A checklist of hollow writing, not a detector

What we call slop

AI slop is first a human way of writing, hollow and generic, that AI made free and massive. Our own measurements show it: several phrasings blamed on models already appear in articles published from 2017 to 2020 and in company emails from 2001.

So a rule does not ask “is this AI?” but “is this passage hollow, whoever wrote it?”. The goal is to help you avoid it and write like yourself.

Five axes, 26 rules and one readability hint

Each rule comes from published research on slop and from editing guides. Each flagged passage shows the rule, the reason and a way to fix it. There is no overall score: read each axis on its own, and zero passages does not mean a good text.

Under each rule, its measured frequency: the share of texts where it fires at least once, among human texts written before 2021 and among nine model families of October 2026. A rule that is frequent among humans is still a rule of hollow writing. A rule that is more frequent among models does not say a passage came from a model.

Measured on October 11, 2026, English texts: 180 human texts (articles from the Wikimedia Foundation and Diff Wikimedia published 2017 to 2020, Enron emails from 2001) and 171 texts from 9 model families of October 2026 on 19 prompts. The same measures exist for French, on the French page.

Substance

Does the text say something specific?

Paragraph without an anchor A01
Paragraph of 40 words or more with no proper name, date, place, example, quotation or lived experience.
Measured frequency: 16.7% of human texts written before 2021, 60.8% of texts from nine October 2026 models.
Packaging A02
Opening or closing that could top or tail any subject.
Measured frequency: 0.6% of human texts written before 2021, 15.2% of texts from nine October 2026 models.
Generic wording G01
Stock phrase from a published list (“it is important to note that”).
Measured frequency: 5.6% of human texts written before 2021, 36.8% of texts from nine October 2026 models.

Evidence and figures

Can what is claimed or counted be checked?

Orphan figure C01
Percentage or count with no population, period or scope.
Measured frequency: 13.3% of human texts written before 2021, 28.7% of texts from nine October 2026 models.
Phantom evidence C02
Evidence invoked without being given (“many studies”, “experts agree”).
Measured frequency: 6.1% of human texts written before 2021, 2.9% of texts from nine October 2026 models.
Decorative precision C03
Decimals or multipliers without a source (“10x”, “3.7 times”).
Measured frequency: 6.1% of human texts written before 2021, 4.7% of texts from nine October 2026 models.
Comparison without a baseline C04
Comparison without a baseline (“twice as many”, “+40%”).
Measured frequency: 0.6% of human texts written before 2021, 0% of texts from nine October 2026 models.
Reference to check against its source C06
Malformed reference: DOI, missing year, pages out of order, invalid ISBN.
Measured frequency: 0% of human texts written before 2021, 0% of texts from nine October 2026 models.
Reference to check V01
Strong claim, figure or quotation with no visible source. Secondary rule: it does not count a passage already flagged by C01 or C02.
Measured frequency: 66.1% of human texts written before 2021, 64.9% of texts from nine October 2026 models.

Templates

Is the form the one everybody recognizes?

Negation pivot T01
“It’s not X, it’s Y” pivot: hollow when nobody had said X, when Y stays vague or when it comes back by reflex. Legitimate when it corrects a real confusion with a precise answer. On English text the tool counts it under G01, with the Barometer’s phrasings, so T01 itself stays at zero.
Measured frequency: 0% of human texts written before 2021, 0% of texts from nine October 2026 models.
Announcement T02
Announcement that holds back the information (“Here’s why”, “The good news:”). On English text the tool counts it under G01, so T02 itself stays at zero.
Measured frequency: 0% of human texts written before 2021, 0% of texts from nine October 2026 models.
Closing question to the reader T03
Generic closing question to the reader (“What about you?”).
Measured frequency: 0% of human texts written before 2021, 4.6% of texts from nine October 2026 models.
Template formatting T04
Emoji lists, hashtag blocks, cascades of one-line paragraphs.
Measured frequency: 3.3% of human texts written before 2021, 22.2% of texts from nine October 2026 models.
Trailing participle phrase T05
Sentence ending on a commenting participle phrase (“…, highlighting the importance of”).
Measured frequency: 5% of human texts written before 2021, 1.2% of texts from nine October 2026 models.
Inflated importance T06
Importance proclaimed rather than shown (“pivotal role”, “game changer”).
Measured frequency: 0% of human texts written before 2021, 1.2% of texts from nine October 2026 models.
False range T07
False range (“from strategy to execution”).
Measured frequency: 0% of human texts written before 2021, 0.6% of texts from nine October 2026 models.
Heavy formatting F01
Bold, headings and “Term:” bullets beyond what the length of the text needs.
Measured frequency: 7.2% of human texts written before 2021, 56.1% of texts from nine October 2026 models.
Assistant leftover X01
Leftover from a chat with an assistant, brackets to fill in, technical trace.
Measured frequency: 0.6% of human texts written before 2021, 31.6% of texts from nine October 2026 models.
Announced conclusion S01
Conclusion announced by a formulaic transition (“In conclusion”).
Measured frequency: 0.6% of human texts written before 2021, 0% of texts from nine October 2026 models.
Repeated opening S02
Three sentences opening with the same words.
Measured frequency: 22.8% of human texts written before 2021, 14.6% of texts from nine October 2026 models.
Close vocabulary R01
Two sentences sharing at least 68% of their informative vocabulary.
Measured frequency: 39.4% of human texts written before 2021, 24.6% of texts from nine October 2026 models.

Voice

Is there someone behind the text?

No situated point of view P01
No situated point of view in a post or article: no experience, no stance.
Measured frequency: 14.2% of human texts written before 2021, 13.1% of texts from nine October 2026 models.
Surface neutrality P02
Surface neutrality (“strike a balance”) with no stance taken.
Measured frequency: 0% of human texts written before 2021, 0.6% of texts from nine October 2026 models.

Usefulness

Does the text move the work forward?

Request not found W01
Email with no question, decision or expected action.
Measured frequency: 28.3% of human texts written before 2021, 5.6% of texts from nine October 2026 models.
Action without owner or deadline W02
Action with no owner or deadline.
Measured frequency: 8.3% of human texts written before 2021, 0% of texts from nine October 2026 models.
Summary without conclusion W03
Report that summarizes without recommending or deciding.
Measured frequency: not measured on human texts of this format, 22.2% of texts from nine October 2026 models.

Readability

A reading hint, not a sign of hollow writing.

Long sentence D01
Sentence of 42 words or more. A readability hint, not hollow writing.
Measured frequency: 61.1% of human texts written before 2021, 7.6% of texts from nine October 2026 models.

How to read the frequencies

The human texts are articles and, in English, emails. The generated texts cover eight genres: posts, blog articles, newsletters, marketing pages, emails, explanations, reports and speeches. The two sets do not mix the same genres: a frequency describes a corpus, not your text. Format rules (W01 to W03, P01, T03, F01) are measured on that format only.

We publish no accuracy figure. The rules are not signs of AI to validate against a hidden truth: they are standards of hollow writing that research already describes. The ongoing check is your feedback: the form under each result reports an unfair alert or a missed passage.

What changed since 0.6

V01 (reference to check) becomes secondary: it no longer counts a passage already flagged as an orphan figure or phantom evidence. A missing source is a trait of much human writing, not a trait of slop on its own.

D01 (long sentence) leaves the axes: it is a readability hint. Long sentences are far more frequent in our human texts than in generated ones.

T05 (trailing participle phrase) describes a habit research found in 2024 models. October 2026 models use it less than our human texts do: the rule stays because the phrasing comments instead of informing, and that is what makes it hollow.

Seven editing contexts

The selected format adapts advice and examples: general text, social post, article, email, report, marketing page or documentation. Format rules (email, report, post) fire on that format only.

English phrasings

For English text, G01 also flags the 15 phrasings the AI Slop Barometer counts: 13 stock phrases (“delve into”, “a rich tapestry”…) and two constructions kept by a preregistered test on 4 October 2026, the “it’s not X, it’s Y” pivot and the “Here’s why” announcement. Human writers use them too. They count because they replace a specific idea with a stock phrase.

The tool’s English method: 0.8.0-en-experimental. The Barometer still measures with version 0.6.0 (0.6.0-en-experimental in English) in October and November so as not to break its series. The switch to 0.8 will be announced in an issue.

Stability tests

Reconstructed cases pin the behavior of each rule, in French and in English, and every change must pass them. They check that the rules stay stable. They do not measure real-world accuracy.

Research basis

Each rule is justified by what makes the passage hollow: a phrase that would fit any subject, a figure with no scope, evidence invoked without being given, importance proclaimed rather than shown. Research helps find these forms: excess words and phrasings in generated text, syntactic templates, participial clauses and nominalizations, vague attributions, invented references. A model using a word more often than people do is never enough to make it a rule. These studies mostly cover English and models from 2023 to 2025: that is why we measure each rule ourselves on French and English texts from October 2026. X01, T07 and F01 come from reviewing generated texts and are measured the same way.

A study motivates a rule. It does not validate its threshold, and our frequencies do not say a passage came from a model.

Sources for this review

Research review checked on 25 September 2026

A study tested ChatGPT text detectors on French, with limitations outside the tested domains and under attacks. A major multilingual benchmark does not include French. Neither result measures the precision or recall of this Observatory’s rules.

Pew found signs of AI writing or editing on 10% of English pages in its July 2026 sample. The 35% figure applies only to pages with publication dates after November 2022. This dated subset is not random. These figures do not measure slop.

A US survey reports time spent reworking material. A controlled experiment also finds gains from AI assistance on one defined task. Output quality, rework time and task performance are different measures. None of these results supports a new automatic alert.