AI Slop Explained: Why It's Now an Engineering Problem

NewsGuard has counted 3,749 websites that publish AI-generated news with no meaningful human oversight. That number, current as of June 23, 2026, is the shortest version of why AI slop explained as a matter of taste misses the point entirely. Slop is machine-generated content produced at volume and published without verification. For two years the complaint about it was aesthetic: the images look wrong, the articles say nothing. Then it started breaking things that engineers have to fix. Three systems in particular, each with a documented failure.

Key Takeaways

• AI slop is machine-generated content published at scale without human verification. The defining property is not that a machine wrote it, but that nobody checked it.
• NewsGuard’s tracker counted 3,749 AI content farm news sites as of June 23, 2026, using four criteria including “strong evidence that the content is being published without significant human oversight.”
• It breaks three categories of system: intake queues that depend on human triage, training pipelines that scrape the open web, and retrieval systems that cite sources.
• The training-data risk has a name and a paper. Shumailov and colleagues described “model collapse,” in which “use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear.”
• The practical skill for a kid is not detection. Detectors are unreliable. The skill is asking what a piece of content would have cost to make, and whether anyone staked anything on it being right.

AI slop explained: the actual definition

Slop is content generated by a model, published in quantity, with no verification step between generation and publication.

Notice what that definition does not say. It does not say machine-written content is slop. A doctor using a model to draft patient instructions that she then reads, corrects and signs has produced a document with an accountable human on it. A site publishing four hundred health articles a day that nobody read has produced slop. Same underlying tool, opposite category, and the dividing line is whether anyone’s reputation is attached.

NewsGuard’s AI tracking center operationalises this better than any definition I have seen, which is why it is worth reading their criteria rather than mine. A site counts when all four hold: “there is clear evidence that a substantial portion of the site’s content is produced by AI”; “there is strong evidence that the content is being published without significant human oversight”; “the site is presented in a way that an average reader could assume that its content is produced by human writers or journalists”; and “the site does not clearly disclose that its content is produced by AI.”

Three of those four criteria are about disclosure and oversight. Only one is about who wrote the words. That is the whole conceptual move: slop is a process failure, not an authorship category.

As of June 23, 2026, the tracker listed 3,749 such sites. NewsGuard notes they carry generic names, publish many articles a day, and frequently originate false claims about brands, health, politicians and celebrities.

The three systems it actually breaks

Here is where the word earns its place in an engineering conversation. Each of these is a distinct failure mode with a different fix.

Intake queues saturate

Any system where a human evaluates incoming submissions has a cost ratio: what it takes to produce a submission, divided by what it takes to assess one. When producing becomes nearly free, the queue does not get smarter. It closes.

The cleanest documented case happened on October 1, 2026, when Google paused its Open Source Software Vulnerability Rewards Program, citing “a significant rise in automated submissions, the vast majority of which are not valid.” The programme is suspended until the first quarter of 2027. We walked through the arithmetic in our piece on the bug-report flood.

Training pipelines eat their own output

Modern models are trained in part on text scraped from the open web. If a growing share of that web is model output, the next generation trains on the previous generation’s approximations rather than on the original human distribution.

Shumailov, Shumaylov, Zhao, Gal, Papernot and Anderson named this in a 2023 paper that later appeared in Nature: “use of model-generated content in training causes irreversible defects in the resulting models, where tails of the original content distribution disappear.” They demonstrated it across variational autoencoders, Gaussian mixture models and language models.

The phrase “tails of the distribution” deserves unpacking, because it is the part that matters and the part that gets lost. The tails are the rare things: unusual sentence structures, minority dialects, obscure facts, odd edge cases. Those are exactly what you need a model to know about, and exactly what averages away first. Model collapse is not a system that stops working. It is a system that becomes blander and more confidently wrong about anything uncommon.

Honest caveat: this is a demonstrated effect under conditions of indiscriminate recursive training, which is not precisely how major labs curate data in practice. Real pipelines filter, weight and mix in licensed human data. The effect is real; the magnitude in production systems is not well established publicly.

Retrieval systems cite the slop

When an assistant answers with sources, it is usually retrieving documents and summarising them. If the retrievable web contains thousands of unverified content farms, some of those documents will be slop, and the citation makes the answer look better-grounded than it is.

The security community catalogues this. The OWASP Top 10 for LLM Applications 2025 lists Misinformation as LLM09, calling it “a core vulnerability for applications relying” on model output, and separately lists LLM08, Vector and Embedding Weaknesses, which covers how retrieval itself can be poisoned.

Failure	Who notices first	What it costs, concretely	What actually fixes it
Intake saturation	Maintainers and triage engineers	Google’s open-source bounty closed for a quarter or more	Raising production cost: deposits, reputation gates, mandatory reproductions
Model collapse	Model evaluators, years later	Loss of rare knowledge; blander, more confident errors	Data provenance, licensed human corpora, filtering generated text
Retrieval poisoning	End users, including students	Confident answers citing content farms	Source allowlists, provenance metadata, human-curated indexes
Reader trust	Everyone	A general discount applied to all online text	Disclosure standards and signed provenance

Why detection is the wrong thing to teach

Parents and schools keep reaching for detectors. I understand the appeal and I would not build a routine around them.

The reason is structural. A detector guesses whether text was machine-generated from statistical features, and those features change with every model release and every prompt that says “write less formally.” False positives fall hardest on non-native English writers and on anyone with an unusual style. And crucially, detection answers the wrong question. If a machine drafted an article that three people verified, the detector flags it and the article is fine. If a human wrote 400 words of confident nonsense, the detector clears it and the article is slop.

The better question is about process, and kids can learn it quickly.

A real alternative exists for images, audio and video, and it is worth knowing the name. C2PA, the Coalition for Content Provenance and Authenticity, publishes an open standard for Content Credentials, which the group describes as “like a nutrition label for digital content, giving a peek at the content’s history available for anyone to access.” Its steering members include Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Sony, TikTok and Truepic, and the specification is at version 2.3. Provenance records what happened to a file. That is checkable in a way that style is not.

How to Teach Your Kid About AI Slop

The target concept is cost and accountability, not vibes.

Ages 5–8: the two-drawing test

Draw two pictures of the same thing. Spend thirty seconds on one and five minutes on the other, and be sloppy on purpose with the fast one. Lay them side by side and ask which took longer and how they can tell. Then ask the better question: “If I had to put my name on one of these and show it to your teacher, which would I pick?” Kids connect effort to accountability long before they can define either word.

Ages 9–12: the four-question screen

Give them a real article from a search result and four questions to answer out loud:

1. Who wrote this? Is there a person’s name, and does that name exist elsewhere?
2. What does this site publish besides this? Scroll the homepage. Does it cover forty unrelated topics?
3. Does anything here have a specific number, date or quote that could be checked?
4. If this is wrong, who gets in trouble?

Question four is the one that lands. Slop has no answer to it. A local newspaper, a university page, a government site and a named blogger all do.

Run this on five results from one search and have them sort the sites into “someone is accountable” and “nobody is.” Twenty minutes, and the skill is permanent.

Ages 13+: build the counter-example

Have your teen generate an article with an assistant on a topic they know well, a game, a sport, a band. Then make them fact-check their own output line by line and mark every claim as verified, unverifiable or wrong. Most teenagers find two or three confident errors in 600 words, and finding them in their own output hits differently than being told it happens.

Then the second half of the project: have them rewrite it with sources attached and their name on it. The point is to feel the gap in effort between generating text and standing behind text. That gap is the entire economics of slop.

The question to ask: “If this turns out to be wrong, who loses something?”

If the honest answer is “nobody,” you are looking at slop, whatever wrote it.

What to do about it at home

Put a provenance habit on one thing, not everything

Pick a single category where your family checks sources every time: health claims, prices, or history homework. One habit practised consistently beats a general instruction to “be careful online,” which no child has ever acted on.

Change the search instruction

Instead of “look it up,” say “find two sites that would be embarrassed to be wrong about this.” That phrasing does the work of a media-literacy unit in nine words, and it points kids toward institutions with something at stake. Our longer guide to teaching kids to spot misinformation builds this into a routine.

Treat image provenance as a separate skill

Text slop and image slop fail differently. For images, the useful moves are checking whether Content Credentials exist, reverse-searching, and looking for the specific physical impossibilities models still produce. Pointing at “it looks AI” is a losing strategy, because the look changes every six months.

Say the quiet part about volume

Kids assume that if something appears a lot, it is probably true, which was a reasonable heuristic in 1995. Tell them directly: one person can now publish four hundred articles a day, so repetition is no longer evidence of anything. That single sentence recalibrates a lot of judgments.

What not to do

Do not install an AI detector and treat its verdict as fact, especially for schoolwork. It will be wrong in both directions, it will accuse a kid who writes unusually well, and it teaches the belief that authorship is the thing that matters. Verification is the thing that matters.

What to Watch For Over the Next 3 Months

• Week 4: Watch whether platforms your kids use start displaying provenance. Content Credentials appearing in an image viewer or a video player is the first consumer-visible sign that the standard is real rather than announced.
• Month 2 red flags: Watch for a search result your child trusts that turns out to be a content farm. When it happens, treat it as a lesson rather than a lockdown: walk the four questions together on the actual site.
• Month 3 self-check: Ask your child to explain the difference between “a machine wrote this” and “nobody checked this.” If they can state it cleanly, the concept has landed and the rest is practice.

Frequently Asked Questions

Is all AI-generated content slop?

No, and that distinction is the whole article. Content becomes slop when it is published at volume with no verification and no accountable person attached. A drafted-then-checked-then-signed document is not slop, regardless of what produced the first draft.

How can I tell if a website is an AI content farm?

Check the homepage breadth, look for a real byline you can find elsewhere, see whether any claim carries a specific checkable source, and ask who faces consequences if the article is wrong. NewsGuard’s four criteria formalise the same instincts: substantial AI-produced content, no significant human oversight, presentation that implies human authorship, and no disclosure.

Do AI detectors work?

Not reliably enough to act on. They infer authorship from statistical style, which shifts with each model release, and they produce false accusations that fall disproportionately on non-native writers. Provenance standards such as C2PA attack the problem from the other end, by recording a file’s history rather than guessing at it.

What is model collapse and should I worry about it?

It is the degradation that happens when models train on their own output across generations, with rare knowledge disappearing first. It is demonstrated in research under conditions of indiscriminate recursive training. Whether it meaningfully affects commercial models, which filter and mix in licensed human data, is not publicly settled.

Does this mean my kid’s search results are getting worse?

Possibly for some queries, and the honest answer is that nobody has a clean public measurement of it. What is measurable is the supply side: thousands of unverified content sites now exist that did not in 2023. Teach the accountability question rather than waiting for the platforms to fix it.

What is the single most useful thing to teach?

That repetition is no longer evidence. For most of human history, seeing a claim in many places meant many people had independently found it true. Publishing became free, and that inference broke. Kids who internalise this are substantially harder to fool.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. NewsGuard. “AI Tracking Center.” Updated June 23, 2026. https://www.newsguardtech.com/special-reports/ai-tracking-center/
2. Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R. (2023). “The Curse of Recursion: Training on Generated Data Makes Models Forget.” arXiv:2305.17493; later published in Nature. https://arxiv.org/abs/2305.17493
3. Ha, Anthony. (2026). “Google froze its open source bug bounty program due to a ‘significant rise’ in AI submissions.” TechCrunch, October 4, 2026. https://techcrunch.com/2026/10/04/google-froze-its-open-source-bug-bounty-program-due-to-a-significant-rise-in-ai-submissions/
4. OWASP. (2025). “OWASP Top 10 for LLM Applications 2025.” https://genai.owasp.org/llm-top-10/
5. Coalition for Content Provenance and Authenticity. “C2PA and Content Credentials,” specification 2.3. https://c2pa.org/
6. Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). “Detecting hallucinations in large language models using semantic entropy.” Nature, 630(8017), 625–630. https://pmc.ncbi.nlm.nih.gov/articles/PMC11186750/
7. Stanford Institute for Human-Centered AI. (2025). “The 2025 AI Index Report.” https://hai.stanford.edu/ai-index/2025-ai-index-report