answer to "AI writing is inherently hollow."
Read →
A Glimpse Ahead
Rhetorica
Marc Watkins
May 3, 2026
Even after four years of intense dialogue and constant conjecture, it's still difficult to see where this is all heading, but there are some signs. Marc Watkins installed Pangram's Chrome extension in May, before any platform adopted it, and quickly found his reading habits were color coded. Reds went unread, while the greens passed the gate. The software flagged vocal AI critics as machines, and his suspicions superseded his judgment. He now rejects technological fixes for technological problems. A label, he writes, is another way of saying "let us do the thinking for you."
Read →
Wired
Jason Parham
Jul 29, 2026
Above, we mentioned how AI-speak is encroaching on how actual humans write, but there's also a different phenomenon taking place. Writers are beginning to reject the strictures of style and grammar, leaving in typos or adopting more formally adventurous prose as a sort of "proof of life." And, of course, there's an app for this: Sinceerly, the anti-Grammarly that will intentionally inject errors into your prose. It's easy to laugh and shrug this off, but it's also potentially exciting (and invigorating) to imagine new forms of writing sprouting up from the ashes of this debacle.
Read →
Maggie Appleton
Maggie Appleton
2023
From the Archives
In 2023, Maggie Appleton imagined a dystopian future where generative text floods the open web and public spaces are rendered as lifeless shells of their former selves, a hauntology of AI bots spewing well-crafted nonsense to one another. She predicted readers would become sceptical of one another's humanity, and grow increasingly paranoid, and considered some type of institutional verification before dismissing it as fraught and impractical. Sound familiar? She called this flooded open web the "dark forest," and imagined a "cozy web" where humans retreat to commune with one another. It's a map for the future, of sorts.
Read →
wreckage/salvage
Erin Kissane
Nov 22, 2024
Towards the end of Appleton's piece, she briefly acknowledges that the retreat she describes is closed to the disabled or those in isolated spaces. Kissane takes this caveat and turns it into a spear she wields against Appleton's metaphor. In her mind, the metaphor dresses up surrender as wisdom. Our predators have names and employers, platform leaders are "gold-rush-addled dorks" rather than shadowy forces, and building these "dipshit accelerators" was a choice somebody made. The "cozy web" assumes a level of privilege, and the dark forest is just a basement huddle. Instead, she wants a real forest, one of interconnected and locally governed ecosystems at many scales. She's here to fight, and argue.
Read →
