What it measures
Three separate indicators, in signals per 1,000 words. Hollow phrasing: generic wording, stock phrases, template structures and hollow turns of phrase widespread in 2026. Sourcing: claims or figures without a visible source. Readability: repetitions and long sentences.
Each rate comes with its 95% interval. An overall index exists in the data, for verification. It is never given alone or in a headline, because it adds up signals that do not mean the same thing.
These three indicators apply to the October and November 2026 issues. From the December 2026 issue at the earliest, with method 0.7, “sourcing” becomes “evidence and figures”, which also checks hollow or unverifiable figures, and a “substance” indicator is added: does the text say something specific? This change will be announced as a series break.
What it does not measure
It does not say what share of texts was generated by AI, or whether a text is true. A signal is a passage an editor should look at again. A human author can produce one, and a generated text can carry none.
The 2026 turns of phrase were chosen on 160 texts generated by four models from a single provider (Anthropic) and on human texts published before November 2022. They are writing flaws AI made common, not proof of origin: human authors write them too. The method is experimental, more so in English.
The sample
Public feeds that publish full text: blogs, open platforms where anyone can post, and a few media outlets. Major news sites, which only publish excerpts, are not included, and the Barometer does not describe them.
A text is included if it has at least 150 words, a date, and is in the edition’s language. Collection respects robots.txt and text and data mining reservations, with at most one request per second per site. Only numbers and a fingerprint are kept: no text, title, author or article address.
Publication rules
An issue is published only if it has at least 300 texts, 100,000 words and 8 sources, if no source exceeds 20% of the words, if all its sources have been validated, if the month is closed and if it is calculated with a single rule version. Otherwise, it states what is missing and gives no rate.
No rate is published per source. The Barometer describes a whole, not an outlet or an author.
Margins of error
The 95% interval comes from resampling the texts (bootstrap, 1,000 draws, fixed seed). Two rates with overlapping intervals are not enough to conclude there is a difference. The English and French editions cannot be compared point by point.
When the method changes
Each measurement carries the rule version that produced it. A month measured with one version is not compared with a month measured with another: the issue announces a series break and calculates no change. When possible, both versions measure the same texts for a while, and the gap between them is published.
Schedule
One issue per month and per language, built on the 1st of the following month and published in the first week. First issue expected: October 2026, published in early November, if the rules above are met.
How to cite
AI Slop Observatory (YIRIWA, LLC), AI Slop Barometer, [month], [English or French] edition, rules [version]. Please give the rate with its interval.
Publisher and interests
YIRIWA, LLC publishes the free Observatory and also sells writing products and services (Kaneme, team workshops). That is why the rules are published before the figures and no rate is published per source.
Report an error
Write to info@kaneme.com naming the issue. A correction states the issue, the date and what changes, under the public corrections protocol. A corrected issue keeps a record of its previous version.