AI Slop Statistics 2026: The Data Behind the Backlash
In December 2025, Merriam-Webster named 'slop' its Word of the Year. I went looking for the numbers behind that, because a vibe is not a reason to change how your team ships. Here is the data I found, what I think it means, and where I'd be careful reading too much into it.
When Merriam-Webster named slop its Word of the Year in December 2025, I assumed it was a content-farm story. Junk articles, AI YouTube channels, the stuff you scroll past. It mostly was. But the part that kept me up was the code, because that is the slop my customers actually ship, and the numbers there are worse than the vibe suggests.
So I went and collected them. Below is the data I trust enough to repeat, split into the internet at large and the part that lands in your repo. I've kept the sources at the bottom, and I've flagged the numbers I'd treat with a pinch of salt, because half of being useful with statistics is knowing which ones not to quote.
The internet got the headlines
The web-content numbers are the ones that made the news, and they are genuinely large. Ahrefs looked at 600,000 top-ranking Google pages and found that 74% of new pages published in 2025 contained at least some AI-generated material. Only a small slice, around 2.5%, was fully machine-written with no human in the loop. The rest is hybrid, which is exactly what makes it hard to point at and call fake.
The part I found more interesting: the correlation between how much AI content a page has and where it ranks on Google is roughly 0.011. That is nothing. Google is not rewarding AI content, and it is not punishing it either. It is grading on quality and ignoring provenance, which is more or less the same bet we make about code. Nobody cares whether a model or a person wrote the function. They care whether it holds up.
The money side is louder than it is important for our purposes, but worth a line: Kapwing's analysis of YouTube's 15,000 biggest channels found 278 that were pure AI slop, pulling 63 billion combined views and an estimated $117M a year. NewsGuard and Pangram were tracking around 3,000 AI content-farm sites by early 2026, growing by a few hundred a month, some of them running influence operations behind local-newspaper façades. None of that touches your codebase. I include it only because it's the backdrop that made "slop" a word everyone suddenly knew.
The numbers that should actually worry your team
Here is where it stops being a media trend and starts being your on-call rotation.
Somewhere north of half of new commits now contain AI-generated code. The exact figure floats depending on who's counting and how they define it, but every serious estimate I've seen lands in the same place: most of the code entering your repository this year was drafted by a model, not typed by a person. That alone changes the math on review. Review used to scale with output because the same human was bounded by the same typing speed. Agents broke that coupling. One engineer can open a hundred-file PR before lunch.
The quality gap is real but it is not catastrophic, and I want to be precise about that because the scary framing helps nobody. CodeRabbit's 2025 benchmark found AI-authored PRs carried roughly 1.7x the issues per PR of human ones, and the gap widened with PR size. That tracks with what I see: small agent changes are fine, large ones accrete shortcuts. Endor Labs put security weaknesses in around 62% of AI-generated code, which sounds enormous until you remember "weakness" is a broad bucket and a lot of it is low-severity. It's still a lot of low-severity work nobody chose to take on.
The number I keep coming back to is the quiet one: a large share of AI code faults are silent. They compile, they pass tests, and they produce the wrong answer under conditions the model never considered. That is the category traditional CI was never built to catch, because your tests check the behavior you thought to write down, and the failure is in the behavior you didn't. An empty catch block does not turn your build red. It turns your dashboard blank three weeks later.
Read these with a pinch of salt
A few of the figures floating around the AI-slop discourse are softer than they look, and quoting them confidently is its own small kind of slop.
"43% of AI code needs debugging in production" gets passed around a lot. The underlying analysis is real, but "needs debugging" is doing heavy lifting, and the methodology behind the percentage is thinner than the precision implies. I'd use it as a direction, not a stat. Same with the social-listening numbers, the 475,000 mentions in a month, the +9,100% search growth. They're true and they capture a real shift in how people talk, but search-volume spikes measure attention, not engineering risk. A topic can trend and still not be your problem.
The honest summary is narrower than the headlines: AI writes most of the new code, a meaningful fraction of it is shallow or unsafe in ways your existing checks miss, and the volume is past what manual review can absorb. You don't need the scarier numbers to act on that.
What I'd actually do with this
If you lead a team shipping agent-written code, the data points at two changes, not ten.
Stop relying on review as the catch. The volume argument is decisive on its own: no human reads a hundred files a day with the care they gave to ten, and pretending otherwise just moves the slop downstream to whoever's on call. The patterns that recur (swallowed exceptions, unsafe casts, hallucinated imports, dead code) are mechanical, so let something mechanical catch them and save the humans for the judgment calls a rule can't make.
Then move the catch earlier than the PR. A gate that blocks before review beats a reviewer who comments after, and a hook that flags an issue while the agent still has the context loaded beats both. That's the whole thesis behind what we build, and it's the one conclusion in this piece I'd stake a number on.
You can see where your own repo lands in a few seconds: npx aislop scan. The score is more useful than any industry average, because it's yours.
Sources
Ahrefs (600K page study, 2025) · NewsGuard & Pangram Labs (content-farm tracking, 2026) · Kapwing (YouTube analysis, 2025) · Endor Labs (security analysis, 2025) · CodeRabbit (PR benchmark, 2025) · Exploding Topics & Visibrain (search and social monitoring, 2025–2026) · scanaislop internal research across 25+ projects. Figures are reported as published; where studies disagree, I've described the range rather than picking the most dramatic point.