--- title: "Is AI Slop Bad? Which Harms Evidence Can Check and Which Are a Matter of Taste" url: https://deadwritingtheory.com/blog/is-ai-slop-bad/ summary: "Partly. 'Bad' bundles four claims: there is a lot of it, it all sounds alike, it is worse, it does harm. The first two have been measured; quality and reader harm have not, here." published: 2026-11-10 updated: 2026-11-10 author: "Jack Stovell" publisher: "Adapt Progress Evolve Limited" language: en-GB --- # Is AI Slop Bad? Which Harms Evidence Can Check and Which Are a Matter of Taste *By Jack Stovell. Published 10 November 2026.* Partly. 'Bad' bundles four claims: there is a lot of it, it all sounds alike, it is worse, it does harm. The first two have been measured; quality and reader harm have not, here. Short answer: "bad" bundles four separate claims, and each needs its own kind of evidence. This court has tested two of them: how much machine text there is, and whether writing is growing alike. It has not measured quality or harm to readers. What is left after that is taste. ## Why "is it bad?" is four questions Start with the word. Merriam-Webster made *slop* its 2025 Word of the Year and defines it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." The definition already includes a judgement on quality, so "is slop bad?" can sound circular. In practice, people who say it usually mean something they could check, and it tends to be one of four things: 1. There is a lot of it. 2. It all sounds the same. 3. It is worse than what it replaced. 4. It harms readers, or the models later trained on it. Arguments slide between these. Someone proves the first and talks as though they have proved the fourth. Below, each claim gets the same treatment: what it means, what evidence would settle it, and what the record here holds. ## Claim 1: there is a lot of it (prevalence) This claim comes with the most numbers, which is why people mistake it for the others. Prevalence studies estimate how much text is machine-written. That is their whole job. They do not grade the text. A machine-written page can be useful and a human-written page can be dreadful, and a counter scores both the same way. Two cases on the docket measure prevalence: [half the articles and a plateau](/cases/half-the-articles-and-a-plateau/), a vendor study of new English articles, and [the websites nobody wrote](/cases/the-websites-nobody-wrote/), a preprint on whole websites. Both rely on AI detectors, and detectors misfire. Graphite, the vendor behind the first, ran its three detectors on articles from before ChatGPT and reports false-positive rates of Pangram 1.844%, Copyleaks 1.836% and GPTZero 1.355%. So some human writing gets counted as machine writing. Rates that low leave a corpus-level estimate standing. They are still a reason to hold the headline loosely, and never to treat one page's detector verdict as proof. There being a lot of it is a fact about quantity. However big the number, it cannot show quality. ## Claim 2: it all sounds the same (convergence) Sameness can be tested, so nobody has to take anyone's word for it. Take a defined set of texts, measure how alike they are, and compare against a baseline: the same kind of writing from before, or written without the tool. If the machine-era texts cluster more tightly, that is convergence. If they do not, there is none to report. Two cases come at this from different directions. In [better stories, more alike](/cases/better-stories-more-alike/), a preregistered experiment, writers given AI story ideas produced stories that were each rated better and, as a group, more alike. The verdict is split. In [the news that didn't flatten](/cases/the-news-that-didnt-flatten/), machine-style words rose in English news while lexical diversity did not fall: a point for the defence on vocabulary, short of a knockout. The caveats on those pages matter as much as the findings. Then comes the step people skip. Sameness evidence tells you texts resemble each other. It does not tell you the resemblance is bad. A house style produces sameness on purpose. Whether sameness is a loss depends on what was lost, and the measurement alone cannot say. ## Claim 3: it is worse than what it replaced (quality) This is the claim most people care about, and this court has not measured it. No case here tests whether machine-written text is lower in quality than the writing it displaced. The nearest is the story experiment, and for individual pieces it points the other way: stories written with AI ideas were rated better. Those writers got ideas rather than finished text, in a lab, writing short fiction. It says nothing about the content-farm page you met this morning. What would settle it? You would need a stated standard (accurate, useful, clear, whatever you decide counts), a fair comparison set, and judges or measures that do not know which text came from where. You would also need to define "replaced". A machine-written page did not always displace a careful human essay. Sometimes it displaced nothing. Sometimes it displaced a worse human page. So when someone says slop is worse, ask two things: worse than what, and who checked? If the answer is "it feels worse", that belongs in the taste section below. ## Claim 4: it harms readers or future models Two claims share this heading, and they need different evidence. Reader harm means something happens to people: they are misled, lose trust, waste time or fail to find what they need. To show it, you would observe readers, or measure outcomes such as whether they come away believing something false. Reading a pile of text and deciding it looks harmful does not show it. Asking people how they feel about AI content measures opinion, which is a different thing from harm. This court has no case on reader harm. Future-model harm concerns training data. The worry is that models trained on machine output degrade over generations. That is model collapse, and [what is model collapse](/blog/what-is-model-collapse/) sets out what the experiments did, and the condition that changes the result: whether the original human data is kept. It is a technical claim with a technical test. It says nothing about whether a person is worse off after reading a machine-written page. Evidence for one of these regularly gets spent on the other. Watch for it. ## What is left over is taste, and that is allowed Take out the testable claims and something remains. It bores me. It feels cheap. Every paragraph has the same rhythm, and reading it is tiring. That is a fair thing to feel and a fair thing to say. The mistake is presenting it as a finding. "I dislike this" is honest. "Research shows this is harmful" needs research. Keep the two labelled and nobody is misled, including you. The court reporter takes no side on whether you should dislike it. [The theory](/theory/) sets out the narrower claim this site does test, and how it tests it. ## How to judge a "slop is bad" claim you meet Run it through five questions. 1. Which of the four claims is it making? 2. If it is about prevalence, does it go on to conclude something about quality? It cannot. 3. If it is about sameness, what was the baseline, and what was measured: vocabulary, structure or ideas? 4. If it is about quality, who judged, against what standard, and did they know which text was which? 5. If it is about harm, is the harm to readers or to models, and where is the evidence for that one? Then watch for three moves: one number doing three jobs, a detector estimate quoted without its false-positive rate, and a sentence that opens with a measurement and closes with a moral. Somewhere in between, a step was skipped. ## Sources - [Merriam-Webster, 2025 Word of the Year: slop](https://www.merriam-webster.com/wordplay/word-of-the-year) - [Paredes, Druck, Benson and Smith, "AI Now Writes as Many Online Articles as Humans", Graphite (vendor study, 2026)](https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do) ## Questions people ask ### Is AI slop actually bad, or just annoying? Those are different kinds of claim. Annoyance is taste, and you are entitled to it. 'Bad' in the sense of lower quality or harmful needs evidence, and this court has measured neither quality nor harm to readers. Annoying, for many people; worse or harmful, untested here. ### Does AI-generated content hurt the internet? It depends which harm you mean. How much machine text there is can be estimated, with detector false-positive caveats. Whether that hurts readers is a separate question this record does not answer. Whether it hurts future models is the model collapse question, which concerns training data rather than readers. ### Is lots of AI writing the same as low-quality writing? No. Prevalence studies estimate how much text is machine-written, not how good it is. A large amount of something can be fine, mediocre or terrible. You would need a quality measure, and the prevalence studies do not supply one. ### Can evidence show that AI slop is harmful? Some of it could, in principle. Sameness can be measured against a baseline, and model degradation can be tested in training experiments. Reader harm would need observation of readers or measured outcomes. Objections of taste cannot be settled by evidence, because they are preferences. ### Is AI slop the same as AI-assisted writing? Not necessarily. Merriam-Webster defines slop as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." Low quality and bulk are part of the definition, so using a tool somewhere in the process does not by itself make the result slop.