What "AI Slop" Means, and How Not to Make It

Slop is usually defined by vibe or by provenance, and both definitions fail: the vibe is unfalsifiable and the provenance is irrelevant. Defined by properties it becomes useful, because a property can be checked on a page in front of you and each one has a specific fix.

A definition worth arguing with

Slop is content that is optimised to exist rather than to be read: produced in quantities that preclude judgement, asserting nothing that could be wrong, and sourced in ways that do not survive being checked.

Five properties follow from that, and they are separable — material can have two and not the others, which is why the fixes are different. Nothing in the definition mentions how the text was produced. That omission is deliberate and is defended below.

One: volume without selection

The first property is quantitative and it is upstream of everything else. If two hundred pieces were published, nobody read two hundred pieces carefully enough to decide which ones deserved to exist. The rejection rate is the editorial function, and above a certain output it is arithmetically zero.

This is why volume is a property of the work rather than of the publisher’s intentions. A reader cannot see your standards; they can see that a site published nine pieces on a Tuesday, and they can correctly infer that none of the nine was chosen over the others.

The fix. Publish what survives a decision. Keep a record of what was rejected and why, because a rejection log is the only evidence that a selection process exists — it is the same artefact that makes an editorial gate auditable. If the log is empty, the process is not slow, it is absent. The cadence arithmetic is on the volume page.

Two: confident vagueness

The second property is the most recognisable and the easiest to test. Slop asserts continuously and commits to nothing. Sentences have the grammatical form of claims and the informational content of weather.

SLOP
Choosing the right model for your use case is essential. Different
models offer different capabilities, and organisations should carefully
evaluate their requirements before making a decision. Cost, performance
and accuracy are all important considerations.

NOT SLOP
Pick the cheapest model that passes your eval set, then re-run the
comparison quarterly. Almost every team over-buys: the expensive model
wins on aggregate benchmarks and loses on your task about as often as
it wins, and you will not know which until you have twenty scored
examples of your own.

The difference is commitment. The second version can be wrong — somebody can say “we did that and the cheap model failed on long inputs” — and being wrong is what makes it worth reading. The first cannot be wrong about anything, which is why it costs nothing to write and returns nothing to read.

The fix. Every paragraph either states a condition, a number, a name, or a consequence. Where you would write “it depends”, write what it depends on. Where you would write “significantly”, write the figure or admit you do not have one. Where you cannot do either, delete the paragraph — it was not carrying anything.

Three: symmetry that flattens the argument

The third property is structural and subtler than the other two. Slop is regular: sections of even length, lists of three, benefits balanced by an equal weight of challenges, every point given the same amount of room.

Real arguments are not shaped like that. One point usually carries most of the weight and deserves four paragraphs; two others are worth a sentence each; a fourth was cut. Uniformity is evidence that the material was arranged rather than thought about, because thinking produces unequal importance and arranging does not.

The specific damage is that symmetry destroys the reader’s ability to prioritise. A page with six equally weighted considerations has told them nothing about which one to act on, and choosing was the job they came to have done. False balance is the same failure applied to evidence: two positions presented as equal when the evidence is not, and a middle-of-the-road conclusion that transfers the decision back.

The fix. Let the structure be lopsided. Give the load- bearing point disproportionate space and let two others be short. Say which side you think is right and under what conditions the other one wins. The mechanical edits for both are on the editing page, as tells one, four and nine.

Four: sourcing that dissolves when checked

The fourth property is the one with actual victims. Slop cites, and the citations evaporate.

• The unnamed study. “Research shows”, “studies suggest”, “experts agree”. Nothing is named, so nothing can be checked, and the authority is borrowed from a body of work that may not exist.
• The statistic with no origin. “Around 40 per cent of companies report” — report to whom, measured how, in what year, on what sample. A figure without those four things is decoration.
• The citation that does not say it. A real paper, correctly titled, that on reading turns out to make a much weaker or entirely different claim. This is the most common failure and the hardest to catch, because the first two checks pass.
• The circular reference. A link to an article that cites another article that cites the first. Common enough on frequently-repeated numbers that it is worth following every chain to its end at least once.
• The fabricated reference. Plausible author, plausible journal, plausible title, no such paper — assembled because every component is individually likely.

The fix. Name the publisher and the year in the sentence, not in a link. Open every source and confirm it says what you claim. Where no reliable public figure exists, write that sentence instead — “no public dataset measures this; it would take X to measure it” is more useful than a number, and it is a sentence slop never contains. The workflow that enforces this is span-verified extraction.

Five: nothing at risk

The fifth property subsumes some of the others and is the best single test. Slop takes no position that could cost its author anything.

Nobody could be annoyed by it. No competitor is named. No approach is called a mistake. No prediction is made that could be checked next year. No experience is described in enough detail to be contradicted by someone who was there. The text is engineered to be publishable rather than to be right, and the two goals are in tension exactly where writing gets valuable.

This is also why the property survives good prose. Text can be beautifully written, well structured, correctly sourced, and still be slop, because none of those qualities requires committing to anything. Skilled hedging is a craft, and it is the craft that produces the highest-quality slop there is.

The fix. Put something at risk in every piece: a recommendation, a criticism, a prediction with a date, an admission of what you got wrong. One is enough. If nothing in a draft could embarrass you in a year, nothing in it was worth writing this year.

Why provenance is the wrong test

The common definition — slop is machine-made content — is appealing and does not work, for three reasons.

It is not extensionally correct. Content farms produced all five properties at scale for two decades using freelancers paid by the word. The incentive structure produced the properties; the technology only lowered the price. A definition that excludes the human version is describing a tool rather than a failure.

It is not decidable. Detection of machine authorship from text is a statistical judgement about style, and style is exactly what a careful writer edits. Any definition whose central term cannot be established from the artefact is a definition you cannot apply to the thing in front of you.

And it produces the wrong instruction. “Do not use a model” leaves you with unselected, vague, symmetrical, thinly sourced, risk-free writing typed by hand. The five properties above give you five things to change, each of which improves the work regardless of how it was produced.

One thing does change with provenance, and it belongs in a different argument: whether the reader was misled about who did the work. That is the disclosure question, it is about honesty rather than quality, and it is handled on its own page.

What slop costs, and who pays

• Readers pay in verification. When a proportion of what you read is unsourced, the cost of establishing anything rises for everyone, including for the people producing good work.
• Search pays, and passes it on. Ranking systems weight signals that were proxies for effort when effort was expensive. Those proxies got cheap, so the signals got noisier, so ranking gets worse for everybody.
• Reviewers pay. Any venue with an open submission queue — magazines, journals, bug bounties, job applications — absorbs volume that costs almost nothing to send and real time to read. Several have closed submissions over exactly this.
• Future models pay. Text scraped for training is increasingly text produced by models, which narrows the distribution that made the next generation possible.
• The publisher pays last and most. A reader who discovers one fabricated citation discounts everything else on the site, including the parts that were checked. Trust is lost per site and earned per page.

The test, applied here

One question, asked of a draft before it goes out: which sentence in this could be wrong?

If the answer is none, it is slop — whatever produced it, however well it reads, however carefully it is sourced. Nothing that cannot be wrong can be useful, because a reader gains nothing from a text that would have been equally true if the world were different.

The question applies to this library too, and it would be dishonest to write nine paragraphs of this and then exempt the page it is printed on. Fifteen hundred pages is a volume that makes property one a real risk, and the honest position is not that the risk was avoided by good intentions but that it has to be checked page by page against the same five properties. The falsifiable claims on this page are the ones to hold it to: that the five properties are separable, that provenance is not decidable from text, that content farms produced the same properties without machines, and that a definition based on provenance yields no useful instruction. Each of those can be argued with. That is the point of writing them down.

The general test for any page, including this one: find the sentence somebody could disagree with. If you cannot find one in the first screen, close the tab. The writer already told you what they had.