Skip to the content
Dead Writing Theory

The Court of Public Prose

From the clerk's desk

Kurzgesagt on AI Slop: Which Claims Can Evidence Check, and Which Are Opinion?

Kurzgesagt's video says AI slop is destroying the internet. Some of its claims can be measured and some are judgement. YouTube's reported mislabelling of the channel shows how fragile a one-item 'slop' verdict is.

By Jack Stovell · Published

Short answer: the video makes three sorts of claim. One is how much of the internet is machine-made, and a study can count that. Another is how hard it is to tell real from fake, which is partly countable. The headline claim, that slop is destroying the internet, is a judgement. The YouTube episode that followed was a verdict on one item, and that verdict went wrong. A one-item verdict is the weakest kind of evidence in this whole argument.

What the video and the reports say

The video is “AI Slop Is Destroying The Internet”, uploaded to the Kurzgesagt In a Nutshell YouTube channel on 7 October 2025. Its description sets out the case in four sentences:

AI-generated content is flooding the internet, from videos and music to news and books. On top of that, generative AI is making false information look more convincing than ever. We’re entering a new era of information overload, and figuring out what’s real and what’s not is becoming more and more of a challenge. While AI slop is slowly changing the internet as we know it, one thing is certain: kurzgesagt videos will always remain human-made.

Then came the twist. On 31 July 2026 the channel posted on X that its latest video, about superpredators, had been its worst upload since 2013, even though viewers clicked it, watched it and liked it. After the channel contacted YouTube, it wrote: “It turned out that YouTube’s automatic AI detection tools wrongly think that our very much human-made videos are AI Slop.” It took the video down to re-upload later and added that “hopefully the bug is fixed for now”. Kotaku and Dexerto both reported the episode. Both rely on the channel’s post. I have not seen YouTube’s own account of what its system did.

One more document helps. In her Guardian column of 8 January 2025, Arwa Mahdawi defined slop as “low-quality text, videos and images generated by AI”. That definition does two jobs. “Generated by AI” is about where something came from, which can be checked in principle. “Low-quality” is an opinion. Most arguments about slop come down to that split.

Three kinds of claim: measurable, partly measurable, taste

Sort the video’s sentences and they fall into three piles.

Measurable. “AI-generated content is flooding the internet” is a prevalence claim: what share of what is out there was made by a machine, and how fast that share is growing. That’s hard to measure well, but it’s a counting question. A related question runs through the wider debate, though the video’s description doesn’t raise it: is writing converging, with more of it sounding alike? You can test that too, by comparing texts over time.

Partly measurable. “Generative AI is making false information look more convincing” has a testable core. You could show people true and false material and record what they believe. But “convincing” to whom, and in what setting, is a choice the researcher makes, and the answer moves with it. “Figuring out what’s real” works the same way. You can measure how accurate detectors and people are. Whether that counts as “more and more of a challenge” depends on a baseline someone has to pick.

Taste. “Destroying the internet” is a judgement. Two people can agree on every number and still disagree here, because “destroyed” depends on what you wanted the internet for. That doesn’t make the judgement silly. It means no study will settle it. The closing promise, that the channel’s videos “will always remain human-made”, is about the future, so evidence can’t check it today.

What the docket can and cannot say about the measurable ones

First, scope. The cases on this docket are about writing. The video’s flood includes videos, music and books, and on those the docket has nothing to say.

On prevalence in text, two cases apply. Half the articles, and a plateau counts articles. It’s a vendor study and isn’t peer reviewed. The websites nobody wrote counts whole sites, and it’s a preprint. The units differ, so the answers differ. Neither one measures how much of what people actually read is machine-made. Both rely on AI-text detectors, and the authors of the second warn that, when trying not to wrongly accuse human writers, “detectors of LLM-generated text perform much worse than advertised.” Both studies check their detectors against material from before ChatGPT, which is the right habit.

On convergence, read better stories, more alike, a lab experiment where AI story ideas made stories more alike, next to the news that didn’t flatten, where the vocabulary of news articles didn’t narrow. Together they don’t give one tidy answer, and they measure something different from the prevalence cases. A web can contain a lot of machine text without all sounding the same, and the reverse is possible too.

What the docket can do is say whether a measurable claim has support, in which corpus and by which method. It cannot say whether any of that makes the internet worse. When an argument moves from “more of it is machine-made” to “so it is ruined”, it has left measurement for taste. The theory page sets out how narrowly this site defines its own question.

When a human-made channel is flagged as machine-made

The YouTube episode works as a lesson in how to read a slop verdict.

An automated system looked at one channel’s videos and decided they were low-quality AI content. By the channel’s account, YouTube agreed the call was wrong. That’s a false positive, and it landed on one item with real consequences: the video’s reach collapsed.

Every classifier makes errors. On a single item there’s nothing to average over, so one wrong call is simply wrong. A study estimating the share of machine-made text in a large corpus works differently. It doesn’t need to be right about every page. It needs to know roughly how often it’s wrong, measure that on material that can’t be machine-made, such as text from before ChatGPT, and report it. Individual mistakes then become a known margin. That’s why a well-built corpus estimate deserves more trust than a confident verdict on one page, one video or one channel, however satisfying the single verdict is to read.

Notice what the flag mixed together. The channel quotes YouTube’s tools as calling its videos “AI Slop”, and Kotaku’s headline says YouTube mistook the channel for AI slop. That label combines an origin judgement with a quality judgement, the same split as in the Guardian definition. When a label combines the two, it’s hard to know which half failed.

The episode doesn’t tell us how much of YouTube is machine-made, or how often the system errs. It is one data point. Its lesson is narrower: a slop flag is a claim, from a person or a machine, and it can be wrong.

For why boredom with machine text can feel like evidence when it isn’t, see what “your AI slop bores me” shows.

How to read a slop claim, a short checklist

  1. Sort it. Is it about how much (measurable), how hard it is to tell (partly measurable), or how bad (taste)?
  2. Ask for the unit. Pages, sites, videos, or what people actually see? Each gives a different answer.
  3. Ask for the method. If a detector did the counting, look for a false-positive check on older material, and for whether the study names it.
  4. Separate prevalence from convergence. “More of it is machine-made” and “it all sounds the same” need different evidence.
  5. Treat a one-item verdict as one data point. That goes for a viral “this is obviously AI” post as much as for a platform’s flag.
  6. Notice the word “worse”. Once it appears, you’re dealing with judgement. Decide whether you share the standard behind it.
  7. Find the primary source. A headline about a channel’s statement is not the statement. Read the post itself.

Sources

Questions people ask

Did YouTube flag Kurzgesagt as AI-generated?

According to the channel, yes. On 31 July 2026 Kurzgesagt posted on X that YouTube had confirmed its "automatic AI detection tools wrongly think that our very much human-made videos are AI Slop", after a video about superpredators became its worst upload since 2013. Kotaku and Dexerto reported it. All of these accounts rest on the channel's own statement.

What does Kurzgesagt say about AI slop?

Its video "AI Slop Is Destroying The Internet" (7 October 2025) says AI-generated content is flooding the internet, that generative AI makes false information look more convincing, and that telling real from fake is getting harder. It ends with a promise: "kurzgesagt videos will always remain human-made."

Is Kurzgesagt itself made with AI?

The channel says its videos are human-made, in the video description and again in its July 2026 post. Nothing we have read contradicts that. YouTube's flag doesn't count as evidence against it, because the channel says YouTube agreed the flag was wrong.

Is AI slop the same as the web's writing sounding alike?

No. Slop is about origin plus low quality, and quality is partly judgement. Sounding alike is convergence, which a study can measure by comparing texts. A web could be full of machine text that varies a lot, or full of human text that all sounds the same.

Can anyone reliably tell AI content from human content one item at a time?

Not reliably enough to treat a single verdict as proof. Every classifier has an error rate, and a single item gets no averaging to soften it. Studies that estimate shares across a large corpus can measure their false-positive rate on older material and report it, which is why they are the sturdier evidence.