A short history of manufactured mediocrity, and why craftsmanship — not the tool in hand — has always been the real dividing line
A word from the mudflats
Long before anyone used it to describe a six-fingered hand in a Facebook photo, “slop” meant something you’d find in a pig trough. In Middle English it named a mud hole; later it stretched to cover any wet mess sloshed about, and eventually the watery swill fed to farm animals. By the 1950s it had drifted into slang for rubbish generally, and by the 1990s it was being aimed at inferior media products. Then came the final mutation. Merriam-Webster named “slop” its Word of the Year for 2025, defining it as low-quality digital content produced in quantity by artificial intelligence. Weeks later, in January 2026, the American Dialect Society independently reached the same verdict in its 36th annual vote — and noted something telling: “AI slop” had been a nominee the previous year, but by 2025 the word could stand alone, the AI part simply assumed.
That assumption is the subject of this paper. The 300-year journey from mudflat to Facebook feed is worth pausing on, because at no point along it did “slop” describe a technology. It described a relationship between a producer and their output — specifically, the absence of one. It is a word for contempt, applied wherever someone stops caring whether the thing they are making is any good.
A great deal of AI output genuinely earns it. But the leap from “much AI content is slop” to “AI content is slop” is a category error, and it is one we have made before, about other tools, in other centuries. This paper traces where the error comes from, examines the pre-AI history the term has largely forgotten, and argues for a diagnostic that actually works — one that catches bad AI content and bad human content alike, and clears the good in both.
The mill before the model
Slop did not wait for large language models. Its most instructive precedent is barely fifteen years old and involved no machine learning whatsoever.
In October 2009, Wired published Daniel Roth’s profile of Demand Media under the headline “The Answer Factory,” subtitled with the phrase that would define the era: the fast, disposable, and profitable-as-hell media model. The operation was elegantly simple. Algorithms mined search data to identify precisely what people were typing into Google. Those queries became article assignments. Freelancers — thousands of them — were paid a few dollars apiece to fill the gap, fast. By 2009 the company was reportedly publishing around 4,000 articles and videos a day; founder and CEO Richard Rosenblatt told Wired that within a year Demand expected to be producing a million items a month, which he framed as the equivalent of four English-language Wikipedias annually. (That larger figure was a projection, not an audited result — but the ambition itself is the point.)
The output was exactly what the incentives predicted. New York Times critic David Carr held up eHow’s Super Bowl party guide, which advised readers to buy beer and keep it in a nearby cooler so they wouldn’t have to visit the fridge mid-game. Other entries explained how to pick blueberries (tie a bucket to your waist) and, notoriously, how to open a refrigerator door. One Demand Studios writer told MediaShift in 2010, on the record, that he was fully aware he was producing garbage.
Google’s response was the February 2011 algorithm update known internally as Panda and initially, tellingly, as “Farmer.” It affected roughly 12% of English search queries and was aimed squarely at thin, low-value content. The aftermath is more interesting than the usual telling admits: eHow initially appeared to survive — its traffic reportedly rose the day after launch, and Demand Media publicly insisted the impact had been overstated — while smaller properties in its network were gutted. The reckoning came later. Demand posted a $6.4 million loss in the fourth quarter of 2011 and $18.5 million across that fiscal year, cut writer assignments, and by 2013 was the subject of a Variety post-mortem titled “Epic Fail.”
Three things are worth extracting from this. First, none of it involved generative AI: the slop was produced by human beings, at scale, because the economics rewarded volume over value. Second, the market did eventually correct, but slowly and imperfectly — bad content is not self-limiting on any useful timescale. Third, and most importantly, the variable that produced the slop was never the tool. It was a payment structure that made caring irrational.
Slop is what happens when content is optimised for quantity at minimum cost, indifferent to whether anyone is served by it. That describes Demand Studios in 2010 as precisely as it describes an image generator pumping out engagement bait in 2026. The instrument changes. The incentive does not.
The cover band problem
There is a second comparison the AI-slop debate keeps skipping past, and it opens up a category the word has no room for.
A tribute act plays music it didn’t write, note for note, for an audience that already knows every lyric. By the strict standard of originality this is about as derivative as culture gets; nobody claims these are visionaries pushing the form forward. And yet the circuit is a durable business. Industry estimates — which should be treated as trade figures rather than audited data — put the number of active tribute acts worldwide in the region of 15,000, with something like 1.7 million annual ticket sales in the United States alone.
The more rigorous finding comes from economics. A study of the German tribute-concert market found that ticket prices are significantly higher when the original act no longer exists — when the copyright holders have retired, or died. That is a precise, measurable statement about what audiences are actually buying. They are not paying to be fooled. They are paying for access to an experience that is otherwise unavailable, delivered competently, close to home, for a fraction of an arena ticket.
This is the missing middle category: work that is unoriginal but conscientious — that fills a real gap, that respects the audience even while borrowing someone else’s material wholesale. A skilled tribute band rehearses the setlist, learns the harmonies, gets the guitar tone right. That effort is exactly what separates it from a bar band that mangles the changes and can’t be bothered to tune up. Both are derivative. Only one is slop.
Generative AI maps onto this distinction far better than the discourse admits. A model prompted carelessly, deployed to flood a platform with volume, and never reviewed by anyone who cares is the algorithmic equivalent of the band that never learned the song. A model used deliberately — briefed properly, checked against sources, edited, structured, aimed at a real reader with a real need — is closer to the tribute act that honours its audience even though nobody involved wrote the melody. Neither is creating something new in the deepest sense. Only one deserves contempt.
Where the term actually came from
It is worth being precise about the history, because the discourse flattens it.
“Slop” as a label for bad AI output goes back to around 2022, emerging in reaction to the first wave of publicly available image generators, with early usage documented on 4chan, Hacker News, and in YouTube comments. It was popularised more widely in 2024 by a writer and technologist posting as “deepfates,” who defined it simply as unwanted AI-generated content — material nobody asked for, pushed into spaces where it wasn’t invited. Simon Willison, the British developer often credited alongside him, has made the operative distinction repeatedly: the objection is to content that is mindlessly generated and thrust upon someone who didn’t ask for it. Both were explicit that not all AI content qualifies. Some AI-assisted work, they noted, is original, surprising, and good enough to hang in museums.
The term was coined by people trying to protect a distinction, not collapse one.
That nuance has not survived contact with the wider public. By 2026 “AI slop” had stretched to cover everything from spam-farm SEO articles to political deepfakes — the latter now with its own coinage, “slopaganda,” for AI content built specifically to manipulate belief at scale. The word’s precision as a diagnostic has been eroded by its convenience as an insult.
From diagnosis to verdict
Which brings us to the mechanism that matters most for anyone trying to use these tools carefully: “slop” has become less a description of a specific piece of work than a reflex applied to a category.
MIT Technology Review, reporting on the backlash in late 2025, put it plainly — the label has become convenient shorthand for dismissing almost any AI-generated output regardless of its actual quality. The creators interviewed for that piece described receiving hostile messages simply for using the tools at all, whatever the finished work looked like. The word had stopped functioning as a quality judgment and started functioning as an identity marker: not this is bad, but this was touched by a machine, and that settles it.
A concrete case makes the mechanism visible. In 2025 the San Francisco International Airport Museum mounted an exhibition, “Women of Afrofuturism,” including AI-assisted portraits by the Boston-based artist Nettrice Gaskins — an established practitioner with a body of work long predating generative tools. The pieces were dismissed online as “AI slop.” Gaskins responded that she uses AI as one instrument among many and that the work is grounded in her own creative practice; curators pointed to positive visitor response. Whatever one concludes about the portraits, the dismissal did not arrive as a judgment about the portraits. It arrived as a judgment about the method.
Some of this is a much older cultural reflex wearing new clothes. Commentators have drawn a direct line from “AI slop” to “kitsch,” the pejorative that acquired critical weight through Hermann Broch in 1933 and Clement Greenberg in 1939, and which was used to dismiss the flood of cheap, industrially reproduced art — colour lithographs, factory-made decor — as inherently lesser than the handmade original. Both terms do double duty. Both identify real low-quality output. Both also police the boundary of what counts as legitimate art in the first place, sometimes with little reference to the specific object under discussion.
The honest counter-argument. It would be too easy to file all of this under gatekeeping, and doing so would be its own kind of laziness. A substantial strand of the skepticism isn’t a quality claim at all — it’s a claim about method. Critics in this camp point out that AI image generation has become markedly more polished over the past few years, the mangled hands and garbled text largely resolved, and yet public sentiment has soured rather than warmed. They take this as evidence that people are not reacting to surface defects but to something structural: the absence of a human hand in each decision, the displacement of working illustrators and writers, the sense of being served a statistical pastiche however convincing the surface. That is a coherent position. It is also an argument about labour and about what art is for, not about whether a given artefact is well made — and it deserves to be met on those terms rather than waved away.
The two uses of the word are doing different jobs, and conflating them causes damage in both directions. When “slop” means low-effort, low-value, indifferent to the audience, it is a sharp diagnostic that cuts a 2009 content farm and a 2026 prompt-and-post operation with equal force. When “slop” means made with AI, full stop, it becomes a loyalty test — one that penalises careful, well-engineered AI work exactly as harshly as the laziest possible use of the same tool, while granting lazy human work a pass simply for having been typed by a person. A term that cannot distinguish a source-checked, heavily edited piece from a one-line prompt dumped onto a blog is no longer measuring quality. It is measuring provenance and calling it taste.
It is worth adding that the reverse error is equally live. “AI-assisted” is not a quality credential either. The most acclaimed AI artwork in a major institution — Refik Anadol’s Unsupervised, which trained a model on MoMA’s digitised collection, drew large crowds from November 2022, and entered the museum’s permanent collection in 2023 — has divided serious critics sharply. New York magazine’s Jerry Saltz dismissed it as crowd-pleasing mediocrity and compared it to a lava lamp; Artforum‘s Lloyd Wise found it spellbinding and read it as a genuine dialogue with modernism. That is what a real aesthetic argument looks like: two critics disagreeing about a specific work on the evidence of the work. It is precisely the conversation that “it’s AI, therefore slop” forecloses — and precisely the conversation that “it’s in MoMA, therefore art” forecloses too.
What slop actually diagnoses
Strip away the medium and a pattern emerges. Slop — whether produced by an underpaid freelancer chasing a keyword quota in 2009 or a model prompted with “write me a blog post about X” in 2026 — tends to share four features:
- No addressee. It isn’t made for anyone in particular. It’s aimed at an algorithm, a quota, or a search query, with the human recipient as an afterthought.
- No editorial judgment. Nobody asked whether it was good, only whether it was finished.
- No accountability. No one is willing to attach their name to the specific claims, or to be embarrassed if they’re wrong.
- Optimisation for output, not outcome. Success is counted in units shipped rather than value delivered.
None of these require a machine. Ghostwritten celebrity memoirs, template-stuffed local news sites, mass-produced motivational posters — human industries have been hitting all four for a very long time. And none of them are automatically triggered by a machine’s involvement. A model used with a clear brief, real source material, a defined audience, and an editor willing to reject a bad draft misses every one.
A working test
For anyone commissioning or reviewing content, the practical questions follow directly from the list above — and none of them mention the tool:
- Who is this for, specifically? If the answer is a search engine or a posting quota, stop.
- What did it cost to check? Not to produce — to verify. Slop is cheap to make and cheaper to fact-check, because nobody does.
- Who is accountable for the claims? A named person who would be embarrassed to be wrong is the single strongest quality signal available.
- What was rejected? Work with no discarded drafts, no cut sections, no killed angles has not been edited. Editing is the difference between output and a decision.
- Would removing it make anything worse? If the honest answer is no, it is filler regardless of authorship.
A piece produced with heavy AI assistance can pass all five. A piece typed entirely by hand can fail all five. That asymmetry is the whole argument.
The limits of this argument
Everything above evaluates artefacts one at a time. That is a real limitation, and four objections deserve to be met head-on rather than buried.
1. A test for one article is not a defence of an ecosystem. The tribute band analogy breaks on scale. A bad cover band ruins an evening for the fifty people in the room; the damage is bounded by the size of the venue. Generative output has no such ceiling. The grievance many people hold is not that any individual piece is unoriginal but that the aggregate pollutes a commons — that finding anything worth reading now carries a verification cost that didn’t previously exist. Five good questions help a reader assess a document in front of them. They do nothing about the cost of arriving at that document in the first place.
That said, the strongest version of this claim does not survive contact with the evidence. Europol forecast in 2022 that 90% of online content would be synthetic by 2026; it isn’t. Graphite’s analysis of Common Crawl, updated through Q1 2026 and cross-checked against three separate detectors, finds the share of primarily AI-generated English-language articles plateaued near 50% in mid-2024 and has stayed there — 50.9% in Q4 2025, 49.9% in Q1 2026. Separately, an Ahrefs study of 900,000 pages found that while roughly 74% contained some AI-generated text, only 2.5% were pure AI with no human involvement; the overwhelming majority were human-AI blends. And the filler appears to be losing: 82% of articles cited by ChatGPT and Perplexity are human-written, and primarily AI-generated articles rank lower in Google. Graphite’s own hypothesis for the plateau is that practitioners discovered this content doesn’t perform.
All of these figures carry real caveats — detector reliability is contested, Common Crawl under-samples paywalled publishers, and the blend of human and machine input is increasingly hard to classify at all. But the shape is clear enough, and it looks less like a dead internet than like Panda happening again, more slowly: a flood, then a correction, then a plateau. The commons problem is real. The extinction narrative is not.
2. The environmental objection is real, and it is not an argument this paper can win. Running these models is not free the way lifting a pen is free. The International Energy Agency puts global data centre electricity consumption at roughly 415 TWh in 2024 — about 1.5% of world demand — and projects it to roughly double to 945 TWh by 2030, with AI-accelerated servers growing near 30% a year. In the United States, data centres consumed around 183 TWh in 2024, over 4% of national electricity. Water consumption for cooling is substantial and rising.
It is tempting to answer that a printing press, a typewriter, and a paper mill are not free either — that their costs are simply front-loaded into manufacturing and distribution rather than amortised across each use. That is true, and it is a fair correction to any framing that treats pre-digital tools as weightless. But it does not dispose of the objection, because the relevant comparison is aggregate footprint at civilisational scale, not per-artefact cost, and on that measure the numbers above speak for themselves.
The more honest response is that this objection is orthogonal to the question at hand. Environmental cost attaches to the technology, not to the quality of what is made with it. It falls identically on a meticulously edited report and on a thousand pieces of engagement bait — and arguably falls harder on the careful work, since real editing means more inference passes, not fewer. It is therefore a serious argument about whether and how much to use these systems. It is not an argument that can sort good output from bad, which is the only question this paper claims to answer.
3. Slop is a rational strategy, not merely a failure of care. The framing so far has located responsibility with the producer, and that is incomplete to the point of being convenient. Actors optimise for the environment they are placed in. When platforms reward frequency, freshness, and zero-friction volume, then flooding a network with generated filler is not an editorial lapse — it is a correctly calculated play for algorithmic territory. Meta’s leadership has acknowledged concern about low-quality content while simultaneously declining to make judgments about what should be allowed to flourish, on the reasoning that formats which once looked like noise later became mainstream. Both halves of that position are defensible. Together they describe a system that will not police itself.
This does not weaken the paper’s thesis; it completes it. The Demand Media story was never about lazy freelancers. It was about a payment structure that made caring irrational, and the writers responded correctly to the incentives in front of them. The same is true now, at greater scale, with the incentive designed by ranking and recommendation systems rather than a per-article rate card. If the objection to slop is genuine, the intervention point is the incentive, not the individual’s conscience — and treating it as a matter of personal craftsmanship lets the parties with actual leverage off the hook.
4. The feedback loop is a genuine disanalogy. Here the Demand Media comparison does fail in a specific and important way. A bad eHow article sat inert. It cluttered search results, but it did not make the next generation of writers worse. Generated text does re-enter the corpus that subsequent models are trained on, and the consequences of that loop are documented: Shumailov and colleagues, publishing in Nature in 2024, showed that models trained recursively on their own output degrade, losing the tails of the distribution and converging toward bland, over-probable, eventually incoherent text.
Precision matters here, though, because the finding is routinely overstated. Collapse was demonstrated under a replacement regime, where each generation trains on its predecessor’s output instead of on real data. Subsequent work has shown that accumulating data — training on the union of real and synthetic material rather than substituting one for the other — substantially mitigates the effect, and that a sufficiently high proportion of genuine human data prevents it. That is much closer to how training corpora are actually curated. The loop is real, it is a legitimate disanalogy with the content-farm era, and it is a strong argument for provenance tracking and data hygiene. It is not a demonstrated trajectory toward inevitable collapse.
What survives. Taken together, these objections narrow this paper’s claim rather than refute it. The argument here is about how to judge a given piece of work — and on that question, the tool used remains the wrong variable. The objections above are about something else: the health of an information commons, the resource cost of an industry, the design of platform incentives, and the integrity of training data. Those are all more consequential than the question of whether a particular essay was worth reading. But they are not answered by refusing to distinguish good work from bad, and a term that flattens that distinction makes them harder to discuss, not easier.
The actual dividing line
So if the tool isn’t the diagnostic, what is? The answer is unglamorous: engineering and intention — the same qualities that separated good work from bad long before any of this existed.
A well-researched, carefully structured report produced with substantial AI assistance — sources checked, drafts rejected, aimed at a specific reader’s actual question — is not slop merely because the author didn’t type every word. A lazily assembled human blog post, dashed off against a publishing quota with no verification and no audience in mind, is not exempt merely because a person wrote it.
This is not a claim that AI and human effort are interchangeable, nor that the flood of genuinely worthless AI content isn’t a real problem. It plainly is, and pretending otherwise would be its own carelessness. The claim is narrower: that “AI-generated” and “worthless” are not synonyms, and that treating them as synonyms lets bad AI content and bad human content off the hook for exactly the wrong reasons — the first by making the critique so broad it stops discriminating, the second by making it so narrow it never lands.
The tribute band that learns every harmony and the Demand Studios freelancer who knew he was writing garbage were using the same instrument — human labour — to opposite ends. Generative AI is no different in kind. It is a tool capable of producing the analogue of a tight, respectful cover version, or the analogue of a mud hole.
None of which settles the larger questions. Whether the commons can absorb this volume, what the energy is worth, who should bear responsibility for the incentives, and whether the corpus stays clean enough to train on are all live and all more consequential than any individual artefact. But they are questions about an industry, and the word “slop” is being used to answer a question about a piece of work — usually without looking at the work. Collapsing the two does not sharpen the critique of AI. It blunts it, by handing the technology’s defenders an easy target and giving genuinely lazy human output a place to hide.
The difference was never in the trough. It was always in who was doing the pouring, and whether they cared what came out.
References
- Merriam-Webster. “Word of the Year 2025: Slop.” Merriam-Webster, December 2025.
- American Dialect Society. “2025 Word of the Year Is ‘Slop.'” Press release, January 9, 2026. https://americandialect.org/2025-word-of-the-year-is-slop/
- Dictionary.com. “Slop — Slang.” https://www.dictionary.com/culture/slang/slop
- “How ‘Slop’ Became the Defining Word of 2025.” Fast Company, December 15, 2025. https://www.fastcompany.com/91460529/how-slop-became-the-defining-word-of-2025
- Wikipedia contributors. “AI slop.” https://en.wikipedia.org/wiki/AI_slop
- Wikipedia contributors. “Slopaganda.” https://en.wikipedia.org/wiki/Slopaganda
- Wikipedia contributors. “Content farm.” https://en.wikipedia.org/wiki/Content_farm
- Wikipedia contributors. “Cover band.” https://en.wikipedia.org/wiki/Cover_band
- Roth, Daniel. “The Answer Factory: Demand Media and the Fast, Disposable, and Profitable as Hell Media Model.” Wired, October 19, 2009.
- Britannica. “Content Farm.” Encyclopedia Britannica. https://www.britannica.com/topic/content-farm
- “The Rise of the Content Mill.” Center for Digital Ethics & Policy, September 28, 2011. https://digitalethics.org/essays/rise-content-mill
- “Epic Fail: The Rise and Fall of Demand Media.” Variety, December 10, 2013. https://variety.com/2013/biz/news/epic-fail-the-rise-and-fall-of-demand-media-1200914646/
- “Google Panda Update Costs Demand Media $6.4 Million in 4th Quarter?” Search Engine Land, 2012. https://searchengineland.com/google-panda-update-costly-112062
- “A Complete Guide to the Google Panda Update: 2011–21.” Search Engine Journal. https://www.searchenginejournal.com/google-algorithm-history/panda-update/
- “Slop Before the Machines: Why the AI Authenticity Panic Misses the Point.” Nearly Right. https://nearlyright.com/slop-before-the-machines-why-the-ai-authenticity-panic-misses-the-point/
- “Cool Cats or Copycats? An Economic Exploration of the Market for Tribute Bands.” 2019. https://www.researchgate.net/publication/330853892_Cool_cats_or_copycats_An_economic_exploration_of_the_market_for_tribute_bands
- “The Economics of Tribute Bands.” ProTributeBands.com. https://www.protributebands.com/the-economics-of-tribute-bands/
- “Why We Can’t Let It Be: The Booming Business of Beatles Tribute Bands.” Beatles Rewind, November 15, 2025. https://beatlesrewind.substack.com/p/why-we-cant-let-it-be-the-booming
- “AI Slop — How Every Media Revolution Breeds Rubbish and Art.” Scientific American, November 9, 2025. https://www.scientificamerican.com/article/ai-slop-how-every-media-revolution-breeds-rubbish-and-art/
- “How I Learned to Stop Worrying and Love AI Slop.” MIT Technology Review, December 23, 2025. https://www.technologyreview.com/2025/12/23/1130396/how-i-learned-to-stop-worrying-and-love-ai-slop/
- “Beyond the Scroll: Museums, AI, and the Value of Attention.” The Harvard Crimson, December 29, 2025. https://www.thecrimson.com/article/2025/12/29/art-ai-museums-thinkpiece/
- Wagner, Michael G. “AI Slop Is the New Kitsch.” May 11, 2025. https://www.theaugmentededucator.com/p/ai-slop-is-the-new-kitsch
- “The Idea of ‘AI Slop’ Is Slop.” The Philosophical Salon, December 1, 2025. https://thephilosophicalsalon.com/the-idea-of-ai-slop-is-slop/
- “AI Image Generation Has Gotten ‘Better,’ But Its Reputation Is Only Getting Worse.” https://oneoftheabove.substack.com/p/ai-image-generation-has-gotten-better
- “MoMA Acquires Refik Anadol’s Unsupervised.” Artforum, October 11, 2023. https://www.artforum.com/news/moma-acquires-refik-anadols-unsupervised-517497/
- “MoMA Acquires Refik Anadol’s ‘Unsupervised,’ Other Digital Artworks.” ARTnews, October 10, 2023. https://www.artnews.com/art-news/artists/moma-acquires-refik-anadol-unsupervised-digital-art-nfts-1234681622/
- “Refik Anadol: Unsupervised.” Museum of Modern Art. https://www.moma.org/calendar/exhibitions/5535
- “AI ‘Slop’ Is Transforming Social Media — and a Backlash Is Brewing.” BBC/AOL. https://www.aol.com/articles/ai-slop-transforming-social-media-152212083.html
- Graphite. “AI Now Writes as Many Online Articles as Humans Do.” May 2026. https://graphite.io/five-percent/ai-now-writes-as-many-online-articles-as-humans-do
- “Exclusive: AI Writing Hasn’t Overwhelmed the Web Yet.” Axios, October 14, 2025. https://www.axios.com/2025/10/14/ai-generated-writing-humans
- Ahrefs. Study of AI-generated content across 900,000 web pages, April 2025.
- International Energy Agency. “Energy Demand from AI,” Energy and AI, April 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
- Pew Research Center. “What We Know About Energy Use at U.S. Data Centers Amid the AI Boom.” October 24, 2025. https://www.pewresearch.org/short-reads/2025/10/24/what-we-know-about-energy-use-at-us-data-centers-amid-the-ai-boom/
- Brookings Institution. “Global Energy Demands Within the AI Regulatory Landscape.” Updated April 2, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
- Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., and Gal, Y. “AI Models Collapse When Trained on Recursively Generated Data.” Nature 631 (2024): 755–759. https://www.nature.com/articles/s41586-024-07566-y
- Gerstgrasser, M., et al. “Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data.” 2024.
- Bertrand, Q., et al. “On the Stability of Iterative Retraining of Generative Models on Their Own Data.” 2023.
- Dohmatob, E., et al. “A Tale of Tails: Model Collapse as a Change of Scaling Laws.” 2024. https://arxiv.org/abs/2402.07043