How plopslop works

plopslop lists news and essays from sources whose recent output passes automated human-writing checks. No person reads articles before they appear here. No person approves sources one by one. The system runs on statistics, and this page explains them in plain language.

What the checks are

Every article of 300 words or more is scored by an ensemble of open source detectors. The primary detectors are zero-shot methods published at ICLR 2024 and ICML 2024. They measure statistical properties of text against live language models instead of pattern-matching against any particular product, which helps them generalize to new model releases. Each detector votes at a calibrated threshold, and the votes are combined by majority. Articles under 300 words are not scored, because every detector known to us degrades badly on short text.

Why decisions happen at the source level

No single-article verdict is reliable, so plopslop never publishes one. Decisions come from rolling statistics over each source's last 50 scored articles, with a 95 percent confidence interval around the rate. A genuine publication produces a low, steady rate of false positives. A publication that switches to generated content jumps to much higher rates. The gap between those two states is large, and the interval math is what separates them.

The number on each card

The human-writing confidence figure is a 0 to 100 integer. It comes from fusing the detector outputs and mapping the result through a calibration table built on labeled reference texts, so the displayed number tracks measured accuracy rather than raw model output. It appears only on published articles, and short articles that were never scored carry no number.

What the scores are and are not

The scores are automated statistical estimates. They are not assertions of fact. plopslop makes no claim of authorship about any specific text or any specific publisher. Absence from plopslop implies nothing about any publication. Detection of machine-written text is an evolving field with real error in both directions, and the methodology here is designed around that error rather than in denial of it.

Privacy

Reading preferences live in your browser and never reach a server. plopslop keeps no accounts and no reading history, and serves no third-party trackers. Request counting is self hosted.

For publishers

plopslop stores an article title and a snippet of 200 characters or fewer, links every card straight to the publisher, and honors robots.txt on every fetch. Any source that asks to be removed is removed. Write to hello@plopslop.com.