Writing reads as machine-generated when it is fluent and evidentially empty: correct sentences carrying no specific detail, no named source, and no claim the writer could be wrong about. The vocabulary tells are the cheapest layer and the easiest to remove. Removing them changes nothing while the paragraph underneath still commits to nothing.
Flow diagram of 5 steps: Delete every paragraph you can delete, Make every general claim name an instance, Put one falsifiable sentence in every section, Break the symmetry deliberately, Rewrite the most hedged sentence as what you actually think.
The word arrived for something specific: text produced in volume because producing it had become almost free, published because publishing it was free too, and addressed to nobody in particular. That is a description of how the writing came to exist and what it does to a reader, not a description of its word choice. A paragraph is slop when deleting it would cost the reader nothing, and that test does not care which tool typed it.
It follows that humans have been producing the stuff for years. Most of the content marketing written between about 2012 and 2022 was slop by this definition — commissioned to occupy a keyword, padded to a word count somebody had decided was necessary, and read by search engines more often than by people. Models did not invent the failure. They removed the last practical constraint on it, which was that somebody had to be paid to sit and type.
Understanding why a model produces it by default matters, because it tells you where the fix has to go. A model asked for eight hundred words on a subject will produce eight hundred words whether or not eight hundred words of substance exist. It has no mechanism for reporting that the honest version of the answer is two hundred words long, and no discomfort about filling the remainder with material that is true, grammatical and useless. The padding is not a bug in the output; it is the output doing exactly what the request specified.
A machine-written paragraph is usually cleaner than a human first draft — the grammar holds, the transitions connect, nothing snags as you read. Absence of friction gets mistaken for quality, and the emptiness only becomes visible when you try to say what the paragraph claimed. If you cannot summarise it without reusing its own phrasing, there was nothing in it to summarise. That test is worth running on your own drafts, because the same smoothness that hides emptiness from a reader hides it from the writer.
The tells split into two layers that get discussed as though they were one. The surface layer is lexical and rhythmic: particular words, particular sentence shapes, a certain evenness. It is the layer everybody can name, and it is almost entirely cosmetic. The structural layer is about what the text contains and risks, and it is the layer that decides whether the writing is worth anyone reading.
Ranking them this way is not an argument for ignoring the surface. A reader who hits three signature phrases in the first paragraph stops reading, and they are then not available to be persuaded by whatever substance came later. But a draft that has been scrubbed of every flagged word and still contains no specific detail has been made harder to diagnose rather than better.
Two layers of tell, and what each one costs
| Tell | What it looks like | How much it matters |
|---|---|---|
| Signature vocabulary | delve, tapestry, testament, navigate the landscape, in an increasingly complex world | Low — a find-and-replace removes it and improves nothing underneath |
| The rule of three, everywhere | Every list runs to exactly three items; every sentence carries three parallel clauses | Low to medium — noticeable rhythm, symptomatic rather than harmful |
| Symmetrical paragraphs | Every paragraph roughly the same length, every section the same internal shape | Medium — signals that nothing was cut, because nothing was weighed |
| The false-contrast construction | It is not just X, it is Y — used to manufacture the feeling of an insight | Medium — a rhetorical shape standing in for an actual distinction |
| No specific detail | No name, number, date, place, price or source anywhere in the piece | High — a reader finishes knowing nothing they did not know before |
| Unfalsifiable claims | Every sentence is one no reader could disagree with or check | High — nothing is being asserted, so nothing can be learned |
| Nothing risked | No position that could embarrass the writer if it turned out to be wrong | High — the reason the piece feels weightless |
| The conclusion restates the introduction | The last section summarises the first, adding no step | High — evidence the argument never travelled anywhere |
Ask somebody why a piece felt like slop and they will usually point at a word, because the word is the part they can quote. What actually produced the reaction is the accumulated absence: three thousand words in which nobody is named, nothing is dated, no figure appears, no source is cited and no claim is made that could be shown wrong. Fixing the quotable part while leaving the absence intact produces writing that reads as competent, forgettable and faintly hollow, which is where most scrubbed drafts end up.
The most popular remedy is a banned-word list. It fails on two counts. The first is that the words are ordinary English with legitimate uses, and a rule against them makes writers work around perfectly good sentences to avoid a false accusation. The second is that the list is a moving target: as soon as enough people avoid a term, its signalling value collapses and the tell relocates to whatever the avoiders reached for instead. You cannot win a game whose rules are defined by what everybody else is currently doing.
There is a fairness problem underneath as well. The features people read as machine-written — limited vocabulary range, predictable phrasing, careful mid-formal register — are also the features of competent writing by somebody working in their second language, or by anybody trained in institutional prose. Accusations of sounding like a machine land hardest on writers who had the least to do with one. That is a reason to treat the surface layer as weak evidence about origin, and to judge a piece on what it contains instead.
The second popular remedy is to hand the draft back to the model and ask it to sound more human. What comes back has contractions, one rhetorical question, a deliberate sentence fragment for emphasis, and possibly a second-person aside. It is a register change with no change in substance, and the register is by now as recognisable as the one it replaced. The third remedy is worse: adding a personal anecdote to prove a person was involved. An invented anecdote is not a fix for empty writing. It is empty writing with a false statement in it.
Automated detection tools return a confident percentage that is not backed by a reliable measurement. They produce false positives on human writing in plain registers and false negatives on machine text that has been lightly edited, and the score moves when you change nothing that matters. Treating that number as evidence — about your own draft or somebody else’s — outsources a judgement it cannot make. Read the piece and ask whether it contains anything specific, checkable and at risk of being wrong. That question is answerable by a person in about a minute.
This pass takes roughly twenty minutes on a thousand words and it is subtractive for the first half. Expect the draft to get shorter before it gets better; a piece that survives losing a third of itself was carrying a third of itself for no reason.
Take them one at a time, remove the paragraph, and read the two paragraphs either side of the gap. If the piece still makes sense and nothing is missing, the paragraph was filler and it stays deleted. Do this before any line editing, because polishing a paragraph you are about to cut is the most common way a revision pass gets spent achieving nothing.
Go through the remaining claims and, for each, supply the specific case behind it — the name, the number, the date, the actual example. Where you cannot supply one, you have found a sentence you do not have grounds for. Cut it rather than dressing it up. This step is where most of the improvement comes from, and it is the step people skip because it is the only laborious one.
Each section needs at least one statement a reader could investigate and disprove. Not an opinion nobody could check, and not a truism nobody would contest — a claim about how something actually works, which is wrong if the world is arranged differently. If a section has none, it is describing a topic rather than saying anything about it.
Vary paragraph length so the shape of the page reflects the weight of the material. A two-sentence paragraph next to a nine-sentence one tells a reader which point carried more. Even paragraphing signals that everything was treated as equally important, which is almost never true and is usually the residue of never having decided.
Find the sentence with the most qualifiers stacked on it — the one with a can sometimes, a may in certain cases and a for many people all in the same clause. Either write the claim you believe, plainly, or cut the sentence. Hedging is where a position goes to avoid being held, and a paragraph of it reads exactly like a machine avoiding commitment.
Before you start, write down in one sentence what the piece claims, without using any of its own phrasing. If you cannot, the draft has no claim and no amount of editing will install one — go back and decide what you think. Run the same test after the pass. If the sentence is now sharper, or if it changed, the revision did structural work. If it is identical and the draft is merely shorter, you tidied rather than revised.
A submission description is short, which concentrates the problem rather than avoiding it. The characteristic failure is a description nobody could object to: a programme that would inspire people to be kinder in their everyday lives, bringing communities together through uplifting stories. Every word of that is agreeable and none of it is information. Nobody reading it can picture an episode, and nobody voting on it is choosing between it and anything else, because it describes the whole category rather than one idea inside it.
The repair is the same as for any other draft, compressed. Name the format and the length. Say who appears and what happens in the first ten minutes. Say what physically has to exist for it to be made — a location, a contributor who has agreed, a piece of footage somebody already has. Say who specifically would watch it, in terms narrower than people who like positive news. A description carrying four concrete facts is more persuasive than one carrying four hundred agreeable words, and it also gives the automated review something to assess, since a criterion like feasibility has nothing to work from when no requirement has been stated.
The form offers an optional AI Description Helper that rewrites what you have typed into something more broadcast-ready. Used as a first draft to edit, it is fine. Accepted unchanged, it tends to smooth exactly the specific and slightly awkward details that made the idea worth submitting, and hands back a description of the general category. That is the mechanism producing slop, running on your idea, with your name on the result.
Two limits worth stating plainly. Kind Channel is new, the community is small and nothing has aired, so nothing here is derived from measuring which descriptions perform well on this platform — there is not yet enough to measure, and any claim otherwise would be invented. And the automated notes decide nothing on their own: community votes determine what gets made, cast by people who mostly read the description and nothing else. Writing to satisfy the review rather than to interest those readers is optimising for the wrong audience.
No. Slop describes writing that is fluent and empty — no specific detail, no checkable claim, nothing the writer could be wrong about — and text acquires those properties from how it was made rather than from which tool typed it. A model used to pressure-test structure or find the weak paragraph in your own draft leaves no trace of that kind. A model asked to generate eight hundred words on a topic you have not thought about produces slop reliably, because there was nothing to say and the word count still had to be met.
The commonly cited ones include delve, tapestry, testament, realm, navigate the landscape, and openings of the in an increasingly complex world variety. Removing them is worth ten minutes because readers do react to them. It is also cosmetic: a draft scrubbed of every flagged term still reads as machine-written if it contains no names, numbers, dates or sources and makes no claim a reader could dispute. Treat the vocabulary as the last pass, not the fix.
Not reliably enough to act on. They return confident percentages while producing false positives on plain human writing and false negatives on machine text that has been lightly edited, and the score shifts on changes that alter nothing meaningful. The penalty falls hardest on people writing in a second language, whose careful, limited-vocabulary prose shares surface features with generated text. Read the piece instead and ask whether it contains anything specific, checkable and capable of being wrong.
Put four concrete facts in it and cut anything nobody could disagree with. For a programme idea that means the format and rough length, who appears and what happens early on, what would physically have to exist to make it, and who specifically would watch — narrower than people who like uplifting stories. Sentences everybody agrees with feel safe to write and carry no information, so they are the first thing to remove when space is tight.