The word arrived faster than the definition did. Merriam-Webster selected “slop” as its 2025 Word of the Year on 15 December 2025, which settled the spelling and the respectability and left everything else open. Academic papers on the subject open by noting that the term has so far resisted formal definition.
So the honest answer has two layers: a plain definition that most people would accept, and a set of live disagreements about where the line sits.
This page is about the concept: what the term refers to, and where its edges are argued over. For the phrase itself, see AI slop meaning; for the short dictionary-style entry, use the glossary definition of AI slop.
The short definition
The dictionary version is the place to start. Merriam-Webster defines slop as “digital content of low quality that is produced usually in quantity by means of artificial intelligence.”
Wikipedia's article on the subject frames it slightly more widely: digital content made with generative AI that is perceived as lacking in effort, quality or meaning, and usually produced in high volume as clickbait to gain advantage in the attention economy, or to earn money.
Both definitions contain three claims: the content was machine-generated, it is bad, and there is a lot of it. A third influential definition comes from the writing that popularised the word, and it drops the volume test. Simon Willison argued in May 2024 that slop was the right name for material that is “mindlessly generated and thrust upon someone who didn't ask for it”, drawing the analogy that not all promotional email is spam and not all AI output is slop.
That difference is worth noticing, because it changes who gets accused. The dictionary definition targets a production process. Willison's targets a social act: sending something you did not check to someone who now has to.
Where the word came from
Slop is an old word wearing new clothes. It meant soft mud in the 1700s and later food waste fed to animals, which is where the sneer comes from. It is what gets poured into a trough.
The AI sense appeared in the early 2020s, and commentary on the term traces its early use to the release of AI image generators in 2022, circulating as in-group slang among communities on 4chan, Hacker News and YouTube. It stayed in those corners for a while.
It became a term of art on 7 May 2024, when the writer and technologist known as deepfates observed that the word was arriving in the language the way “spam” had arrived, as the label for unwanted AI-generated content. Simon Willison picked it up the next day and gave it its shape, arguing that sharing unreviewed machine output with strangers is rude, and that slop was the ideal name for the anti-pattern.
Institutional recognition followed within a year. Oxford's word-of-the-year shortlist had already included slop, in a year where the publisher reported a 332 per cent increase in usage. Merriam-Webster then made it the 2025 Word of the Year, with the dictionary's president, Greg Barlow, telling the Associated Press that the word was “part of a transformative technology, AI” and “something that people have found fascinating, annoying and a little bit ridiculous.” The publisher's editors read the choice as a tonal shift: a word that is less fearful and more mocking, aimed at a year of talking cats and off-kilter advertising. The American Dialect Society picked the same word for the same year.
Three properties that survive scrutiny
Researchers who tried to pin the concept down came away without a clean boundary, and with three qualities that most examples share. A paper published in the ACM's journal on the topic identifies three:
Superficial competence. The output looks like it would have taken real skill to make. The prose is grammatical, the image is photorealistic, the layout is clean. What is missing is underneath, and that gap is the tell. A generated image can be beautiful and still contain a hand with six fingers or a horizon that bends the wrong way.
Asymmetric effort. It took far less effort to produce than the equivalent human work, and it asks more of the reader than it gave. This is the part people actually feel. The resentment in complaints about slop tracks the ratio between how long it took to make and how long it takes to check.
Mass producibility. Slop is built to exist at volume, distributed through systems that reward quantity. Some of it is personalised to a single reader, and it is still mass producible by design.
Those three properties are more useful than the dictionary entry, because they tell you what to look at. A well-written AI-assisted article that a human edited and stands behind fails all three tests, which is why calling it slop would be a category error.
What the label is actually doing
Slop is a word that describes and judges at the same time, which is why arguments about it heat up so fast.
Writing in npj Artificial Intelligence, one philosopher argues the term smuggles a verdict on the technology in alongside the verdict on the output. Optimists use it to quarantine bad output from AI's promise: that is not real AI, that is slop. Critics use it to indict the whole enterprise: this is what AI amounts to. Both moves skip the harder questions about what roles we want the technology to play, questions the word lets everyone avoid naming out loud.
The other reason the label lands is that it names a transaction. Cory Doctorow's phrase for the practice, reported in a Columbia University convening on AI slop and the information ecosystem, is non-consensual slopping: sending unreviewed machine output to strangers in order to make them do the unpaid labor of evaluating something the sender did not evaluate. Read that way, slop is less a genre than a way of behaving.
A research report on the term's normative status makes a related point about where the real objection sits. Complaining that a specific output is bad is a weak objection, because quality depends on purpose and standard. Complaining that the aggregate is restructuring the attention environment, raising the cost of finding anything worth reading, is the stronger one. The harm is in the flood rather than in any single drop.
Where the edges are contested
Three disagreements keep coming up.
Does it need volume? The dictionary definition says quantity. Legal scholars who have drafted a provisional definition argue the opposite: that a single unreviewed AI output can degrade a field, and that slop is better understood as a spectrum defined by negligible exertion, asymmetrical imposition and the damage done to the domain it enters. Under that reading, a filing full of invented case citations is slop whether or not anybody generated ten thousand of them.
Does it need AI at all? Everyone agrees there was slop before AI, and that people write carelessly without any model's help. Wikipedia's entry makes the point directly. The word as used today points at machine-generated material specifically, and stretching it to cover everything low-effort drains it of meaning.
Is the word doing more harm than good? That is now a live argument. The npj piece argues that slop-talk lets people feel they are discussing AI's social role without discussing it, and that the term's real function is to postpone the question of what we want. It concludes that the word is worth keeping, as long as people say which verdict they mean.
AI-assisted work is not the same thing
The most common mistake in this argument is treating any AI involvement as contamination. The definitions do not support that, and neither does the practice.
| Using AI as a tool | AI slop | |
|---|---|---|
| Effort spent | Substantial, spread across drafting, editing and checking | Close to none beyond the prompt |
| Review | Someone read it and is accountable for it | Nobody checked, or nobody has to |
| Purpose | Answer a specific question for a specific reader | Occupy a slot, rank, or farm attention |
| Surface | Reads like a person wrote it, rough edges included | Reads competent, says nothing specific |
| Who bears the cost | Shared | Pushed onto the reader |
The hard case is the workplace one. A study by Harvard Business Review with Stanford and BetterUp coined the word workslop for AI-generated content that looks good and lacks substance, and found that a substantial share of employees had received some, with each incident costing hours to resolve. That is the pattern worth watching: the output is not posted publicly, it is forwarded, and the bill arrives as somebody else's afternoon.
Why this matters
If slop were only annoying, the word would be a complaint. The reason it has attracted academic and policy attention is that the cost lands unevenly.
Readers pay in verification time. Creators pay in lost reach, because a flood of cheap material competes for the same slots. Institutions pay in trust, once enough plausible-looking fabricated material circulates that people stop believing anything without checking. And everyone pays in the slow degradation of the shared places where finding something good used to be easy, which is the concern the research on slop's normative status keeps returning to.
There is no tidy fix to hand over. The practical move is the one the definitions point at: judge the artifact in front of you rather than its provenance. Did someone check this? Does it tell me something specific? Would it exist if nobody were counting clicks? Those questions work on a generated paragraph and on a human one.