What is AI slop? How to spot it, why it spreads and how to avoid making it

AI slop is low-value content made wholly or substantially with generative AI and released with too little human judgment, verification, craft, or accountability. It is commonly produced because speed, scale, traffic, engagement, or cost reduction matters more to the publisher than usefulness to the person receiving it.

That definition includes the disposable articles clogging search results, but also synthetic videos built to hold a child’s attention, fake product reviews, hollow business memos, and realistic images created to bait reactions. It does not include everything touched by AI. A reporter who uses software to transcribe an interview has not produced slop, and neither has an expert who uses a model to organize original research that an editor later checks.

The distinction matters because AI use is becoming ordinary. Quality, contribution, purpose, verification, and accountability tell us far more than the presence of a machine somewhere in the workflow.

The short definition: AI slop is generative content that looks more complete than the work behind it deserves. It is cheap for the producer because readers, editors, colleagues, platforms, or advertisers inherit the cost of checking and cleaning it up.

With over 20 years of experience as a publisher, I understand firsthand the harm that producing such content can cause to your audience. It’s not just publications, individual influencers can also be negatively impacted by the content they create.

Where did the term “AI slop” come from?

The word “slop” is old, but its current technological meaning took hold in the early generative AI era. Developer Simon Willison helped popularize it in May 2024 after sharing an earlier observation that slop could become for unwanted AI content what spam had become for unwanted email. Willison did not claim to invent the term, and early uses predate his post. His contribution was helping a scattered complaint become a useful label. Simon Willison’s original post remains the cleanest starting point.

The label spread because it named something that “AI-generated content” could not. The latter describes a production method. Slop expresses a judgment about the result, the effort behind it, the way it is distributed, and the burden it creates for other people.

By the end of 2025, the term had moved from internet slang into the dictionary. Merriam-Webster selected “slop” as its 2025 Word of the Year and defined it as low-quality digital content usually produced in quantity by artificial intelligence. Merriam-Webster’s announcement reflected how quickly the public vocabulary had caught up with the production boom.

Researchers still disagree about the exact boundary. A June 2026 Columbia University report described AI slop as high-volume synthetic text, images, audio, or video generated quickly and optimized for engagement rather than depth. It also warned that the label can flatten important differences between harmless experimentation, artistic expression, annoying clutter, manipulation, and fraud. The Columbia SIPA report is the most comprehensive map of those tensions so far.

AI-generated content and AI slop are not the same thing

This is the most important boundary in the entire discussion.

AI-generated content is a broad category that can include a translated support page, a synthetic background in a film, a cleaned transcript, an automatically written weather summary, or an article nobody checked. Some examples are useful and carefully supervised. Others are slop.

Research into 900,000 English-language pages created in April 2025 estimated that 74.2% contained some AI-generated text. That result does not mean three quarters of new pages were slop. It means AI assistance was widespread according to the detector and thresholds used in the study. Ahrefs published its methodology and results, and that methodological detail is the difference between a useful statistic and an alarming but false conclusion.

Three characteristics identified in the 2026 paper “Why Slop Matters” help explain why some synthetic content earns the label:

• Superficial competence: the work has the appearance of skill, polish, or authority but lacks the substance that appearance implies.
• Asymmetric effort: producing it requires far less effort than the output would have required without generative AI.
• Mass producibility: the same process can create huge quantities of similar material.

Those are family resemblances, not a legal definition. They explain why a polished report with fabricated sources can feel more slop-like than a clumsy but honest human note.

What counts as AI slop?

The following matrix handles common borderline cases.

The same asset can also change character through distribution. One bizarre AI image posted as an obvious joke is different from the same image copied across hundreds of accounts, paired with a false story, and used to push people toward scam pages.

Why AI slop is spreading

Generative AI did not invent junk content. Content farms, scraped pages, clickbait, fake reviews, template videos, and made-for-advertising sites existed long before ChatGPT. AI changed their economics.

A publisher once needed writers, designers, editors, and time to manufacture a convincing imitation of a publication. Now a small operation can generate text, images, narration, code, headlines, and translations through connected tools. Each additional item can cost almost nothing, while search engines, social feeds, creator programs, and ad exchanges provide immediate distribution.

The result is a simple economic mismatch:

The cost of producing another item approaches zero, but human attention and verification remain expensive.

Columbia’s 2026 analysis argues that slop often reflects incentive systems more than individual creative failure. Monetization programs, engagement metrics, ranking systems, and automation pipelines reward scale, speed, and emotional response. Under those conditions, generative abundance becomes a rational strategy for the producer even when the overall system becomes worse for users.

The AutoBait operation documented by DoubleVerify makes that business model unusually visible. Researchers found a network of more than 200 made-for-advertising sites using templated prompts to generate clickbait articles and realistic-looking images. Some slideshow pages contained dozens of slides and hundreds of potential ad placements. DoubleVerify estimated that a page could cost less than $2.25 to generate. Its investigation includes the exposed production prompts.

In that model, the article is not really the product. Ad inventory is the product, and the article is inexpensive wrapping designed to acquire a few seconds of attention.

How much AI slop is there? Read the numbers carefully

There is no reliable census of AI slop across the internet. The category is subjective, changes by context, and cannot yet be measured accurately with a single automated tool. The most quoted figures answer narrower questions.

NewsGuard’s count is particularly useful because its process combines automated detection with analyst review and requires substantial AI content, a lack of clear disclosure, and repeated low-quality output. Its tracker had identified 3,749 sites across 16 languages by June 2026. NewsGuard publishes the current tracker and criteria.

The feed studies from Kapwing’s YouTube report and TikTok report are vivid, but their fresh-account samples should be treated as demonstrations rather than universal estimates. Recommendation feeds personalize rapidly, and one account cannot represent a platform.

The main forms of AI slop

Slop adapts to the system distributing it. A search engine rewards query coverage, a short-video feed rewards retention, an ad network rewards impressions, and a workplace may reward the appearance of productivity. The surface changes while the production logic remains familiar.

1. Search slop

Search slop consists of pages built to capture queries rather than resolve them. Common examples include mass-generated location pages, rewritten definitions, old articles with a new date, comparison pages assembled from manufacturer copy, and troubleshooting guides whose authors never used the product.

The most deceptive version borrows signals of experience. A page says “we tested,” “our experts found,” or “after 30 days” without a test record, original photographs, measurements, named reviewer, or any observation that could only come from actual use.

2. Social and video slop

Feed slop is optimized for a reaction before a viewer has time to inspect it. It often uses exaggerated peril, rescue stories, sentimental scenes, celebrity fabrications, or a chain of unrelated synthetic clips held together by narration. Children’s content is particularly attractive to this model because repetition, bright imagery, familiar characters, and constant novelty can hold attention without a coherent story.

YouTube’s monetization rules now explicitly reject mass-produced, generic, repetitive, or manipulative channels, while allowing AI tools when the result carries original creative or educational value. YouTube’s current policy captures the distinction better than a blanket ban would.

3. Image slop and synthetic empathy bait

Image slop includes impossible architecture, fake historical photographs, nonexistent disasters, imaginary products, and emotionally loaded scenes designed to solicit comments. The famous “Shrimp Jesus” images were strange enough to become a joke, but investigations described networks using similarly surreal religious imagery to generate engagement and route people toward ad-heavy or scam sites.

Visual quality is improving quickly, so malformed hands and mangled text are no longer dependable screening methods. Source, context, posting behavior, and provenance increasingly matter more than anatomy.

4. Commercial and review slop

Generative systems can produce thousands of plausible product descriptions, testimonials, buying guides, and comparison tables without anyone touching the products. That makes commercial slop especially damaging because the content imitates independent judgment at the moment a reader is preparing to spend money.

A summary based on disclosed specifications can be legitimate. A first-person verdict based on imaginary use is not. The problem is fabricated evidence, not polished grammar.

5. Workslop

Workslop is AI-generated workplace material that appears finished but lacks the context or thinking needed to move a task forward. It includes a strategic memo without a decision, code without an explanation, a research summary with unsupported claims, or a long email that leaves the recipient to discover the actual request.

Research from BetterUp Labs and Stanford Social Media Lab surveyed 1,150 full-time US desk workers in September 2025. Forty percent said they had received workslop during the previous month, and respondents estimated that resolving an incident took about two hours. BetterUp publishes the sample and headline findings. The exact cost will vary by organization, but the mechanism is clear: one person records a productivity gain while another person receives the rework.

6. Slopaganda

“Slopaganda” describes synthetic political communication that borrows the cheap production, emotional intensity, and feed-native style of commercial slop. It can be obviously absurd and still shape what feels normal, threatening, popular, or worthy of ridicule.

Not every synthetic political joke is misinformation. The risk rises when realistic content hides its origin, impersonates someone, invents an event, targets a vulnerable audience, or overwhelms the information environment through coordinated repetition.

7. Knowledge and research slop

Academic papers, educational videos, reports, and reference material gain authority from their format. Generative output can reproduce that format while inventing citations, flattening uncertainty, repeating retracted findings, or creating a trail of sources that all lead back to the same unsupported claim.

This form of slop is harder to discard because later writers, databases, and models may treat the published item as evidence. A weak article can become an input, then a citation, then an apparent consensus.

How to spot AI slop without accusing every tidy writer

No single clue proves that content was generated or that it is slop. Look for clusters of evidence and examine what the publisher did, not merely how the sentences sound.

Stronger warning signs

• The page claims personal testing, travel, interviews, or ownership without any evidence unique to that experience.
• Citations exist, but they do not support the sentence beside them.
• Sources cite one another in a loop, while the original claim cannot be found.
• Current dates sit above stale screenshots, discontinued products, old prices, or obsolete instructions.
• The content covers every subtopic but never makes a difficult choice, states a limitation, or explains what would change the recommendation.
• Several pages use the same structure, examples, verdict, and phrasing with only the entity names replaced.
• One author publishes implausible volumes across unrelated areas of expertise.
• The page contains prompt fragments, model disclaimers, placeholder text, invented quotations, or instructions meant for an editor.
• A supposed review offers only public specifications and stock or manufacturer images.
• The site has no credible authorship, correction route, editorial policy, or ownership information.

Weaker clues that people overvalue

Perfect grammar, tidy headings, transition words, short introductions, repeated sentence patterns, certain favorite phrases, and enthusiastic punctuation may raise suspicion. Excessive em dashes have become a running joke for a reason. None of these proves AI use, and none tells you whether the work is accurate or useful.

Writers can sound formulaic. Models can be prompted to sound irregular. A serious assessment must reach beneath the voice.

Why AI detectors cannot decide whether something is slop

AI detectors attempt to estimate how likely it is that text or media came from a model. Slop is a judgment about quality, usefulness, context, intent, and production behavior. These are different questions.

Research summarized in Columbia’s 2026 report found that professional editors’ slop judgments reflected usefulness, accuracy, framing, coherence, relevance, and writing quality. The importance of each dimension changed by context, and automated methods did not reliably reproduce the human judgments. The underlying study, “Measuring AI Slop in Text”, describes fully automated measurement as an open challenge.

Detectors also face edited output, mixed authorship, translated text, new models, false positives, and an obvious incentive for publishers to evade them. Use a detector as one signal during an investigation, never as proof that a writer cheated or a page lacks value.

Provenance tools answer a narrower and increasingly useful question. C2PA Content Credentials can record an asset’s origin, edits, and use of AI in a cryptographically bound history. They work like a digital nutrition label. However, C2PA explicitly warns that credentials do not decide whether a claim is true, and the absence of credentials should not make an asset automatically untrustworthy. The C2PA explainer describes both the standard and its limits.

For an ordinary reader, the best approach remains lateral reading: leave the page, inspect the publisher, find the original source, compare independent coverage, and check whether the apparent evidence exists.

Is AI slop bad for SEO?

AI content is not automatically against Google’s rules. Google’s public guidance says generative AI can help with research and structure, while generating many pages without adding user value may violate its scaled content abuse policy. Google’s generative AI guidance focuses on the result and the purpose, not a blanket ban on a production tool.

The spam policy is even more direct. Scaled content abuse involves producing many pages primarily to manipulate rankings rather than help users, regardless of whether the pages were made by AI, humans, scraping, translation, or stitched sources. Google lists mass AI generation without added value among its examples.

That distinction creates three practical lessons for publishers:

1. AI percentage is a poor SEO target. A page does not become useful when a detector labels it human, and it does not become useless when AI helped edit it.
2. Scale multiplies weak decisions. One poorly reviewed draft is an editorial error. Ten thousand templated pages can become a policy and site-quality problem.
3. Originality must exist in the inputs. A longer prompt cannot manufacture firsthand evidence. Give the process proprietary data, reporting, testing, expert decisions, or a genuinely new synthesis before asking for prose.

Slop can rank temporarily. Spam has always found temporary openings. A strategy that depends on ranking systems failing to recognize the absence of value is still a bet against the search engine’s stated objective, the reader’s memory, and the publisher’s reputation.

How to use AI without making slop

The answer is not to sprinkle quirks into a generated draft until it passes as human. That changes the disguise while leaving the work untouched.

1. Start with evidence rather than a blank prompt

Collect the interview, dataset, screenshots, test notes, product access, source documents, customer questions, or subject-matter judgment first. Ask AI to help transform material you possess rather than imitate knowledge you never gathered.

2. Give every factual claim a route home

Maintain a claim ledger for high-value work. Record the claim, its source, the exact supporting passage or observation, the date checked, and the editor who verified it. If a source does not support the final wording, narrow the claim.

3. Separate generation from verification

Do not ask the same model response to create a claim and certify that claim. Check important facts against primary documents, real software, raw data, recordings, or independent experts. High-stakes subjects require subject-matter review.

4. Make the human contribution visible

Include methods, exclusions, photographs, screenshots, calculations, dissenting findings, uncertainty, and the reasoning behind a recommendation. A reader should be able to identify what the publisher learned or decided.

5. Assign one accountable owner

An editor or author should be willing to put a name on the final output and answer for its accuracy. “The AI wrote it” is a description of process, not a correction policy.

6. Tie publishing volume to review capacity

If a team can rigorously review ten articles a week, the ability to generate 500 does not create capacity for 500. It creates a queue of unverified output. The safe production ceiling is set by evidence gathering and quality control, not generation speed.

7. Disclose AI use when it changes what the audience needs to know

Disclosure is especially valuable for realistic synthetic media, public-interest information without human review, automated updates, or content where readers may reasonably assume a person witnessed or created the underlying material. A label should be specific enough to help: “AI-generated illustration,” “automated summary reviewed by Jane Smith,” or “voice translated with AI” says more than “AI may have been used.”

Disclosure does not repair invented evidence, weak analysis, or careless review. Transparency and quality are separate duties.

8. Preserve corrections and updates

Make it easy to report a problem, show when the material was last checked, and correct the record visibly when a consequential error occurs. Automation raises the speed of publication, so the correction system must be able to keep up.

What platforms and regulators are doing

Responses increasingly target behavior and incentives rather than the mere presence of AI.

• Google Search prohibits scaled content produced primarily for ranking manipulation and lacking user value, regardless of how it was created.
• YouTube makes mass-produced, generic, repetitive, and manipulative content ineligible for monetization, while permitting meaningful AI-assisted creation.
• Meta says it is prioritizing original creator work and reducing the distribution of duplicative or minimally transformed material. In March 2026, the company reported that views and time spent on original Facebook Reels had roughly doubled in the second half of 2025 compared with the same period a year earlier. Meta’s policy explanation also says repeat offenders can lose recommendation and monetization.
• The European Union began applying Article 50 transparency obligations on August 2, 2026. Providers of relevant generative systems must add machine-readable marks for AI-generated or manipulated content. Deployers must inform people about deepfakes and AI-generated public-interest text published without human review or editorial control, among other covered uses and exceptions. The European Commission published updated guidance on August 6, 2026.

These measures can reduce certain incentives, but labels alone cannot measure usefulness. A clearly labeled page can still waste time, while a carefully reviewed AI-assisted page may be excellent.