What is AI slop?

"AI slop" is a term for the large volumes of low-effort content produced with Generative AI and pushed at people who did not ask for it: fake feel-good stories with machine-made images, absurd short videos, and jumbled text across social media and search results. It is cheap to produce, so it arrives in bulk and gets measured in billions of views.

Where the word came from

The term crystallised in May 2024. An X post by the account @deepfates proposed that "slop" would become to unwanted AI content what "spam" was to unwanted mail, and the software developer Simon Willison wrote a widely read post agreeing. Willison's definition is still a good starting point:

Not all promotional content is spam, and not all AI-generated content is slop. But if it's mindlessly generated and thrust upon someone who didn't ask for it, slop is the perfect term for it.

Willison drew a personal rule from it: attach your name to what you publish and stand behind it; do not publish slop. He also flagged that the boundary is fuzzy — the day after, he proposed "slom" for spam that happens to be made with AI, the overlap of the two categories. The thing itself is older than the word: Forbes notes that Facebook de-prioritised news and brand content in 2018, laying the algorithmic groundwork later filled by machine junk.

Shrimp Jesus

Examples appeared immediately. In April 2024 Forbes explained "Shrimp Jesus": AI images of Jesus fused with shrimp and crabs, flooding Facebook. The accounts behind them buy follower bots to inflate reach (a Stanford Internet Observatory study), post glitchy output as engagement bait, and steer real users toward content farms. A 404 Media investigation, cited by Forbes, found the images were pushed onto timelines by Facebook's "Suggested for You" engine, launched in 2022. Older users were the easiest marks: distorted hands and twisted limbs are easy to miss, and while younger users drifted to TikTok and elsewhere, older people stayed on Facebook — what Forbes called "the Dead Internet Theory coming to life".

How big it is

By early 2026 the scale was being measured. Kapwing research relayed by the BBC found that 20% of the content a freshly opened YouTube account was shown was "low-quality AI video", including 104 of the first 500 Shorts clips. The most-viewed such channel, Bandar Apna Dost, had 2.07 billion views and estimated annual earnings of about $4m (£2.9m). A single fake feel-good image on Facebook drew close to a million reactions. YouTube's chief executive, Neal Mohan, wrote that in a single month (December) more than a million channels used the platform's AI tools. The figures are journalists' relays of third-party research, so treat them as order-of-magnitude.

What the research says

A small academic literature on slop is forming, and much of it is properly cautious. Screening biomedical-science teaching videos on YouTube and TikTok in early 2025, a study in JMIR Medical Education judged 57 of 1,082 (5.3%) to be probably AI-generated and low quality, concluded slop was "not especially prevalent" at the time, and — on a small sample — found its views, likes and comments comparable with those of ordinary videos (Jones et al.). The gap between that and the Kapwing figure above comes from measuring different things: topic searches rather than a recommendation feed, and early 2025 rather than late 2025.

Kommers and colleagues, writing in ACM AI Letters in 2026, make the steelman case that slop deserves study rather than dismissal. It meets a demand for more content than humans can supply and, they argue, often carries its own aesthetic value; they define it by three features of a family resemblance: superficial competence (a veneer of quality over a lack of substance), asymmetric effort (far less work than a non-AI equivalent), and mass producibility.

Other strands push the subject further. An AoIR study documents AI-generated videos aimed at children on YouTube, together with tutorials on profiting from low-effort AI video (Mintz and Bozzi). Slop has reached scholarship itself: Nature reported that 21% of the manuscript reviews submitted to the 2026 ICLR machine-learning conference were found to be AI-generated (Naddaf). And two working papers — not peer-reviewed — frame the economics: Madsen and Puyt catalogue "generative waste" under seven "V" headings, and Tullis reads flooded text markets through Akerlof's "market for lemons", in which buyers who can no longer tell quality apart stop paying for it.

Why it spreads, and what platforms do

The economics reward it: engagement feeds advertising, and, in the BBC's phrase, "all engagement is good engagement". Meta's Mark Zuckerberg told investors that social media had entered a "third phase" in which AI adds "another huge corpus of content"; when the BBC asked whether Meta was reining in slop, the company pointed to an earnings call that mentioned no clampdown. Mohan acknowledged "low-quality content, aka AI slop" and said YouTube was working to remove "low quality, repetitive content", but he ruled out judging what should flourish. Pinterest now lets users opt out of AI content, but the system relies on detecting AI output or users admitting their posts are generated, and complaints continue, the BBC reports.

What it costs

The noise has a price. The engineer Robin Moffatt wrote in May 2026 that slop is "driving up the noise, and making the signal more and more difficult to discern", strangling online communities "like bindweed" until members leave and the spaces wither or shrink to bots talking to bots. He invoked "the asymmetry of bullshit": refuting bad content costs an order of magnitude more effort than producing it, so moderation becomes the community's unpaid tax. Researchers quoted by the BBC link the flood to "brain rot" — content people know is fake being consumed anyway. Slop is also the mild end of a spectrum: the BBC notes that beyond it lie AI-made pornographic deepfakes of real women and children, and fake war footage shared in the wake of the US raid on Venezuela, all against a backdrop of platforms cutting their content-moderation teams.

The word is blunt

The term is easy to misuse, and the edge cuts in both directions. Willison stressed that not all AI-generated content is slop. The BBC notes that presentation matters: obvious jokes get tolerated, content meant to deceive draws anger; its academic sources separate harmless absurdity (fish with shoes, gorillas lifting weights) from deceptive material. Moffatt reports the opposite complaint: some people "brand anything written about AI as 'AI Slop', even if not written by AI". Measurement backs him: a study of 25 million comments on Hacker News and Reddit (2023–2026) found that accusations of "AI slop" rose more than tenfold, now make up 94% of pejorative labelling, and — in a matched test — were not predicted by the features that actually separate AI prose from human writing. The accusation now works as social gatekeeping, not detection (Miklian and Katsos, 2026). His line is "built with AI, not by AI" — the tool is not the problem, care and intent are. The word has also been pointed inward: Amazon employees mock their own AI coding tool's output as "slop", 404 Media reported. Whether something counts as slop is a judgement about quality, intent, and whether anyone asked for it, not about the model that made it — and different people draw the line in different places.

Théodore Cazals, a 20-year-old student from Paris, ran the X account "Insane AI Slop" to catalogue the worst examples; it passed 130,000 followers before he largely gave up, telling the BBC that "the battle is lost". His definition of slop: "fake, unconvincing videos and pictures, made quickly".

The site you are reading is called Sloppy Research. The name is a joke; the articles are not. Sources are listed below so you can check.

Sources

Reporting and commentary

Research

Slop is low-effort AI content pushed at people who did not ask for it. The word caught on in 2024; the flood it names has since grown.