We already know that artificial-intelligence (AI) has a hallucination problem: generating confident responses unjustified by the AI’s training data.
Sometimes, though, an AI will cough up a response that’s built on training data, all right, but the data is dubious or downright mendacious. Jokes. Pranks. Lies.
Such responses are becoming common enough that a word has emerged to describe them: slop.
I first encountered AI slop back in May, in Mignon Fogarty’s AI Sidequest newsletter. She linked to a May 9 Mastodon post from Steve Williams that cited Google’s “anti #Slop policy,” which Williams called “a masterclass in cognitive dissonance.”
Did Williams coin “slop”? If he did, I haven’t been able to confirm it. (And he didn’t do it recently. Keep reading.) He didn’t get the term from the Google guidelines, which use “spam,” not “slop.”1 There is in fact a direct connection between the terms, two four-letter words beginning with S that occupy a sense continuum between “unwanted” and “unappetizing.”
The New York Times legitimized the new sense of “slop” in a June 11 article by Benjamin Hoffman headlined “First Came ‘Spam.’ Now, with A.I., We’ve Got ‘Slop’” (gift link):
Google suggesting that you could add nontoxic glue to make cheese stick to a pizza? That’s slop. So is a low-price digital book that seems like the one you were looking for, but not quite. And those posts in your Facebook feed that seemingly came from nowhere? They’re slop as well.
The term became more prevalent last month when Google incorporated its Gemini A.I. model into its U.S.-based search results. Rather than pointing users toward links, the service attempts to solve a query directly with an “A.I. Overview” — a chunk of text at the top of a results page that uses Gemini to form its best guess at what the user is looking for.
(In case you’re skeptical about the pizza-glue story, there are, as they say, receipts.)
“Slop,” wrote Hoffman, “conjures images of heaps of unappetizing food being shoveled into troughs for livestock. Like that type of slop, A.I.-assisted search comes together quickly, but not necessarily in a way that critical thinkers can stomach.” He found examples of “slop” from as early as 2022, “seemingly . . . in reaction to the release of A.I. art generators.”
In September, New York magazine published
’s article “Drowning in Slop,” which defined “slop” as “a term of art, akin to spam, for low-rent, scammy garbage generated by artificial intelligence and increasingly prevalent across the internet — and beyond.” He issued a warning:There is a fear that as slop takes over, the large language models, or LLMs, that train on internet text will “collapse” into ineffectiveness — garbage in, garbage out. But even this horror story is a kind of wishful thinking: Recent research suggests that as long as an LLM’s training corpus contains at least 10 percent non-synthetic — that is, human — output, it can continue producing slop forever.
And in October, Kate Knibbs wrote in Wired that “AI slop is flooding Medium” (a platform on which I myself publish stories, written 100 percent by me and me alone):
Earlier this year, WIRED asked AI detection startup Pangram Labs to analyze Medium. It took a sampling of 274,466 recent posts over a six-week period and estimated that over 47 percent were likely AI-generated. “This is a couple orders of magnitude more than what I see on the rest of the internet,” says Pangram CEO Max Spero. (The company’s analysis of one day of global news sites this summer found 7 percent as likely AI-generated.)
Slop is an oldish English word, dating back to about 1400 but with roots in Old English sloppe (“dung”).2 Slop originally meant a mudhole or puddle; slip is a related word. In the mid-1600s slop took on the meaning of “liquid or semi-liquid food of a weak, unappetizing kind”; in the mid-1800s its meaning was extended metaphorically to mean “sentimentality.” The OED tells me that slops — refuse fed to pigs — is “originally U.S. dialect and colloquial,” which was news to me.
Sloppy, from the late 1600s, originally meant “covered with water or mud”; it can also mean “slovenly” or “careless” or “disagreeably effusive.” Sloppy Joe, which I know only as “ground beef cooked in a thick sauce and served on a bun” — I used to love this dish back when I still ate mammals — can also mean “a loose-fitting sweater.” There’s a British sweatshirt brand, established in 1993, called Sloppy Joe Essential Clothing; one of its products is called the Diana Zip Neck as a tribute to Princess Diana, a fan of the brand and the style.
The standard dictionaries haven’t yet added the AI sense of “slop” to their lexicons. To nudge them along and give them something to work with, I prompted Perplexity.ai to write 100 words about AI slop in the style of Stephen King. Here’s what it served me:
In the dim glow of the computer screen, AI slop seeped through the circuits like a thick fog, a synthetic mire born from data gone rogue. Each pixel and byte twisted into grotesque forms, reflections of humanity’s darkest fears and desires. It whispered secrets of those who fed it—an echo of our own chaotic thoughts, unleashed and spiraling. Writers and artists swam through this murky digital swamp, grappling with the uncanny creations that emerged. Was it inspiration or insanity? The line blurred, and they realized that what lay in this disturbing algorithm was not just code but a mirror, reflecting back the monstrous aspects of their own souls.
Slop or literature? You be the judge.
UPDATE, November 12:
wrote about “slop” and “the linguistic observer effect”: “‘Slop’ has reached sustained usage in the media, and I’ve even heard my friends casually use the phrase (though I know they weren’t saying it earlier this year).” I’m tagging “slop” as a potential word of the year.More wordy news!
I have two new posts up on Strong Language, the sweary blog about swearing, both about the fourth edition of The F-Word, by Jesse Sheidlower, which was published last week:
“Spam” has meant “unsolicited internet mail” since around 1993. Capitalized, the word is a trademark for the Hormel’s tinned meat product. It’s a portmanteau of spiced and ham.
The plant called cowslip was originally cu-slyppe, or “cow dung,” which was its habitat. There’s also an obsolete “slop,” from the early 1300s, that meant “a charmed bag employed to steal milk from cows” — I really want to know more about this — and another obsolete “slop” or that mean “a loose outer garment.” In the 16th and 17th centuries “slops” were loose breeches.
The “Stephen King AI-generated slop” contained some (accidental?) insight:
“they realized that what lay in this disturbing algorithm was not just code but a mirror, reflecting back the monstrous aspects of their own souls.”
Weirdly, this “truthlet” seems broader than slop, which I am happy to now know about. Perhaps the worst part of slop is that, like all disinformation or misinformation, separating out the grains of truth is like pouring brothy soup—or slop—through a strainer.
Wasn’t it Bannon’s goal to “flood the zone with shit” so as to seed doubt about even that which was true? Now AI can automate that process. Powerfully.
Hard to stay hopeful. Thank you for your 100% person-written insights. Truly.
Damn! I have been mildly creeped out by examples of AI "writing" before, but I am now right on the edge of terrified. I LIKED that "Stephen King" thing. The elements of poetry, logic, and sly humor all seemed to be there.