AI slop

AI slop is low-quality content made in quantity with generative AI and put in front of people who did not ask for it. The word arrived in 2024 and became a word of the year in 2025, because it named something everyone had started to see: the saint made of crustaceans in a Facebook feed, the foraging guide nobody wrote, the report citing papers that do not exist, the memo from a colleague that looks finished and is not. The cost of making it has fallen close to zero. The cost of judging it has barely fallen, and it falls on whoever receives it. That transfer is the real subject of this page. Slop is less a story about bad content than a story about where the work of judgement goes when production becomes free.

Where the word came from#

The word was in circulation as online slang before anyone wrote it down. On 7 May 2024 a user called deepfates posted that they were watching "slop" become "a term of art", and the next day the developer Simon Willison wrote a short post, Slop is the new name for unwanted AI-generated content. Willison credited deepfates and claimed no coinage. His test is still the cleanest one: content that is "mindlessly generated and thrust upon someone who didn't ask for it". He added a qualification that is often dropped: not all AI-generated content is slop, in the same way that not all promotional email is spam. Reports that the word was used earlier on 4chan, Hacker News and YouTube are repeated widely; I have not found a dated first use to support them.

The Guardian took up the term on 19 May 2024 and The New York Times on 11 June 2024, under the headline "First Came 'Spam.' Now, With A.I., We've Got 'Slop'". By the end of 2025 the dictionaries had followed. The Macquarie Dictionary chose "AI slop" as its 2025 word of the year on 25 November 2025, defining it as "low-quality content created by generative AI, often containing errors, and not requested by the user". Merriam-Webster chose "slop", defined as "digital content of low quality that is produced usually in quantity by means of artificial intelligence", and described the word's tone as "less fearful, more mocking". The American Dialect Society voted the same way in January 2026 and noted that slop had become a productive combining form, turning up in new words such as slopification.

Three features recur in every serious definition. Quantity: slop is made in volume, because volume is what the economics reward. Low quality: it does not meet the standard a reader would apply to the same thing from a person. And no request: the receiver did not ask for it. Remove any one and the word stops fitting. A single bad AI answer to a question you asked is an error. A thousand competent AI product descriptions are automation.

What it looks like#

Slop now has a recognisable shape in every medium. The examples below are the ones with a documented source behind them.

Images in the feed. Renee DiResta and Josh Goldstein studied 125 Facebook Pages that each posted at least fifty AI-generated images, and published the result in the HKS Misinformation Review in August 2024. The Pages averaged 146,681 followers. One AI image post was among Facebook's most viewed posts in the third quarter of 2023, with 40 million views. A "crab Jesus" image drew 209,000 reactions. The structural finding matters most: the share of feed views coming from posts by accounts people did not follow rose from 8 per cent in 2021 to 24 per cent in 2023, so the images reached people who had never chosen the Pages.

Disasters. After Hurricane Helene in October 2024, an image of a crying girl in a boat clutching a puppy was shared widely on X and Facebook as a photograph from North Carolina. Full Fact showed it was AI-generated: in the uncropped version the girl's hand has four fingers. Emergencies are where slop does most damage, because attention is high and checking is slow.

Books. In August 2023 404 Media reported AI-written mushroom foraging guides on Amazon. One listed author claimed a doctorate the named university had no record of. The New York Mycological Society warned that the content "can literally mean life or death". Amazon removed the titles it was told about and, the following month, limited self-publishers to three new titles a day and required them to tell Amazon when content was AI-generated.

News sites. NewsGuard's tracking centre listed 3,749 AI content farm news and information sites in 16 languages at its update of 23 June 2026. These sites publish AI-generated articles with little or no human oversight, usually to carry advertising.

Music. Deezer reported on 21 July 2026 that tracks it detected as fully AI-generated passed half of daily uploads at their peak in June 2026, while drawing only 1 to 3 per cent of streams, and that up to 85 per cent of the streams on such tracks in 2025 were fraudulent. Spotify said in September 2025 that it had removed more than 75 million "spammy" tracks in the previous twelve months. That figure is often reported as 75 million AI tracks; Spotify did not say so. In the summer of 2025 a band called The Velvet Sundown built an audience of hundreds of thousands of monthly listeners before describing itself, as The FADER reported, as "a synthetic music project guided by human creative direction".

Video. The editing company Kapwing set up a new YouTube account and classified the first 500 Shorts it was shown. 104 of them, 21 per cent, were AI slop. A further 165 were what Kapwing called brainrot, attention-bait that may or may not be machine-made. Its report, published in November 2025, estimated that the largest AI slop channel had 2.07 billion views.

Science. In February 2024 the journal Frontiers in Cell and Developmental Biology retracted a review because "concerns were raised regarding the nature of its AI-generated figures". One of them, as widely reported, showed a rat with impossible anatomy.

Work. Researchers at BetterUp Labs and the Stanford Social Media Lab named workslop in the Harvard Business Review in September 2025: "content that appears polished but lacks real substance, offloading cognitive labor onto coworkers". In their US survey, 40 per cent of workers had received some in the previous month.

The professions. Damien Charlotin's database of court decisions that note hallucinated material in filings stood at 2,207 cases on 23 October 2026, 877 of them involving lawyers and most of the rest self-represented litigants. In Australia, Deloitte refunded more than A$97,000 of a contract worth about A$440,000 after its report to the Department of Employment and Workplace Relations was found to contain a fabricated quotation from a court judgment and citations to papers that do not exist, as CFO Dive reported after the department confirmed the refund in October 2025. In May 2025 the news outlet NOTUS found that seven of the more than 500 studies cited in the Make America Healthy Again Commission's report, published by the White House, did not appear to exist.

Text slop has features that researchers have started to list. Chantal Shaib and colleagues interviewed experts and built a taxonomy of text slop in September 2025: low information density, poor relevance to the question, factual errors, repetition and templated structure, and a tone that is fluent, verbose and inflated. Readers say they notice the same things. Three paragraphs that could have been one sentence. Lists of three. Headings that announce rather than say. Confident claims with no source, or with a source that turns out not to exist. The evidence below suggests these tells are less reliable than readers think.

How it is made, and why it pays#

Slop comes from incentives that existed before generative AI and were waiting for production costs to fall. Every distribution system that pays for attention, measures volume or ranks by engagement will fill with whatever meets its measure most cheaply. Generative AI lowered that price to almost nothing.

On social platforms the mechanism is a payout. Jason Koebler of 404 Media traced where Facebook's AI slop comes from in August 2024: creators, many in India, Vietnam and the Philippines, learning a method from YouTube and Telegram tutorials and earning from Facebook's performance bonus programme, with American audiences targeted because advertising against them pays more. One tutorial promised "$100 for 1,000 likes". Meta said in October 2024, as TechCrunch reported, that creators had earned more than $2 billion on Facebook that year.

On the open web the mechanism is advertising. NewsGuard found 393 programmatic adverts from 141 major brands on 55 AI content farm sites in mid-2023, more than 90 per cent of them served by Google's ad systems. The brands had not chosen the sites. The ad market chose for them, by price. In music it is royalty fraud: cheap tracks streamed by bots to collect a share of the pool. In books it is volume on a marketplace that pays per sale and cannot read everything it lists.

The platforms now make it themselves. Google's AI Overviews told users in May 2024 to add glue to pizza sauce and to eat rocks; Google's head of Search, Liz Reid, explained that the answers drew on satirical and forum content where little reliable material existed. In September 2025 Meta launched Vibes, a feed of AI-generated short videos, and OpenAI launched a Sora app built around one; OpenAI's page now says the app has been unavailable since April 2026. And the answer engines now read the slop. NewsGuard's audit of August 2025 found the ten leading chatbots repeating false claims on news prompts 35 per cent of the time, up from 18 per cent a year earlier, and blamed web search polluted by sources including "AI-generated content farms that the models fail to distinguish from credible outlets".

At work the mechanism involves no fraud at all. A person under pressure to show output uses a tool that produces something that looks finished. The sender saves an hour. The receiver spends the hour, and more, finding out what is missing. BetterUp's figures put the rework at 1 hour 51 minutes per incident and the cost at $186 per employee per month. Workslop is made by ordinary people responding sensibly to how their work is measured. Of all the forms of slop, it is the one most organisations are paying for.

In every case the same asymmetry holds. Production is cheap and scalable. Judging is slow, skilled and human. Slop is what happens when the first is rewarded and the second is not paid for.

What slop does to judgement#

At this point the subject stops being about content and becomes about people. Slop is a tax on human judgement, collected in seven ways.

1. Spotting it is not a skill most people have. Sophie Nightingale and Hany Farid found that people identified AI-generated faces at 48 per cent accuracy, below chance, and rated them more trustworthy than real faces (PNAS, 2022). Maurice Jakesch and colleagues, across six experiments with 4,600 participants, found people unable to detect AI-written self-presentations because they relied on cues that do not work (PNAS, 2023). Brian Porter and Edouard Machery found readers did worse than chance at telling AI poetry from human poetry (Scientific Reports, 2024). "Train people to spot it" is a weak strategy, and detection tools have their own failure modes; see does AI detection work.

2. Even experts disagree on what it is. In the Shaib study, expert annotators asked to label texts as slop or not agreed little better than chance: their agreement scores ran from -0.15 to 0.29 on a scale where 0 is chance and 1 is full agreement. Language models asked to judge showed almost no agreement with the humans. Slop is a judgement that resists being written down as a rule. That matters for any organisation hoping to automate quality control on AI output with more AI.

3. The cost of judging moves to the receiver. Workslop shows the mechanism in an office. The sender makes the decision to send. The receiver does the checking, rewriting and chasing. Spread across an organisation, the checking becomes the limit: generation scales and verification does not. See the verification bottleneck.

4. It changes how people judge each other. In the BetterUp survey about half of the people who received workslop saw the sender as less creative, capable and reliable, 42 per cent as less trustworthy, and almost a third were less willing to work with them again. Slop spends the sender's reputation without telling them.

5. It teaches people to doubt true things. Cristian Vaccari and Andrew Chadwick found that deepfake video produced uncertainty more than deception, and that the uncertainty reduced trust in news on social media (Social Media + Society, 2020). Sacha Altay and Fabrizio Gilardi found that labelling a headline as AI-generated made people less likely to believe it even when it was true and human-written, because they assumed full automation (PNAS Nexus, 2024, 4,976 people in the US and UK). Kaylyn Jackson Schiff and colleagues showed that politicians can escape a text-based scandal by claiming it is fake, the "liar's dividend" (American Political Science Review, 2024). Where slop is everywhere, doubt becomes a default, and a default is not judgement. Judgement is telling which claims deserve trust. Blanket doubt is as lazy as blanket belief.

It reaches memory too. During the pandemic my family played with an app that swapped our faces onto film characters, and my youngest daughter still remembers seeing me as Tom Cruise. As I wrote in The Internet is Eating Itself, years from now she might remember the scene rather than the joke. Children growing up now will have more of those memories than any generation before them, and fewer ways to tell which are real.

6. It wears away the habit of checking. In a Microsoft Research survey of 319 knowledge workers, presented at CHI 2025, people who trusted generative AI more reported thinking less critically, and people more confident in themselves reported thinking more. It is self-reported, and it fits decades of research on automation bias. A day spent skimming plausible text trains the eye to skim. People who spend their working hours accepting or forwarding output they did not make lose the repetitions that built their judgement, the mechanism behind capability debt and the missing rungs.

7. It becomes the evidence decisions rest on. The Deloitte report was written to inform a government's welfare compliance system. The MAHA report was meant to shape health policy. The court filings ask judges to decide cases. In each, slop did not stay in a feed; it entered the record a decision was built on, and nobody downstream could see where. Model collapse is the same problem inside the machines: Ilia Shumailov and colleagues showed in Nature in July 2024 that training models indiscriminately on model-generated data causes irreversible defects, with the tails of the original distribution disappearing first. See what is model collapse. The tails are the unusual cases, the ones human judgement exists to handle.

Put together, the risk that can be measured sits where it sits for AI in general: in the handover of decisions, and of the inputs to decisions, to machines, and in what happens to human judgement afterwards. What an organisation controls is which decisions a machine may make or feed, who can stop each one, what people must remain able to do themselves, and how anyone would know if it had gone wrong.

Is all AI content slop?#

No, and a serious page on slop has to say so. Porter and Machery's readers did not only fail to spot the AI poems; they preferred them. The Velvet Sundown found real listeners. Plenty of AI-assisted writing, code and design is good, and plenty of human work is slop by any test. The word can be used as a sneer at anything machine-made, which empties it of meaning.

The definition survives this because it is about the relationship, not the tool. Content is slop when it was made in volume, falls below the standard the receiver needs, and was pushed at someone who did not ask for it. The same poem is a pleasure to a reader who chose it and slop in an inbox that did not. The same report is useful if someone checked it and slop if nobody did. Quality is judged against a purpose, and the receiver's purpose sets the standard.

Slop and enshittification#

The word slop sits beside an older one. Cory Doctorow coined "enshittification" in his Pluralistic newsletter in late 2022 and set out its stages in his January 2023 essay on TikTok: "Here is how platforms die: first, they are good to their users". Then they abuse their users to serve business customers, then abuse the business customers to claw back the value for themselves. The American Dialect Society made it the word of 2023, the Macquarie Dictionary the word of 2024, and Doctorow's book on it was published in October 2025.

The two words describe different things. Enshittification is a theory of platform decay driven by market power. Slop is a kind of content. Doctorow himself keeps them apart: in a November 2025 interview with The New World he said enshittification describes "something that's real, and AI is fake". The connection is still close. A platform in decline ranks for what keeps people scrolling and what its business customers pay for. Slop meets that ranking most cheaply, so a decaying platform fills with it, and the slop speeds the decay by driving away the people who made the platform worth using. When AI-written knock-off workbooks of his own book appeared on Amazon in October 2025, Doctorow wrote, as Futurism reported, that the "slop merchants" were using his book to "(further) enshittify Amazon".

The dead internet theory, a 2021 forum post that The Atlantic wrote up the same year, claimed that most online activity was already bots. As stated it was wrong. It lasted because it described a feeling that slop has since made more reasonable.

The backlash, and the fixes on offer#

The word itself is the first sign of the backlash. Merriam-Webster's editors called it mocking rather than fearful. People who had spent two years being told AI was salvation or catastrophe found a word for the thing in front of them, and it was a word for pig feed.

The platforms have moved, mostly by labelling and demonetising. YouTube renamed its "repetitious content" monetisation rule "inauthentic content" from 15 July 2025, to "better identify mass-produced and repetitious content". Pinterest added "AI modified" labels in April 2025 and gave users a control to see less generative AI content in some categories. Meta and TikTok label AI content, and Spotify introduced a spam filter and AI credit disclosures in September 2025. Amazon capped self-publishing volume. None of these removes the incentive. Each raises the cost of the cheapest slop at the margin.

Provenance is the technical answer. The C2PA standard attaches a tamper-evident record of where a file came from and how it was edited, and Google's SynthID embeds an invisible watermark in content its tools make. Both have limits their makers admit. C2PA says its metadata can be stripped and that provenance "alone cannot tell you whether the digital content is true". SynthID only finds SynthID. The law is following: Article 50 of the EU AI Act requires providers of generative AI to mark output in a machine-readable way from 2 August 2026, with a transitional period, which law firms including Mayer Brown report runs to 2 December 2026, for systems already on the market. Provenance tells you who made something. It cannot tell you whether it is any good, and that remains a human judgement.

Brands have met the backlash directly. Coca-Cola's AI remake of its "Holidays Are Coming" advert was called "soulless" in coverage collected by Social Samosa in November 2024, and AdNews described its 2025 version as "last year's rehash, with better fidelity". John Oliver gave the subject a main segment of Last Week Tonight in June 2025.

The public mood is measurable. Pew Research Center found in 2025 that 76 per cent of Americans think it extremely or very important to be able to tell whether content was made by AI or a person, and 53 per cent are not confident they can. The Reuters Institute's Digital News Report 2026, across 48 markets, found concern about what is real or fake online at 62 per cent, up four points, and linked the rise to deepfakes, AI slop and misinformation. The backlash carries its own cost, covered above: the more people expect slop, the less they believe anything.

June to October 2026: what the consultancies found#

The large consultancies do not use the word slop, but their surveys describe its effects inside companies. BCG's study of deskilling, based on 70 senior executives, found almost 90 per cent citing "overreliance on AI outputs without stress testing or challenge" and judged that "judgment and decision making carries the highest de-skilling risk" of any skill. Its description of the mechanism is a description of slop at work: "Teams start treating AI-generated work as 'good enough' and skip the critical thinking step entirely." Only one company in ten had an organisation-wide strategy to deal with it.

PwC's Hopes and Fears survey (29 September), of 49,364 workers in 48 countries, found that 31 per cent of daily AI users name accuracy or quality as their top barrier to using it. The people closest to the output are the ones who see the slop.

What an organisation can do#

I read slop as the purest example of drift. Nobody in any organisation decided to fill the shared drive with polished documents nobody can rely on. It happened through a thousand small choices, each sensible on its own: use the tool, ship the draft, hit the volume target. Design asks the question first: what standard does this kind of output have to meet, and who checks it before it reaches someone else? The responses that work move the cost of judging back to the person who chose to generate, and they are decisions about rules before tools.

Set the standard per kind of output. A board paper, a client deliverable, a customer reply and an internal note need different standards. Write them down. Slop thrives where the standard is unstated.

Make the sender the first checker. Whoever sends AI-assisted work vouches for it as if they wrote it. The working rule fits on one line: if you would not sign it, do not send it. It puts the hour back where it was saved.

Use the four receipts. Hand, Head, Hours, Heart read as a test for slop. Hand: who stands behind this and will answer for it? Head: does anyone learn when this kind of output is wrong? With slop nobody does, so the same errors recur. Hours: where did the saved time go? Workslop is saved time moved onto someone else's desk. Heart: who bears the consequence? With slop it is the reader, the forager, the litigant, the colleague.

Measure rework, not volume. If the dashboard counts documents, messages or tickets closed, it is paying for slop. Count what had to be redone, and how long checking took.

Reward rejection. A reviewer who sends work back is doing the job. If nobody ever rejects anything, nobody is looking.

Disclose what was checked, and do not label everything. Labels reduce trust in true content too, so a blanket "AI was used" stamp spends credibility without protecting anyone. Disclosure works best when it says what a human checked.

Protect the repetitions. Juniors need to do by hand work the tool could produce, often enough to learn what good looks like. Without that, the organisation will one day have nobody able to tell slop from substance.

For a board the single question is: what share of the AI-assisted work that reaches customers, regulators or this board was checked by someone who could have done it themselves, and how would we know?

What this does not show#

No study has measured the total effect of slop on human judgement. The evidence here is a set of parts: detection experiments in laboratory conditions, surveys of attitudes, and counts by platforms and trackers. The workslop figures come from a single survey by a company that sells coaching; the economist Jason Collins has pointed out that it has no counterfactual and does not measure the value of the AI output, so it cannot show that productivity is being destroyed, and its published figures vary slightly between sources. Deezer's and Spotify's figures are each company's own and are not audited; Spotify's count is of spam, not of AI. Kapwing's feed came from one new account, and recommendations depend on history. Charlotin's database counts court decisions noting hallucinated material, not sanctions, and depends on what is reported to it. The dictionary choices show attention, not harm. And the claim that slop erodes trust in true content rests on studies of deepfakes and labels, not of slop as such. The direction is consistent across all of them. The size is not yet known.