What is AI slop?
Engineering culture and software practice
AI slop is low-value AI-generated material produced in bulk because producing it is nearly free: text, images, code, music, reviews and research that is fluent, plausible and not worth the reader's time. The term spread through 2024, championed by developer Simon Willison in a May 2024 post, and was named 2025 word of the year by Merriam-Webster and the American Dialect Society. Slop is a value and volume problem, not a truth problem: it can be entirely accurate and still worthless.
What this means
Slop is the flood of generated stuff that fills a channel because it costs almost nothing to make: articles, images, reviews, applications, even research, all fluent and plausible and none of it worth your time. The key feature is not that it is false. Slop can be perfectly accurate and still be slop, because the problem is volume and low value, not error.
The word caught on by analogy with spam. Just as spam named unwanted email regardless of whether any given message was a lie, slop names unwanted, mindlessly generated content thrust on people who did not ask for it. That framing is why the term stuck: it gave people a single word for something they were all suddenly experiencing.
There is a workplace cousin. When the burden of fixing low-value AI output falls on a colleague rather than a stranger, it has acquired its own name, workslop, coined in 2025 to describe AI-generated work that looks done but pushes the real effort downstream onto whoever receives it.
Why it matters
Slop matters because it moves a cost rather than removing one. The cost of producing content has collapsed, but the cost of evaluating it has not. Every fluent, low-value item still has to be read, checked and judged by a human, and that time lands on inboxes, application piles, review queues, procurement processes and public consultations. The producer saves minutes; the readers, collectively, lose hours.
Inside an organisation the effect is corrosive. When colleagues send AI-generated work that looks polished but lacks substance, the receiver has to decode it, supply the missing context, or redo it, and trust erodes. Research on the workplace version found this is common and expensive, and that it damages how people see the sender, not just their time.
There is also a risk of producing slop yourself. An organisation that floods its own customers, applicants or regulators with generated filler pays in reputation and relationships. And there is a slower, collective harm: as generated material fills the web, it degrades the shared pool that search and future models draw on. For a leader, the point is that volume, not falsehood, is the mechanism, so the usual defences against inaccuracy do not address it.
How it works
Where the term came from
Slop as a label for low-grade AI output circulated informally from around 2022, after image generators became widely available. It broke into the mainstream in May 2024 when the British developer Simon Willison championed it in a blog post dated 8 May 2024, arguing it should become the standard word for unwanted AI content the way spam became the word for unwanted email. Willison was explicit that he was amplifying, not coining, the term, crediting a post by the account deepfates, and that not all AI content is slop, only that which is mindlessly generated and pushed on someone who did not ask for it.
The evidence that the term has stuck is unusually strong for something so recent. Merriam-Webster named slop its 2025 word of the year, defining it as digital content of low quality produced usually in quantity by artificial intelligence. The American Dialect Society also selected slop as its 2025 word of the year, in a vote held in January 2026, noting that AI slop had been a nominee in 2024 but that by 2025 slop could stand alone, and recognising it as a productive combining form. Sustained mainstream and business press use accompanied both. This is a fad-versus-fixture question the reader can now answer: two established dictionaries and a scholarly society is the opposite of a passing meme.
The economics
The engine of slop is a simple asymmetry. Generating a plausible article, image or review now costs almost nothing, while evaluating whether it is any good still costs a human's attention. When production is nearly free and evaluation is not, the rational move for anyone chasing clicks, applications or apparent output is to produce more, and the cost of sorting it lands on everyone downstream. That is the whole mechanism, and it explains why slop can be accurate and still harmful: the burden is the volume, not the errors.
The varieties
Slop appears as content on social and search, as filler articles built to attract clicks, as generated code that compiles but adds little, as padded research, and as mass-produced reviews and job applications. The workplace variety, workslop, was described in a Harvard Business Review article on 22 September 2025 by researchers at BetterUp Labs and Stanford's Social Media Lab, who defined it as AI-generated work that masquerades as good work but lacks the substance to advance a task. Their survey of 1,150 full-time United States desk workers found forty per cent had received workslop in the previous month, with each instance taking on average one hour and fifty-six minutes to sort out.
Why it is not the same as hallucination
This is the distinction the term most needs. An AI hallucination is content that is false: the model asserts something untrue. Slop is orthogonal to truth. A slop article can be entirely accurate and still be worthless because it is low-value filler produced in bulk. You can have accurate slop and false non-slop. Confusing the two leads organisations to fight slop with fact-checking, which misses the point, because the problem is that the material should not have been produced or sent at all.
The downstream effect on the commons
There is a longer-term worry about the shared pool of information. As generated material accumulates online, it degrades what search returns and what future models train on. The peer-reviewed study most cited here is by Ilia Shumailov and colleagues in Nature in 2024, which found that training models indiscriminately on model-generated data causes irreversible defects, a phenomenon termed model collapse. Nature published an author correction to that paper in 2025, so the 2024 finding should be read together with its correction rather than on its own.
Examples
A charity posts a modest job opening and receives hundreds of applications, many of them fluent, generic and clearly generated in bulk. Each still has to be read. The applicants spent seconds; the hiring panel loses days, and good candidates are harder to find in the noise. Nothing here is false; it is simply slop, and the cost has moved onto the charity.
A professional services firm finds an analyst circulating internal memos that read smoothly but fall apart on a close read, leaving colleagues to reconstruct the missing reasoning. This is workslop: the effort was not saved, only shifted onto the receivers, and trust in the sender quietly declines. The firm responds with a disclosure norm and a rule that work is reviewed before it is sent.
A council runs a public consultation and receives a wave of near-identical generated responses. Officers must now separate genuine public views from mass-produced filler, which is slow and undermines the consultation's purpose. The council adds a verification step and asks about provenance, because the volume, not any single false claim, is the problem.
Common misunderstandings
The first and most important misconception draws the line against AI hallucination, the closest existing idea. Slop is not the same as a hallucination. Unlike a hallucination, which is content that is false, slop can be entirely accurate and still worthless, because it is a value and volume problem rather than a truth problem. Fighting slop with fact-checking therefore misses the target.
The second is that slop means all AI-generated content. It does not. The coiner was explicit that not all AI content is slop, just as not all promotional email is spam. Slop is specifically the mindless, low-value, unsolicited kind. Plenty of AI-assisted work is good.
The third is that slop is just spam under a new name. It overlaps but differs: spam is unwanted messaging, often commercial; slop is unwanted generated content across media, defined by low value and bulk rather than by a sales motive.
The fourth is that the danger is that slop is untrue. The danger is that it is worthless and voluminous, so it consumes the scarce resource of human attention regardless of accuracy. Volume is the mechanism.
The fifth is that slop is a property of the tool. It is not. Slop is a description of effort withheld, of output generated carelessly and pushed on others, not a verdict on AI itself. The same tool, used with care and review, produces work that is not slop at all.
Risks and boundaries
This is a recent term, and this article was written in September 2026. Its meaning is still moving: slop began as a label for AI content specifically and, by the American Dialect Society's 2025 vote, had broadened into a combining form for anything of little value made in bulk. Attribution is genuinely contested; Willison is credited as an early mainstream champion but has said the word was in use before him and credited others, so the honest position is that it was popularised rather than cleanly coined.
The term is misapplied when it is used to dismiss all AI-assisted work, or as a generic insult for anything a person dislikes. It does not cover accurate, carefully made AI-assisted work, and it is not a synonym for error. The evidence base is uneven: the dictionary and word-of-the-year listings are solid and verifiable, the workplace survey is a single self-reported study, and the model-collapse research, while peer reviewed in Nature, carries a 2025 correction and describes a risk under specific training conditions rather than a certainty. The fair counter-point stands: slop describes effort withheld, not a property of the technology.
What to do next
First, name the real cost. Treat slop as an attention tax, not an accuracy problem. The question to ask of any AI-assisted output is not only whether it is true but whether it is worth the receiver's time and whether it should have been produced at all.
Second, set disclosure and review norms internally. A simple expectation that AI-assisted work is reviewed by a person before it is sent, and that its use is disclosed where it matters, prevents most workslop and protects trust between colleagues.
Third, measure rework. If you want evidence of a slop problem, track how much time people spend decoding, correcting or redoing work they receive. That number, not output volume, tells you whether AI is helping or simply shifting cost.
Fourth, guard your external channels. Where you receive applications, submissions or consultation responses, add verification and ask about provenance, so mass-produced filler does not crowd out genuine engagement. Where you send, make sure you are not the source of someone else's slop.
Fifth, keep the counter-point in view. Do not respond by banning the tools or treating all AI-assisted work as suspect. The distinction that matters is care: reviewed, valuable work is not slop, and saying so protects the people using AI well.
FAQs
Who coined AI slop and when?
The word circulated from around 2022 and was popularised in mainstream use by developer Simon Willison in a post dated 8 May 2024. He said he was amplifying an existing term, not coining it, so it is best described as popularised rather than coined.
Is slop the same as an AI hallucination?
No, and this is the key distinction. A hallucination is false content. Slop can be entirely accurate and still worthless, because it is a problem of low value and high volume, not of truth.
What evidence is there that the term has stuck?
Merriam-Webster named slop its 2025 word of the year, and the American Dialect Society did the same in a vote held in January 2026. Two established dictionaries and a scholarly society mark it as a fixture, not a fad.
What is workslop?
The workplace form of slop: AI-generated work that looks done but lacks substance, shifting the effort onto colleagues. It was described in a September 2025 Harvard Business Review article by BetterUp Labs and Stanford researchers.
How costly is workslop in practice?
In the BetterUp and Stanford survey of 1,150 United States desk workers, forty per cent had received workslop in the previous month, and each instance took on average nearly two hours to sort out. It is one self-reported study, so treat the figure as indicative.
Does slop mean all AI content is bad?
No. The coiner was clear that not all AI content is slop, just as not all promotional email is spam. Slop is the mindless, low-value, unsolicited kind. Careful, reviewed AI-assisted work is not slop.
How does slop harm search and future AI?
As generated material fills the web, it degrades the shared pool that search and models draw on. A 2024 Nature study by Shumailov and colleagues described model collapse from training on generated data, with an Author Correction published in 2025.