“AI slop”一词如何掩盖了我们真正该进行的 AI 讨论

The term “AI slop” hides conversations we need to have. When we call something “AI slop,” we’re often doing more than complaining about one output. Optimists use the term to quarantine failures away from AI’s promise; critics use it to indict the whole technology. Slop-talk thus lets us feel like we’re discussing important social issues without actually doing it. I argue that we should have those discussions openly.

This spring, arXiv started banning authors whose submissions bear the marks of unchecked AI use, the latest of several key institutions to move against “AI slop^{1, 2}.” The reactions make sense, since cheap synthetic text now floods not only research repositories but also traditional and social media, art, advertisements, and even legal briefs. But what worries me is less the slop itself than how we’re using the term. Sometimes we use the term just to complain about a bad AI output, but often we use it to smuggle in one of two much broader views about AI’s proper role in society, taking sides in an important social debate while seeming only to sneer at a single output. The move is slippery, often unintentional, and it allows us to avoid important discussions we should be having.

One attitude “AI slop” can express is an attitude of quarantine. In this use, the slop is seen as unfortunate and not representative. To call it “slop” is to try to cordon it off from the good stuff that AI can produce, the useful translation, the genuinely helpful code, or the new mathematical discovery³. This is how Sam Altman used the term when OpenAI launched Sora 2. Creativity, he wrote, could be headed for “a Cambrian explosion,” while the “RL-optimized slop feed” was a degenerate case they put great care into avoiding⁴. It is also the attitude Google tried to take when its AI search suggested using glue to keep cheese on pizza. The company said such results came from uncommon queries⁵. The embarrassing case was sealed off so the rollout could go on.

A second attitude is one of contagion. In this use, the slop is not a defective specimen but a revelation. The bad output shows us what AI really is, a technology inherently derivative, careless, and hostile to the culture it feeds on. The critic says, look at the synthetic children’s books, the student essays no one wrote, and the apology emails no one stands behind. Those aren’t unfortunate by-products but perfect representatives of what AI can do for us. This is the move Madsen and Puyt make⁶, treating the flood of derivative, recycled output not as a defect to be cleaned up but as what AI amounts to. And it is how the Atlantic’s Charlie Warzel answered Sora 2’s launch three weeks later. “AI slop is winning,” the piece announced, and what Altman was describing was “a world of creativity without craft⁷.”

There’s something commendable in both attitudes. The quarantine attitude rightly reminds us that not every AI system or output is the same. The contagion attitude reminds us to be on the lookout for slop, since many incentives reward cheap, low-value production. My concern is not with either attitude but with how the term lets us avoid saying which way we intend it. In the optimist’s mind it is only slop, with the emphasis falling on the “slop”; in the critic’s mind it is all slop, with the emphasis falling on the “AI.” And importantly, two people on opposite sides of the bigger social questions can walk away from a discussion thinking they agree because they’ve settled that some output is AI slop, when really the larger disagreement was hidden by their differing intentions in using the term.

The trick is that “AI slop” presents itself as a verdict on an output while implicitly suggesting a verdict on the technology^{8, 9}, the same move behind the charge that a text “sounds like AI¹⁰.” Beyond calling an output cheap and low-value, it invites disgust and suggests we recoil (philosophers of language would call it a “thick” term, one that describes and evaluates at once). That recoil can mean two different things. In the optimist’s mouth, it’s a push to quarantine the output from AI’s valuable work; in the critic’s mouth, it’s a push to scuttle the whole enterprise. But disgust at a particular output is one thing, and a verdict on the whole category is another, and the term lets us slide from the first to the second without noticing.

We can see that this slide is a cheat by recalling that there was slop long before there was AI¹¹. There were bad novels, formulaic films, empty lectures, and academic committee reports that said nothing at length. But no one thought that bad novels settled the worth of literature or that a worthless report settled whether academia should keep committees. This point cuts both ways. The slop coming out of our machines doesn’t show that AI is worthless any more than its good outputs show that it’s good on the whole. So when discussions of “AI slop” seem to bear on this, it must be some background view about AI’s proper role that’s really doing the work. It’s that background view we really need to be discussing, and when we get there via “AI slop,” it’s typically carried in quietly and left undefended.

None of this is to say that there’s nothing unique about AI slop. AI does change the scale, speed, and cost of producing low-value outputs, and makes some of them harder to spot, more profitable to spread, and more apt to poison the data future models learn from¹². Those are real worries, but they are worries about how the technology is designed, owned, deployed, and used. They are not proof of its overall value or worthlessness.

The natural contrast with slop is quality, and often quality is all that is at issue. Ask whether a particular manuscript is slop and you are usually just asking whether it is any good, and that may be all an editor needs to know. But quality is not a stance on how we should live. The better AI gets, the more tempting it is to think the quality settles something bigger. It doesn’t. A machine can write a flawless condolence note, and whether such a note ought to be machine-written is left exactly where it was. Quality answers the easy question and stays silent on the hard one.

So instead of focusing on AI slop, we should ask directly the questions slop-talk lets us postpone by seeming to settle them, questions about the kind of society we should be building and what roles AI should play in it. When someone is sick, lonely, or frightened, do we want a real person in the room, or will a bot do?¹³ Do we want our children taught and shaped mostly by people, or are we content to hand that to machines? When a court weighs whom to believe, or an employer whom to hire, do we want a person behind the call, or can a machine sort it?¹⁴ Do we want the systems that increasingly run our lives steered by a few private companies, or held more widely accountable? And do we want a shared world we can make sense of together, or one so flooded with synthetic stuff that no one can tell what is real?¹⁵ None of these can be answered by asking whether AI produces quality or slop. They are about the communal lives we want, and when we reach for “AI slop,” we’re sloppily pushing each other on exactly these questions without ever confronting them out loud.

I am not saying we should ban the term “AI slop.” Sometimes we want to name what should be shovelled into the trough, and that term will do. But we should stop letting it stand in for the tough collective discussions we should be having. There are small ways to start. When we reach for “AI slop,” we can try to be explicit about whether we mean to give a verdict on an output or a verdict on the technology. And our institutions can say, when they ban slop, what they are and are not deciding.

Consider arXiv again. To its credit, the rule it implemented is narrower than the headlines suggested. Announcing it, the chair of arXiv’s computer-science section clarified that the rule is only about the quality of the submitted piece, writing that when a submission shows “incontrovertible evidence that the authors did not check the results of LLM generation,” the upshot is that “we can’t trust anything in the paper¹⁶.” But to my bigger point, the ban was read as a stand anyway. One commentator cheered it as a blow against “the slop machine,”¹⁷ while another countered that AI-generated text “is not a problem in itself” and that the answer to the flood of slop may be “more AI, not less¹⁸.” So while a policy can keep to the quality question, the term drags the bigger issues in, and it gets used in the two different ways without the difference ever being marked. That is exactly why saying what you do and do not mean by the term matters, and when others reach for the term, we should ask what they mean as well. This is how we’ll make progress on tackling those bigger societal questions about how much of our shared lives we want to turn over to the technology and how much we want to keep for ourselves.

Data availability

No datasets were generated or analysed during the current study.

References

1. Chawla, D. S. Researchers who use hallucinated references to face arXiv ban. Nature 653, 988–989 (2026).

   Article  Google Scholar 

2. Giray, L., Sevnarayan, K., Maphoto, K. B. & Wider, W. AI slop in academic publishing: history, characteristics, manifestations, causes, and mitigation strategies. Internet Ref. Serv. Q. 30, 221–244 (2026).

   Google Scholar 

3. Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature 625, 468–475 (2024).

   Article  Google Scholar 

4. Altman, S. Sora 2. https://blog.samaltman.com/sora-2 (2025).

5. Reid, E. AI Overviews: about last week. Google https://blog.google/products/search/ai-overviews-update-may-2024/ (2024).

6. Madsen, D. Ø. & Puyt, R. W. When AI turns culture into slop. AI Soc. 41, 1287–1288 (2026).

   Article  Google Scholar 

7. Warzel, C. A tool that crushes creativity. Atlantic https://www.theatlantic.com/technology/2025/10/ai-slop-winning/684630/ (2025).

8. Shanahan, M. Talking about large language models. Commun. ACM 67, 68–79 (2024).

   Article  Google Scholar 

9. Foster, J. & Ichikawa, J. Normative inference tickets. Episteme 22, 47–73 (2025).

   Article  Google Scholar 

10. Ciriello, R. On the troubling rise of generative AI suspicion in academic publishing. Nat. Mach. Intell. 8, 136–137 (2026).

    Article  Google Scholar 

11. Kommers, C. et al. Why slop matters. ACM AI Lett. 1, 1–6 (2026).

    Article  Google Scholar 

12. Shumailov, I. et al. AI models collapse when trained on recursively generated data. Nature 631, 755–759 (2024).

    Article  Google Scholar 

13. Turkle, S. Alone Together: Why We Expect More from Technology and Less from Each Other (Basic Books, 2011).

14. Huq, A. Z. A right to a human decision. Va. Law Rev. 106, 611–688 (2020).

    Google Scholar 

15. Rini, R. Deepfakes and the epistemic backstop. Philos. Impr. 20, 1–16 (2020).

    Google Scholar 

16. Dietterich, T. G. Post on X. https://x.com/tdietterich/status/2055000956144935055 (2026).

17. Gerard, D. arXiv bans academic authors for AI slop papers. Pivot to AI https://pivot-to-ai.com/2026/05/15/arxiv-bans-academic-authors-for-ai-slop-papers/ (2026).

18. Kovanovic, V. A key science publishing platform is cracking down on AI slop. The Conversation https://theconversation.com/a-key-science-publishing-platform-is-cracking-down-on-ai-slop-283136 (2026).

Download references

Acknowledgements

I am grateful to Luca Garzino Demo, Jennifer Morton, Nate Sharadin, and Frank Saunders Jr. for early conversations about the idea that inspired this essay. The author used Claude Opus 4.8 and Fable 5 for editorial suggestions, revising wording, and identifying references. The argument, all substantive choices, and the final text are the author’s own, and the author takes full responsibility for the content. No funding was received for this research.

Author information

Authors and Affiliations

1. Department of Philosophy, School of Arts and Sciences, and Legal Studies & Business Ethics, The Wharton School, University of Pennsylvania, Philadelphia, PA, USA

   Daniel J. Singer

Authors

1. Daniel J. Singer

Contributions

D.J.S. was responsible for all aspects of the production of the essay.

Corresponding author

Correspondence to Daniel J. Singer.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Singer, D.J. ‘AI slop’ hides the conversations we should be having. npj Artif. Intell. 2, 68 (2026). https://doi.org/10.1038/s44387-026-00161-8

Download citation

• Received: 06 July 2026

• Accepted: 05 September 2026

• Published: 24 September 2026

• Version of record: 24 September 2026

• DOI: https://doi.org/10.1038/s44387-026-00161-8