Topics
Slop
I'm Spwashi. Slop is the cheap residue of a generator: fluent, cheerful, and interchangeable. The word was Merriam-Webster's word of the year for 2025. By 2026 it had become the ordinary accusation for anything that feels unread.
The feeling
People are tired of it, and the platforms have started to say so. On 17 August 2026 the New York Times reported that Spotify had removed 75 million spammy tracks, and that LinkedIn had named AI slop a top priority. LinkedIn added a "seems like AI slop" control on 30 July. By 21 August, The Verge reported more than a million uses, and LinkedIn said views of what it classifies as slop were down about 40 percent. A Pangram count around then had flagged 41 percent of long LinkedIn posts as fully generated.
A Wired account of an Imperial College, Stanford, and Internet Archive preprint said about 35 percent of new websites from 2022 to 2025 were generated or assisted, and that those pages scored about twice as positive in tone as the rest. The cheerfulness is part of the complaint. A Spectator essay on 22 September 2026 called the result homogenization: the same safe average on a menu, a flyer, or a paragraph. The writer would rather see a crooked handwritten sign.
The accusation has pulled away from the evidence. A Villanova study, covered by Metro on 9 August 2026, found readers often cannot tell the stories apart, and they rate the machine story higher when they are told a person wrote it. An analysis of 25 million Hacker News and Reddit comments, posted 23 September 2026, found that "slop" had risen from about 13 percent of pejorative accusations in 2023 to about 94 percent by 2026, and that the tone had moved from mockery toward gatekeeping. The surprise in that paper: the prose features that statistically separate machine text from human text do not predict which human comments get accused.
Tells in the writing
The tell is not a grammar mistake. It is a paragraph you can finish without being able to say what you learned. These are the patterns the essays and detectors keep naming. They describe a style. They do not reliably name an author.
Hollow fluency
Sentences are smooth and the claim is missing. If you cannot restate the point, the paragraph did not carry one.
Glue
"In today's fast-paced world." "It's important to note." "A testament to." Connective tissue with nothing attached.
Unsourced authority
"Studies show." "Many experts." A percentage with no name, date, or link.
Balance with no case
Every point is paired with its opposite. No particular day, place, or decision.
The same shape
Lists of three. Paragraphs of equal length. A closing sentence that repeats the opening. The Spectator also notes a small word set, including "delve," "adept," and "meticulous," that human writers have started to copy.
The picture
Block letters on a dark field, a fake handwritten script tilting up to the right, and every corner filled.
Patterns to consider
Two Claude engineering notes are useful here as patterns, not as a product to adopt. How we made claude.ai 3x faster in two weeks (23 September 2026) and Reducing cost and improving performance (8 September 2026, updated after the 22 September model) describe how a team spent less waiting and less money without asking the writer to be "more thorough." Thoroughness as a ritual is one way slop gets made. These are the moves I would try on another project.
Measure the journey, not the server
They timed four journeys that were about 95 percent of use, from the person's action to the painted result, and they split client time from server time. A fresh page became typeable in 0.55 seconds instead of 3.1, at the 75th percentile. Another project can pick the few paths people actually walk and time those, end to end.
A count that can only fall
Wall-clock time is what a person feels, and it is too noisy to gate a change. They paired it with deterministic counts: CPU instructions, React commits, style recalculations, DOM mutations. If a lab number did not move the felt time, they threw the number out. Once a count proved itself, a test refused to let it rise. That ratchet is the receipt.
A first surface copied from the real one
They put a static composer in the HTML so a person can type before the application is ready. The static markup is rendered from the real component, and a test fails if the two drift by a pixel. The pattern for another project: the first thing a person can use should be the document, generated from the same source as the live view, not a sketch that rots.
Name the region that jumped
A global layout-shift score stayed "good" while the sidebar rearranged, because each jump was about 0.008. They mapped each shift to a named region and a phase, then failed a test on any shift in a named region. A single site-wide score can hide the one place a person notices.
A convenient selector can tax every change
One :root:has() rule was adding 24 milliseconds to every DOM change. Em dashes and curly quotes were forcing whole strings onto a slower text path and freezing highlighting for about a second. Before adding a root-level rule, count the style work. A character you like in prose can be a cost in a hot loop.
Delete the ritual that once patched a weaker writer
"Verify twice." "Be maximally thorough." A mandatory scratchpad. On their support example, those lines made a newer model do extra searches, follow contradictory rules too literally, and sometimes write the action inside its reasoning without doing it. Removing them cut cost about another 9 percent and raised accuracy slightly. Hollow fluency is the prose version of the same ritual. A project can search its own instructions for glue that asks for volume instead of a checkable result.
Effort is a dial
More deliberation helps only while there is still evidence to find. Past that, it adds cost and can make the answer worse. A stronger writer at a lower setting beat a weaker one working hard, on their coding bench, at about a third of the cost. Their later cost pass reported roughly 67 percent lower cost on LegalBench, 73 percent on a retail bench, 72 percent on OfficeQA, and 24 percent on SWE-bench Verified, with scores essentially flat. For another project: decide the evidence you still need, then stop. Do not treat maximum effort as the careful choice.
Refuse a small win that costs the system
They declined a 900-line change because 2 milliseconds per send was not worth keeping the plugin. Someone still has to say when a measured gain is the wrong kind of gain. That refusal is part of the pattern, not a failure of the measurement.