← back

How it works

plopslop is a game and a feed. This page says what is behind each of them, with the real numbers. Nothing here is a summary of something more precise kept somewhere else.

why a game

Two things are true at once. Readers without practice cannot tell AI writing from a person's: in blind tests of news text published in 2025, they did no better than a coin flip. And no detector can be trusted on its own: a peer-reviewed comparison of four commercial detectors published in June 2026 found every one of them weakest on edited and humanised text, and independent benchmarks from 2025 and 2026 show detectors losing 30 to 70 points once a text has been run through a humanizer.

The third thing is the reason the game exists. People who practise get very good. In a study published in 2025, five readers who used these tools daily missed 1 article in 300 by majority vote, better than most detectors, and they stayed that good on paraphrased text. Practice is the variable, and a game with an instant answer is a practice loop. The calls it collects are how plopslop learns where its own software is wrong.

The Slop Game

You get an article. You call it slop or top. We tell you what our model called it, at once, and you score for a correct call and nothing for a wrong one. Nothing waits for anybody: your score lands on your first round, with nobody else playing, and nothing about it ever depends on other players.

Three modes, one rule each:

• Normal. A headline, the publication, and about 200 characters, with a link to read the article first. 10 points.
• Quick bites. A headline and one passage of at most 200 characters chosen from anywhere in the body, and no publication, so the source can never be the tell. 6 points.
• Deep. Normal, and then you say why: tap the reasons that fit, add a line if you like. 15 points. A line of text has to be a real line, six words or more that read as words, not a copy of the card and not a repeat of something you already wrote. When it is not, the game says “Say a bit more” and waits.

Correct means reading the article the way our model did. The reveal says so in the biggest type on the screen, then what we called it, then the points. When we called it slop, it also says what the software noticed: up to three of the documented signs of machine writing found in the text, each with a quote of 15 words or fewer. AI vocabulary, the rule of three, “not just X but Y”, em dashes everywhere, the tidy summarising closer, negative parallels, unnamed experts, and the tacked-on participle ending. People write all of these too, which is why a sign only counts on density and why the reveal frames them as what the machine noticed, never as proof.

The percentage behind the call is there for anyone who wants it, under Settings, and off by default. The verdict is the answer. The number is a detail.

What that costs, since it is worth saying. A player chasing points drifts towards guessing what the software says instead of looking, and those calls are partly a copy of it. The answer is that the payout and the record are different things. What is paid for is agreement. What is KEPT is every call, and above all the ones that disagree: a round where a person read something the other way from the software is a round that says where the software is wrong, and those are the valuable ones. No call is ever discarded for disagreeing.

How sure you were is read off the gesture, a short swipe or a tap against a long swipe or a held button. It is recorded and it changes nothing. What other players said about the same card is counted and stored and you are never shown it: seeing a count of strangers before you look changes what you see, and independent calls are the only ones worth keeping.

Streaks, badges and the leaderboard sit on top of all this and change none of it. A badge is a fact about your own rows, worked out when you ask. The leaderboard shows a handle, never a name, and the percentage beside it is how often that player read a card the way our model did.

Bring your own

You can paste a link or a text and see what plopslop would call it. A link is fetched the way everything else is, honouring robots.txt and the site's pace, and a site that keeps bots out stays out. The text is read by the same detector the feed uses, with the same floor: under 300 words, nothing is read at all. The estimate is shown to you and to nobody else, it is never listed anywhere, and the copy plopslop keeps to score it is private and deleted with your account. Then you argue: your own call, with the same reasons the game offers. That argument goes into the record with the rest.

The feed

Everything at /read is there because the site that published it keeps passing automated checks. No person reads an article before it appears and no person approves a site.

One detector reads every article and returns how sure it is that a machine wrote it, from 0 to 100. Past 80 it is set aside on that read alone. Below 50 the article goes up and nothing else is consulted. Between 50 and 80 a second detector, a different model that makes different mistakes, reads the same article, and it takes agreement at 85 or more to set the article aside. One uncertain read never decides anything by itself.

Both detectors are open models running on plopslop's own server. No article text leaves it. Articles under 300 words are never scored at all, because every detector gets worse on short text.

A single read can be wrong either way, so no article is ever judged on its own. A site is judged on its last 50 scored articles with the statistical room for error that a sample of 50 deserves, and it keeps its place while enough of them pass. A new site joins once its first 5 scored articles clear the same bar, which allows a stray flag or two. A site that drops below the bar stops appearing, is rechecked after 30 days, and comes back on its own when its newer work passes.

The badge on each card

top 87% means the article ranks above 87 percent of everything plopslop published in the last 14 days. Three things go into it: an estimate that a person wrote it, measured prose quality, and the standing of the publication. It orders a list. It says nothing about who wrote anything.

The first of those starts as the detector's reading and listens to players as they call the article in the game. With no calls the detector decides alone. At 8 calls people carry a quarter of that estimate. The more calls, the closer to half, and never past it, so the detector and the players end up equal partners and neither can shout the other down. A Deep mode call counts one and a half times, and a call made in under 1.5 seconds counts nothing, because that was not a reading. A card that has been called says how many player calls went in, and never what they said.

The number inside the game

The game can show the other side of the same reading: the detector's estimate that a machine wrote the text, 0 for clearly human, 100 for clearly machine. It appears in the game and nowhere else, only when a player switches it on, always as an estimate, and every round you play is a check on it.

What every call records

A call keeps the article, the mode, what you called it, how you made the call (a swipe, a tap or a key), how sure you were, the estimate the system had whether or not you looked at it, whether you matched it, how long you took measured by our own clock, whether you opened the article first, which of the deck's pools the card came from, in Deep mode the reasons you tapped and any line you wrote, and the signs the reveal showed you. In Quick bites it also keeps which passage you were shown.

Those rows do two jobs. They give you a score and a place on the leaderboard, and together they are the human check that shows where the software is wrong. plopslop intends to release them as a dataset, free or under a paid licence, with no usernames and no email addresses in it, and may share them with other efforts against machine slop. Nothing here trains a language model. The privacy policy lists every field, how long it is kept, and what a release means for you.

Pictures

Pictures are next. The picture game is closed while its deck is rebuilt on a pool whose origin is certain: 602 pictures declared machine made by their own uploaders and 601 photographs confirmed by the camera make and model in their own files, every one with the page that says so. The detector the old game scored against caught 16 percent of pictures whose uploaders had declared them machine made, and none at all from the three most widely used tools, so it will decide nothing when the game returns.

What these numbers are not

The scores on plopslop are automated statistical estimates, and not one of them is a finding. plopslop makes no claim of authorship about any specific text or publisher, and none about any specific picture or whoever made it. Absence from plopslop implies nothing about any publication.

Detecting machine-written text is a young field with real error in both directions, and the checks above are built around that error rather than around a hope that it is small.

Privacy

Signed out, plopslop records nothing about what you read. An account is optional for the feed and needed for the game. Signing in saves your score and preferences and lets your feed learn from what you open. No third-party trackers, ever, and no analytics product on the site at all. The privacy policy lists every field plopslop stores and how long it keeps it.

For publishers

plopslop keeps a title and 200 characters or fewer of any article, shows nothing more, links every card straight to you, and honors robots.txt on every fetch. Any site that asks to be removed is removed. Write to hello@plopslop.com.

The same address takes picture removals. Send the address of the picture or of the page it is on and it leaves the pool. The terms set out that route for rights holders and for anyone who appears in a picture.