Some of my oldest habits as a human creator have become markers of AI slop.
Example one: I love the em dash. I use it all the time. But for some reason LLMs, especially early chat models, developed a conspicuous fondness for the em dash—often in places where I would use an ellipsis or a colon—and suddenly thousands of my farm-to-table human-written words began setting off self-appointed “AI slop detectors.”
I sometimes find myself changing punctuation not because it improves the prose but because it improves the signal of the prose: it helps people understand that a human wrote it.
Example one point five: Some readers have similarly learned that LLMs seem to lean toward long, syntactically complex sentences—presumably excepting the LinkedIn posts that try to manufacture profundity with line breaks. My teachers from about fourth grade onward had the same comment about me.
Example two: VerseNotes, the website I started in 2018, has an off-white background intended to be reminiscent of parchment while remaining highly legible. (The first internal version, code-named Siloam, had a soft green color scheme reminiscent mostly of algae.) In 2025 and 2026, LLM-based coding tools—especially Claude Code—have used that off-white parchment-y background on all kinds of vibe-coded sites.
It may be that LLM “taste,” such as it is, runs to the mean. But the models arrived at that mean by absorbing choices humans were already making for reasons of their own. My web-design skills are modest at best, but the choices I made for VerseNotes were considered and relevant to the work: parchment for ancient manuscripts; purple to evoke royalty. The machine arrived later, imitated the choice at scale, and made the earlier human work look derivative.
So I am dismayed when I see implications—as I did this morning when I finally read this piece by Nolen Royalty on legibility of effort from June—that the more than 200,000 words I’ve published won’t get a fair hearing because they live on an eight-year-old color scheme that resembles vibe-coded sites from the last eighteen months or so.
Royalty is right that LLMs have blurred many of the cues by which we once inferred that someone cared. But “legibility of effort” identifies the loss too narrowly. Effort was always a weak and error-prone proxy for care.
“This was hard to create” and “this is good” and “this brings me joy” describe different things. They sometimes coincide, but none guarantees either of the others. Some people enjoy the fact of the effort itself—I occasionally count myself among them—but laborious creations can be joyless or poor, while easy creations can be delightful or enormously valuable.
One creator may use generative AI tools with great care, choosing, verifying, revising, and accepting responsibility for every part of the result; another could type every word on a Remington and emit careless slop. When creators must demonstrate legibility of effort to signal the worth of their work, we should expect performative friction: typos, ugliness, conspicuous inconvenience, and other difficulties added primarily to prove that someone suffered.
Readers and audiences of all kinds inevitably use heuristics to decide what deserves their attention. The epistemic laziness begins when those heuristics harden into verdicts—when an em dash, a complex sentence, or a cream background becomes proof that a work was machine-made or careless. At that point, you’re mistaking a shallow stylistic stereotype for evidence about provenance.
Style itself can be substantive—Emily Dickinson’s punctuation and Aaron Sorkin’s cadence are part of what their work means—but style is not a reliable authorship detector. An em dash may shape the meaning and rhythm of a sentence, but it cannot tell you whether a person wrote it. If you engage with the work rather than treating isolated stylistic features as clues to provenance, you may find value where you did not expect it.
Royalty describes himself as paralyzed as a creator by the tension among the claims at the beginning of his piece. But ironically, one of those claims almost contains the solution: “Making a good thing still requires making lots of decisions.”
The way forward is to make care legible without turning our work into a CAPTCHA for humanity: judgment, specificity, coherence, revision, accountability, humility, and respect for the person receiving the work. Those elements may involve tremendous labor, or good tools may reduce the labor required to practice them. The value lies in the care exercised and the good received, not in the number of hours visibly sacrificed.
I intend to keep the em dashes.
