Perell asks Macfarlane how he feels about ChatGPT and AI in writing. The answer is fast and clear: Oh my goodness. Look at those eyes. I mean, I don't and I cannot ever imagine doing so.1 He doesn't use the LLMs. He does tolerate spell-check.
Then he turns to Grammarly. My writing often is highly violating of grammatical norms... they're, like, the grammar check goes wild, does not like it, and I'm like — good. This is excellent.1
He demonstrates with a passage. The aerial-river description from over New Brunswick:
Seeing this huge Canadian river for the first time from the sky, my first sight of the river itself catches my breath. A world snake in the green. Cliffs dropping near sheer to water. House-sized boulders on the banks. Time falls from the rock faces above. Water blue black and glossy in the deeper calmer runs, peat brown where it stretched towards and away from rapids, churning green, gold, and cream in the falls.1
Five paragraphs, five sentences. Sentences doing weird stuff. Cliffs dropping near sheer to water — that's verbless. House-sized boulders on the banks — also verbless. Time falls from the rock faces above — that one has a verb but is doing something weird with abstract time.
Grammarly does not approve. Which I take to be a badge of success.1
Grammarly makes the average piece of writing, I think, much better. But it hurts a kind of writer. It takes away from distinctiveness and individuality that underlies so much great writing.1
The paradox is precise. Grammarly is genuinely useful for writing that is at-or-below-average. The average writer makes the kinds of grammatical errors and unclear constructions that Grammarly catches; the writer who deploys Grammarly's suggestions improves their average sentences. Net positive.
But Grammarly is calibrated against grammatical norms — the average of what counts as correct. Great writing routinely violates those norms in productive ways. The verbless sentence; the run-on that mirrors a state of being-run; the parallel-structure violation that produces a specific cognitive jolt. These violations are not errors; they are deliberate choices that produce specific effects no norm-compliant alternative could produce.
Grammarly cannot tell the difference between a productive violation and an error. The tool treats both as errors. The writer who follows Grammarly's suggestions regularizes the productive violations into the norm-compliant equivalents — and the prose loses what made it specifically theirs.
The cost operates at three layers:
Layer 1 — Sentence-level distinctiveness erosion. Each Grammarly correction smooths out a specific stylistic choice. Across a manuscript with hundreds of corrections, the cumulative smoothing produces prose that reads cleanly but unmemorably. The writer's voice is the sum of micro-choices; smoothing the micro-choices kills the voice.
Layer 2 — Learned regression toward the norm. A writer who uses Grammarly consistently begins to internalize its preferences. Over time the writer's own pre-revision drafts become more Grammarly-compliant because the writer has trained themselves to write the kinds of sentences Grammarly approves. The tool reshapes the writer at the input level, not just the output level.
Layer 3 — Norm-convergence at scale. When most writers use Grammarly, most published prose converges toward the same norm-pattern. The distinctiveness erosion is no longer just per-writer; it is industry-wide. The aggregate effect is a cultural reduction in stylistic diversity. This is the Why All Writing Sounds the Same Now claim in the title of the Macfarlane episode — though Macfarlane attributes the title to Perell, not to himself.
Macfarlane on LLMs is uncategorically against. The Grammarly-erosion argument scales: if grammatical-norm enforcement smooths distinctiveness, then full-text LLM generation does the same thing at much larger scale because the LLM is trained on the distribution of all prose and generates the statistical center.
Salman Rushdie said (Macfarlane quotes him approvingly): writers have nothing to fear from AI until it can do comedy. The hardest thing of all is to make us laugh; the hardest specifically-human craft is comedy. Macfarlane is less sanguine. I think writing broadly does have a lot to fear from AI and we should resist it, including the absolute piracy of our books that we now know large-language-model composers have been undertaking for years now and feeding without any kind of recompense or licensing or permission into the creation of LLMs.1
The piracy concern is independent of the distinctiveness-erosion concern but they reinforce each other. The LLMs that erode distinctiveness are trained on the very writers whose distinctiveness they erode. The structural double-bind is acute.
The implications are not "ban Grammarly" or "ban LLMs" — those are policy positions Macfarlane isn't directly making. The implications are at the individual-writer level:
The discipline is not against tools per se. The discipline is against unconscious tool-use that operates on the writer below the level of conscious craft.
Macfarlane converges with Karlsson's LLM-as-standardization-pressure on the principle that LLM-tools pull prose toward the statistical center and erode distinctiveness. Both writers reach the same finding from different craft traditions. Karlsson framing is from essay-craft (the wildness defense); Macfarlane framing is from lyric nature-writing (Grammarly-erosion). The convergence is striking and recent — both writers came to this position in 2025-2026 as LLM tools became unavoidable in the prose-craft conversation.
Macfarlane diverges from the LLM-assisted-drafting mainstream position (which treats LLMs as productivity tools that augment rather than diminish craft). The split is about whether the prose-as-craft and prose-as-output frames are the same project. The mainstream position implicitly accepts that prose is output; Macfarlane's position insists prose is relation, and relation cannot be augmented by tools that have no relation.
Grammarly-as-distinctiveness-erosion maps onto two adjacent domains where the same norm-pressure-as-individuality-loss mechanism appears: behavioral-mechanics (the homogenization pressure of algorithmic recommendation systems on content production) and creative-practice (the historical fate of folk-craft traditions that converged toward industrial-norm and lost their regional distinctiveness).
Behavioral-Mechanics: Algorithmic Content Homogenization — content creators on algorithmic-recommendation platforms (YouTube, TikTok, Instagram) face structural pressure to optimize for what the algorithm rewards, which is by definition the statistical center of what already performs. Creators who follow the algorithm's signals converge toward the platform's prevailing aesthetic; creators who resist the signals see lower distribution. Grammarly operates structurally identically at the prose-craft level. The structural identity: both crafts experience norm-pressure that systematically reduces distinctiveness at scale while improving median quality at the margin. The insight neither domain generates alone: Grammarly is not merely a writing tool — it is part of a broader algorithmic-homogenization pressure that touches every creative practice with norm-feedback. Writers who understand the algorithm-homogenization mechanism in content-creation can see Grammarly as the prose-craft instance of the same dynamic and resist it more consciously.
Creative-Practice (Folk Craft): Folk Craft and Industrial Norm Convergence — the historical fate of folk-craft traditions (rural pottery, regional weaving, vernacular building) under industrial-pressure shows the same homogenization mechanism over longer timescales. A region's pottery had distinctive forms because the makers learned from neighbors and worked with local clay; industrial pottery, with its uniform clay and pattern-book designs, displaced the regional distinctiveness as pottery production became more efficient. The folk forms survived where they were deliberately preserved against the industrial pressure. Macfarlane's Grammarly resistance is structurally identical: defend the regional/personal distinctiveness against the industrial-norm pressure. The structural identity: both crafts have experienced norm-pressure that improves the average product while extinguishing the distinctive forms, and both have generated preservation traditions to resist the pressure. The insight neither domain generates alone: Grammarly and AI text-generation are doing to prose what industrial pottery did to regional pottery — and the historical lesson from folk-craft preservation is that the distinctive forms only survive when there is deliberate cultural commitment to preserving them. Writers who understand folk-craft history have a longer-arc vocabulary for what they are defending.
A third connection worth marking: Grammarly-erosion has structural similarity to place-words and onomasticon collection — both treat the disappearing distinctive vocabulary as worth deliberate preservation work against the standardizing pressure.
The Sharpest Implication
If Grammarly's distinctiveness-erosion mechanism is real and operates below conscious craft, then writers who use Grammarly throughout their drafts are slowly losing the very voice that would make their work distinctive — and they are losing it without noticing because the tool's interface frames the corrections as helpful. The uncomfortable implication: most early-career writers in 2026 use Grammarly or LLM-tools as a default; many of them will reach mid-career as prose-writers whose voice has been pre-smoothed before they had a chance to develop it. The cohort effect could be a generation of writers who cannot find their voice because the tools they trained on prevented it from emerging. Macfarlane's verbless-sentence-badge-of-success isn't anti-tool stance for its own sake; it is a generational warning.
Generative Questions