AI slop

AI slop, often shortened to slop, is low-quality digital content made in large quantities with generative artificial intelligence, usually without being requested by the people who end up seeing it. Merriam-Webster defines it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence."^[1] The term is modeled on spam: just as spam names unwanted email rather than all promotional email, slop names unwanted AI output rather than all AI-generated content.^[2]

The word spread widely in 2024 as AI-generated images, text, and video filled social media feeds. By early 2026 it had been chosen as word of the year by Merriam-Webster, the Macquarie Dictionary, and the American Dialect Society.^[1][3][4] Well-documented examples include engagement-bait images on Facebook, AI-generated channels on YouTube, fabricated disaster photos, AI-assisted work documents that shift effort onto colleagues, and invalid security reports sent to open-source software projects.

Definitions

Definitions of slop share three features: the content is low in quality, it is produced in bulk at little cost, and it reaches people who did not ask for it.

The definitions differ in emphasis. The dictionary definitions focus on quality and volume, Willison's on the fact that the content is unwanted, and Kapwing's on the motive — farming views or influencing politics.

Origin and spread of the term

Before its AI sense, slop was associated with pig feed and worthless material; the American Dialect Society described the word as having "moved from the pig sty, to the algorithm."^[4]

On May 8, 2024, the British programmer Simon Willison wrote a blog post, "Slop is the new name for unwanted AI-generated content", which endorsed an observation by the social media user @deepfates that slop was becoming a term of art in the way spam had for unwanted email.^[2] Willison noted that "not all AI-generated content is slop" and proposed "don't publish slop" as a basic rule for personal use of AI: sharing unreviewed machine output with other people, he wrote, is rude.^[2]

The term then moved into general use:

• 2024: "AI slop" was a nominee in the American Dialect Society's word-of-the-year vote for 2024.^[4]
• November 2025: The Macquarie Dictionary named "AI slop" its word of the year for 2025.^[3]
• December 15, 2025: Merriam-Webster named "slop" its 2025 word of the year, describing it as "all that stuff dumped on our screens, captured in just four letters."^[1]
• January 9, 2026: The American Dialect Society chose "slop" as its word of the year for 2025. The society noted that by 2025 the word no longer needed "AI" in front of it, because the AI context was usually understood, and that it had begun forming new compounds such as slopification, sloppunk, and friend slop.^[4]

A related coinage, workslop, was introduced in a September 2025 Harvard Business Review article for AI-generated work documents (see the workplace section below).^[6]

Examples

Social media engagement bait

Facebook was one of the earliest widely reported sources of slop. In December 2023, the journalist Jason Koebler of 404 Media reported that Facebook was being flooded with AI-generated variations of photos of wood carvings stolen from an artist, Michael Jones; one AI image of a man with a carved bulldog received more than one million likes, and many commenters appeared to believe it was real.^[7]

In March 2024, Renee DiResta and Josh A. Goldstein published a study of Facebook pages that posted AI-generated images to grow their audiences. They found that spammers and scammers were already getting substantial reach with such images, and that Facebook's recommendation algorithm showed unlabeled synthetic images to users who did not seem to know they were artificial.^[8] 404 Media reported alongside the study that the images — including a recurring series showing Jesus combined with shrimp, known as "Shrimp Jesus" — had drawn hundreds of millions of interactions. Many of the pages later sent their audiences to ad-heavy websites or sold products.^[9]

In August 2024, Koebler traced much of this content to creators in countries including India, Vietnam, and the Philippines, who were paid through Facebook's Performance Bonus program for engagement on their posts. YouTube tutorials and Telegram channels taught them how to generate images with AI image generators, with emotionally manipulative subjects such as starving children and disaster victims chosen to drive likes.^[10]

Fabricated news imagery

Slop can also mislead people during real news events. After Hurricane Helene struck the southeastern United States in 2024, an image of a crying child holding a puppy on a rescue boat spread widely on social media. The fact-checking site PolitiFact rated it false on October 11, 2024. Versions of the image contradicted each other: the child's clothing and the puppy's markings changed between copies, one version showed a malformed hand, and no credible news source had reported the rescue.^[11]

Online video

In November 2025, the video-editing company Kapwing published a report estimating how much AI slop YouTube shows its users. Using data from October 2025, researchers checked the top trending channels in each country by hand for channels made up of AI slop. They also created a new YouTube account and classified the first 500 Shorts (short vertical videos) it was shown.^[5]

The report was published by a company rather than peer-reviewed, relied on manual labeling of channels and videos, and used revenue figures that were estimates rather than disclosed earnings.^[5]

In the workplace

In September 2025, researchers writing in the Harvard Business Review introduced the term workslop for "AI generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task." In a survey of 1,150 full-time U.S. employees, 40% said they had received workslop in the previous month.^[6] The authors argued that workslop "shifts the burden of the work downstream, requiring the receiver to interpret, correct, or redo the work."^[6]

Open-source software

Open-source projects have also received AI-generated bug reports that look plausible but lack substance. The curl project, a widely used data-transfer tool, had run a bug bounty program since 2019. In January 2026, its founder and lead developer Daniel Stenberg announced the program would end on January 31, 2026, after a steep rise in low-quality submissions during 2025. Early in 2026 the project had received seven reports in a 16-hour period and 20 in total by the time of the announcement.^[12] Stenberg said the goal was "to remove the incentive for people to submit crap and non-well researched reports," which put "a high load on the curl security team."^[12]

Effects on AI training data

Slop that stays online can end up in the data used to train future AI systems. A 2024 study in Nature by Ilia Shumailov and colleagues found that "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear" — an effect they named model collapse.^[13] The authors concluded that data about genuine human interactions will become more valuable as AI-generated content spreads across the web.^[13] How much slop currently makes up training datasets, and how much it affects deployed models, has not been established.

Open questions

Some questions about slop are not yet settled by the available evidence:

• Where the line falls. Willison and the dictionary definitions agree that AI involvement alone does not make something slop, but they offer no clear test for when AI-assisted content becomes slop.^[2][1] Kapwing's report quotes the view that the share of AI in a video does not by itself make it better or worse.^[5]
• How much there is. Current estimates, such as Kapwing's, come from small samples, one platform at a time, and manual labeling.^[5] Better estimates would need shared definitions, larger random samples, and data from the platforms themselves, which the platforms generally do not publish.
• How much harm it does. Evidence of harm so far is mostly case studies: fact-checks of individual images, reports from individual projects, and self-reported surveys.^[11][12][6] Controlled studies of how exposure to slop affects trust, attention, or decisions would help settle the question.

Analysis: effects on human agency

This section contains value judgments based on the evidence cited above.

Slop reduces human agency in three ways.

First, it undermines people's ability to understand their situation. Readers of Facebook posts and viewers of disaster photos have been shown fabricated images without labels and, in many cases, without realizing they were fabricated.^[8][11] When fake images spread during a disaster, they compete for attention with accurate information that people need in order to act.

Second, it transfers labor without consent. Workslop and AI-generated bug reports are cheap to produce but expensive to check. The cost falls on colleagues and volunteer maintainers, who lose time they would otherwise spend on their own goals.^[6][12] curl's decision to end its bug bounty shows that this burden can also remove options for everyone else: legitimate security researchers lost a paid channel for reporting real problems.^[12]

Third, the Facebook evidence shows that slop can be sustained by incentive systems — engagement-based recommendation and payment for engagement — that reward content chosen to provoke emotional reactions rather than to inform.^[10][9] Systems that profit from attention regardless of whether it serves the viewer reduce agency even when the content is enjoyed.

Willison's suggested rule — do not publish AI output you have not reviewed and would not stand behind — matches a principle of personal accountability: the person who publishes content is responsible for it, whatever tools produced it.^[2]