AI Slop in 2026: A Wall of Real Quotes and Numbers

Where the word came from

The word arrived on 8 May 2024, in a post by Simon Willison, a software developer and co-creator of the Django web framework. He was careful about credit: he did not coin the term, he popularized it, and his post points at an earlier tweet. What made it stick was that he supplied a test rather than an insult. Slop, he wrote, is content “mindlessly generated and thrust upon someone who didn’t ask for it”.

Notice where the offense sits in that sentence. Not in the tool. In the consent and the missing review: nobody asked for it, and nobody looked at it before it went out.

Eighteen months later, a dictionary ratified the same idea. On 24 November 2025 the Macquarie Dictionary made “AI slop” its Word of the Year, defining it as “low-quality content created by generative AI, often containing errors, and not requested by the user”. Fifteen words, and the last five carry the argument. It took both the committee’s pick and the People’s Choice - only the fourth time that has happened in the award’s history.

Macquarie was not alone in noticing. In December 2025 Merriam-Webster made the bare word “slop” its own Word of the Year. Two dictionaries, one year, the same complaint - one naming the compound, one naming the stuff itself.

So this is a described phenomenon with a fixed origin date, not a mood. Here is what people have counted since - sixteen notes, in the order the story happened.

How much of it is there? Nobody can honestly tell you

The best available measurement comes from the research team at Graphite - Gregory Druck PhD, Jose Luis Paredes PhD, Bevin Benson and Ethan Smith - published in May 2026: roughly half of all new articles published on the web are primarily AI-generated. The sample was 55,400 English-language URLs drawn from Common Crawl. The trend is a plateau rather than a spike: 49.6% in the first quarter of 2025, 50.9% in the fourth, 49.9% in the first quarter of 2026.

Now the methodology, in the same breath, because it is not a footnote. Three AI detectors did the classifying, and the study averaged them. That is a deliberate improvement on Graphite’s own October 2025 study, which used a single detector and, by the team’s own account, ran about 3.3 points high. It is still a measurement taken with the instrument section six examines.

Two more limits worth stating plainly. “The web” is Graphite’s generalization from English-language articles in one crawl. And per the team’s related work, these articles largely do not surface in Google or ChatGPT results - so this counts what gets published, not what anybody reads.

Not every figure depends on a machine reading prose. NewsGuard’s AI Tracking Center counts sites rather than sentences: 49 AI content farm news sites in May 2023, and 3,749 as of 23 June 2026, spanning sixteen languages. That is a census of what NewsGuard’s analysts have found and rated, not an exhaustive map of the web, and the counter is revised upward roughly monthly. Read it as a floor.

Platforms counting their own intake produce the same shape. On 21 July 2026 Deezer said AI-generated tracks had passed half of its daily uploads for the first time, on peak days, with June 2026 - the peak month - averaging roughly 90,000 a day. The same company is careful about what that does and does not mean: AI music accounts for only 1–3% of actual streams, and up to 85% of those streams were fraudulent and demonetized. Uploads are not listening.

And the flood reaches the top of the chart, not just the bottom of the pile. A July 2026 arXiv preprint by Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg and Paramveer Dhillon examined 14,419 self-published genre-fiction books with sales through June 2026 and found that 31% of new entrants to its Top-25 lists carried substantial AI text - over 25% of the text, as classified by the Pangram detector - against near zero in 2023. Two qualifications travel with that number: it is a preprint, not yet peer-reviewed, and Top-25 is the paper’s own within-sample, per-genre ranking of self-published fiction, not Amazon’s public bestseller chart.

What you will not find here is the familiar claim that 90% of the internet is AI-generated. It circulates constantly and traces back to no published methodology. A number with no sample and no method is a mood with a percent sign on it.

The platforms have not been idle. Amazon capped self-publishers at three new titles a day in September 2023 - its own stated rationale was abuse generally, and it noted it had not seen a publishing spike; the AI framing came from the Authors Guild. Spotify said in September 2025 that it had removed over 75 million spammy tracks in twelve months. And Google, in an update to a post by Elizabeth Tucker, its Director of Product Management for Search, said people would “now see 45% less low-quality, unoriginal content in search results” after its March 2024 update - Google’s own measure of its own cleanup. Whether any of that changes what actually surfaces is a separate question with a post of its own: what this is doing to search rankings.

It reached the places whose whole job is checking

Open-web numbers are easy to hold at arm’s length. These are harder, because somebody was supposed to be reading.

In summer 2025 Jenna Russell and Mohit Iyyer at the University of Maryland, working with Pangram Labs, put 186,000 articles from about 1,500 US newspapers through a detector. About 9% came back partly or fully AI-generated. The rate was 9.3% at small papers against 1.7% at papers with circulation above 100,000, and only seven of the newspapers had any public AI policy at all. The work was published at ACL 2026.

The sharpest figure in that study is not the 9%. It is the disclosure rate: of a hundred flagged pieces the researchers examined, five said so. The problem the evidence keeps describing is not that a tool was used. It is that nobody mentioned it.

In May 2025 a 56-page summer supplement went out through King Features Syndicate and ran in the Chicago Sun-Times and the Philadelphia Inquirer. Its summer reading list carried fifteen books. Ten of them did not exist. The authors were real, living writers; the titles were invented and hung on their names. A freelancer admitted using AI and not checking the output, the syndicator ended the relationship, the newsrooms’ own staff had no involvement, and the Inquirer’s editor, Gabriel Escobar, called it a violation of the paper’s internal policies and a serious breach.

Then there is the case that folds in on itself. Ars Technica retracted a story containing fabricated quotes and fired the reporter - and the retracted piece was itself about an AI agent fabricating an attack on an engineer. In a report dated 2 March 2026, Futurism quoted the site’s editor-in-chief, Ken Fisher, describing it as “a serious failure of our standards”.

All three are citable for the same reason: each institution counted, corrected or fired in public. The failures anyone can name are the ones somebody owned, which is worth saying before treating them as villains. For the practical newsroom side of this, a journalist’s guide to content authenticity goes further than this post can.

Readers already adjusted

None of this is happening to an unsuspecting audience.

Pew Research Center surveyed 5,023 US adults between 9 and 15 June 2025 and asked how important it is to be able to tell AI-generated pictures, video and text from the human-made kind. Seventy-six percent said extremely or very important. In the adjacent sentence of the same report, 53% said they are not confident they could actually do it. Near-universal demand to know which is which; a majority admitting they cannot tell on their own. One survey, one sample, so that gap is a finding rather than two studies stitched together.

The Reuters Institute’s Digital News Report 2026 - 48 markets, about 2,000 respondents each - put trust in news at 37%, a record low within its own series, which begins in 2015, and falling in 29 of those markets. Trust in news that arrives via an AI chatbot sat at 20% across all audiences; among people who already use chatbots it was 44%.

Its separate Generative AI and News Report, published in October 2025 across six countries, found comfort rising in a staircase as the human share rises: 12% comfortable with news made entirely by AI, 21% with AI under human oversight, 43% with human-led work using AI help, and 62% with news made entirely by humans - the last figure up four points on 2024.

Read those together and the demand turns out to be modest. Nobody in these samples is asking for a purity test. They are asking to know which one they are reading.

The bill lands on the people still writing

Asking costs a reader nothing. Being the person who has to work out the answer, one submission at a time, is where the volume actually lands.

By noon on 20 February 2023, Clarkesworld Magazine had taken in about 700 legitimate submissions and 500 machine-written ones. Neil Clarke, its publisher and editor-in-chief, closed the inbox. The baseline is what makes that land: in the magazine’s history, spam had never exceeded 25 submissions in a month, and plenty of months had none. Against a normal intake of roughly 1,100 submissions a month, the flood was not the whole slush pile - it was simply enough to make reading the pile impossible.

The obvious objection is that February 2023 was a novelty spike around a new toy. Clarke answered it himself in an August 2025 editorial: submissions more than doubled at their peak, and a second wave - lightly polished rather than raw - now jeopardizes the magazine’s ability to properly prioritize human writing and keep submissions open.

Smaller presses are still finding out. In July 2026 Bona Books, a UK queer speculative-fiction press, published a post explaining that it had bought an AI-written story out of 606 submissions and caught it before publication. The anthology was pushed to autumn 2026.

The money has moved too. The UK Society of Authors surveyed its members in January 2024 - 787 responses from a body of about 12,500 - and 36% of translators and 26% of illustrators said they had already lost work to generative AI. That is self-report, collected by the trade body of the people affected, which is both its weakness and its point: these are the people who would know first.

The books preprint gives the structural version of the same thing. Books with substantial AI text made up 20.0% of its sample but 12.1% of sales and 11.3% of revenue. The market gained a great many more titles to sell without gaining a proportionate amount of money to pay for them. Dilution is the paper’s own word for it.

The instrument everyone reaches for is the one that misfires

Go back over the last four sections and count the detectors. Graphite classified the web with three of them. The Maryland newspaper study used Pangram. The books preprint used Pangram. Platforms screen uploads with them, editors screen submissions with them, teachers screen essays with them. The census-takers and the judges are holding the same instrument.

Here is what that instrument does. In 2023, Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou at Stanford ran 91 human-written TOEFL essays, by non-native English writers, through seven GPT detectors. On average the detectors wrongly flagged 61% of them as AI-generated; individual detectors ranged from 48% to 76%. Eighty-nine of the 91 essays were flagged by at least one detector, and all seven agreed on eighteen of them. Publishing in the peer-reviewed journal Patterns, the authors strongly caution against using GPT detectors in evaluative or educational settings, and describe their own study as a pilot.

They also showed the mechanism, which is the part that should worry anyone quoting a detector. Enrich the essays’ vocabulary and the false-positive rate falls to 11.6%. The detectors were not finding a machine. They were finding a smaller vocabulary.

The cost of that lands on individuals, one at a time. Marley Stevens, a student at the University of North Georgia, was put on academic probation and required to pay $105 to attend a seminar on cheating after Turnitin flagged her work; Copyleaks flagged it too. She says she used nothing but Grammarly’s standard spelling and grammar suggestions. “AI detectors are garbage”, she told EdSurge in April 2024 - and the same article discloses that she later became a Grammarly student ambassador and that the company donated to her fundraiser, which you should know when weighing the quote.

Moira Olmsted, a student at Central Methodist University, has autism spectrum disorder and writes in the kind of formulaic style detectors punish. Her grade was reversed after she was flagged, but she was warned that a second flag would be treated as plagiarism. She now screen-records herself writing and drafts in Google Docs so there is a change history behind the words. “I have so much to lose”, she told Bloomberg Businessweek in October 2024. In the same investigation, Bloomberg obtained 500 Texas A&M application essays written in summer 2022 by public records request - old enough to effectively guarantee they were not AI-generated - and ran them through GPTZero and Copyleaks. Between 1% and 2% came back falsely flagged, sometimes with a claimed certainty near 100%.

This is the point where an article like this usually hands you a checklist: uniform paragraph lengths, no typos, a flat register, an intensifier where a number should be. Please do not use it. Tells are heuristics, not evidence. They miss careful generated text, and they punish precisely the people the Stanford paper identified - writers working in a second language, writers with a formal register, writers who proofread. A tell is a reason to look closer. It is never a finding.

0 of 12 cleared

Reveal the restReset

• Graphite research team, May 2026Source
  50% of new articlesPeel
• NewsGuard AI Tracking Center, June 2026Source
  49 to 3,749 sitesPeel
• Deezer newsroom, July 2026Source
  ~90,000 AI tracks a dayPeel
• Chakrabarty, Liu, Ginsburg & Dhillon, July 2026Source
  31% of new bestseller entrantsPeel
• Russell & Iyyer, University of Maryland, Nov 2025Source
  9% of newspaper articlesPeel
• King Features syndicated supplement, May 2025Source
  10 of 15 books did not existPeel
• Ken Fisher, editor-in-chief, Ars Technica, 2026Source
  "a serious failure of our standards"Peel
• Pew Research Center, Sept 2025Source
  76% want to tell. 53% can't.Peel
• Neil Clarke, Clarkesworld, Feb 2023Source
  500 AI submissions. Inbox closed.Peel
• Society of Authors survey, 2024Source
  36% of translators lost workPeel
• Liang et al., Stanford, Patterns, 2023Source
  61% of human essays flaggedPeel
• Marley Stevens, student, 2024Source
  "AI detectors are garbage"Peel

Twelve of the notes on this page, squeezed down to what fits on a sticky. Peel one and you get the half that matters: who counted it, and when.

Every note you just cleared had a name, a date and a link behind it. That is the whole difference: not how the words read, but whether anything stands behind them.

All twelve notes as text, with names, dates and sources

• 50% of new articles - Graphite research team, May 2026 · Source
• 49 to 3,749 sites - NewsGuard AI Tracking Center, June 2026 · Source
• ~90,000 AI tracks a day - Deezer newsroom, July 2026 · Source
• 31% of new bestseller entrants - Chakrabarty, Liu, Ginsburg & Dhillon, July 2026 · Source
• 9% of newspaper articles - Russell & Iyyer, University of Maryland, Nov 2025 · Source
• 10 of 15 books did not exist - King Features syndicated supplement, May 2025 · Source
• "a serious failure of our standards" - Ken Fisher, editor-in-chief, Ars Technica, 2026 · Source
• 76% want to tell. 53% can't. - Pew Research Center, Sept 2025 · Source
• 500 AI submissions. Inbox closed. - Neil Clarke, Clarkesworld, Feb 2023 · Source
• 36% of translators lost work - Society of Authors survey, 2024 · Source
• 61% of human essays flagged - Liang et al., Stanford, Patterns, 2023 · Source
• "AI detectors are garbage" - Marley Stevens, student, 2024 · Source

The individual-harm half of this problem has a post of its own: why detectors misfire on human writing. This section treats detectors as unreliable census-takers; that one treats them as unreliable judges. It is the same instrument in both jobs.

Nobody asked for this

Return to the last five words of the dictionary definition: “not requested by the user”. Macquarie’s committee put the exhaustion where it belongs. Their observation was that readers now have to behave something like prompt engineers simply to wade through what is in front of them. The work of filtering has been handed to the person who only wanted to read something.

The same burden has arrived at the other end of the pipe. In March 2026, discussing the Authors Guild’s certification work, its chief executive Mary Rasenberger described “growing pressure on authors and publishers to differentiate human-written work from the slop”. Set the certification program aside and look at the sentence. The head of the largest writers’ organization in the United States is describing a world in which writing the thing is no longer the whole job. You now also have to establish that you wrote it.

That is a strange bill to hand to the person who did the work. It is also, at this point, the bill.

The part of writing that stays expensive

Section six is the reason this post cannot end by teaching you to spot slop by reading. Nothing inside a finished text reliably separates a person’s work from a machine’s, and the tools that claim otherwise fail hardest on the most careful writers.

The making of it is not ambiguous at all. Slop’s defining property - the one every source on this wall circles - is that it costs almost nothing to produce. That is what makes 3,749 tracked sites, and 90,000 tracks a day in a peak month, arithmetically possible. What stays expensive is the other half: a record of the work has to be made while the work is. Research notes, dated drafts, a version history, the record of a paragraph being written and rewritten - none of it can be attached to a finished page afterwards.

That is an asymmetry, not a test. Someone could sit and retype generated text to manufacture a record, and nothing described here would catch them - it would simply cost them the time the shortcut was meant to save. Which is the point: a finished document is a claim, and a claim is cheap. A record of it being made is evidence, and evidence takes as long as the work did.

Moira Olmsted worked that out for herself, before anyone sold her a tool for it.

The methods, including the ones that need no product at all, are compared in the complete guide to proving human authorship.