DropSlop research · Method

How we did it.

One detector reads every text. What it flagged in each source before ChatGPT (released ) counts as false alarms, what it misses on known AI writing is put back, and the result is the share wholly written by AI.

− − =

False alarms on real writing.

What we collected.

We collected published writing from public web archives, search results, public registers and official feeds, and scored unique published texts, plus AI-written examples made to measure catch rates. Four sources show how the counts fall out; every source’s counts are in its own file.

Only whole texts, cleaned the same way.

The detector reads prose, so everything around it is cut first, the same way for every source.

  • The -word ruleA text is scored only if it has words or more after cleaning. Shorter texts give a detector too little to read, so they are counted and set aside, never scored. English only.
  • Each text onceA text that appears twice, in one source or across sources, is scored once. Copies of the same page in several crawls are dropped, oldest kept.
  • Boilerplate strippedLines repeated at the top or bottom of many pages of one site, or across many sites (menus, sign-up boxes, cookie notices), are removed before scoring. That changed new web pages, top-ten pages and AI-cited pages. Public comments carry no site template; changed: .

The detector, and its benchmark.

DropSlop’s own detector, version (file fingerprint ), scored every text in one run. A text counts as flagged at a score of or more. On its clean benchmark it raised no false alarms; on real writing it does (above), which is why every source is corrected against its own rate.

  • Core benchmarkOf held-out texts, it flagged of the written by people and caught of the AI-written ones.
  • Hard human textsOf human texts chosen because detectors often get them wrong, it flagged .
  • Short paragraphsOn short paragraphs its ranking accuracy (AUROC, where 1 is perfect) was .
  • Mixed textsOf texts mixing human and AI passages, it caught only . This is its weak spot.

Catch rates, measured on known AI writing.

We had texts written for this study by widely used AI model families, across kinds of writing, from web pages to public comments. Each source’s catch rate comes from the examples of its own kind. They are unedited text generated with five 2026 models; text from older models, or AI text a person edited, may be caught at a different rate.

    Ranges that respect where texts come from.

    Pages of one site, comments on one rule and releases of one company tend to rise and fall together. So every range resamples those whole groups, not single texts.

    • Clustered bootstrapEach 95% range comes from resamples of whole groups, by source as listed below. The baseline and the catch rate are redrawn in the same resamples, so their uncertainty is in the range.
    • What each source resamples
    • The bar for a groupA group (a page type, a topic, a year) is published only with or more texts in each era, or is shown as a stated limit with its wider range. Headline groups were fixed before scoring (studies/FINAL_CELLS.md); others are exploratory and marked on each page.
    • Each source’s own baselineFalse alarms come from the same source before ChatGPT, with three exceptions. Job postings: today’s postings come from the live hiring boards and the earlier ones from archived copies of those boards, so the two eras were collected differently. Apps: an older app’s description is read as it is today, and many were updated since, so the app figure is a floor. AI-cited pages and top-ten search results: old pages are read against archived 2022 copies of the same pages, since live pages may have been rewritten.

    What these numbers cannot tell you.

    1. Commercial pages look machine-made to it.

      Even archived copies of top-ranking commercial and SEO pages from before ChatGPT were flagged at (AI-cited) and (top ten), and copies from 2019 or earlier at . That is why each source is corrected against its own baseline; a share without it would be badly wrong.

    2. Only wholly AI-written text counts.

      Texts mixing human and AI writing are caught only of the time, so every share here is about texts written wholly by AI. The share with any AI help is higher.

    3. Live pages may have changed.

      A page read today may have been rewritten since it was first published. Archived copies fix this for the baselines of the search stories; the pages published since ChatGPT are read as they are now.

    4. It can be wrong about any single text.

      The findings are about shares of many texts. No page, comment or release here should be judged AI-written on this detector’s word alone.

    The data behind every number.

    Aggregate figures only: no texts, addresses or names. Every number on these pages links to its file and key.

    Run the same detector on your text.

    Paste a draft to see which phrases read machine-made, or read a whole site page by page.