AI Slop

AI slop, often shortened to slop, is low-quality digital content produced in bulk with generative AI and little human care: filler articles, uncanny images, narrated stock-footage videos, and machine-made songs posted to capture search traffic, engagement, or ad revenue. Merriam-Webster chose slop as its 2025 Word of the Year, defining it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence," and the Macquarie Dictionary picked "AI slop" for the same year. What makes content slop is how and why it is made, at near-zero cost and in volume, rather than any single error it contains.

Where the Term Comes From#

Slop originally meant soft mud and later food scraps fed to pigs. In May 2024, programmer Simon Willison argued that slop should become the name for unwanted AI-generated content the way spam became the name for unwanted email, crediting the online writer deepfates for noticing the word turning into a term of art. The same year, Facebook feeds filled with surreal text-to-image pictures such as "Shrimp Jesus," images of Jesus fused with shrimp, posted by pages chasing engagement, and by 2025 the word covered AI-made video, music, books, and academic writing as well.

Why AI Slop Spreads#

Slop is an economic product. A large language model (LLM) or an image generator built on diffusion models can turn out thousands of posts for the cost of one human-made piece, while feeds ranked by a recommendation system and ad networks pay for attention regardless of who or what made the content. A producer only needs a small fraction of posts to catch on:

The pattern now appears across the web:

• Content farms and search spam: Sites mass-produce LLM-written pages to rank for queries. Google's spam policies define scaled content abuse as generating many pages that add little value for users "no matter how it's created," which targets the volume and purpose of slop rather than the tool behind it.
• Streaming platforms: Spotify said in September 2025 that it had removed more than 75 million spammy tracks in the previous 12 months, and YouTube renamed its "repetitious content" monetization rule to "inauthentic content" in July 2025 to cover mass-produced videos.
• Scientific publishing: An analysis of nearly 951,000 papers found that up to 17.5% of computer science abstracts showed signs of LLM modification by February 2024. Low-effort submissions with fabricated references add strain to peer review.

AI Slop vs. Related Terms#

• Hallucination: A false or unsupported statement in a model's output. Slop describes how content is produced and used, so an accurate but pointless article is still slop, and a carefully reviewed article can contain a hallucination.
• Deepfakes: Synthetic media built to make a real person appear to say or do something. Slop usually chases attention rather than deceiving viewers about a specific person or event.
• Model collapse: The degradation of models trained repeatedly on generated content. Slop scraped from the web into training data is one way that loop starts, which is why data cleaning pipelines for web-scale datasets deduplicate near-identical samples and filter low-quality ones.
• Synthetic data: Generated data created on purpose, filtered, and validated to train models, the controlled counterpart of slop.

Detecting and Filtering AI Slop#

Provenance signals are the most reliable evidence when they exist. AI watermarking embeds a detectable signal at generation time, and Content Credentials (C2PA) attach signed metadata recording which tool made or edited a file, so platforms can label content whose generator cooperated. Neither helps with output from tools that add no mark, and metadata is lost when a file is re-encoded or screenshotted.

Classifiers that judge content after the fact fill the gap with mixed results. Detectors for AI-written text have a poor record: OpenAI withdrew its own text classifier in July 2023 after it correctly flagged only 26% of AI-written text. Images are easier, because generators leave statistical traces; some are visible to people, such as warped lettering or malformed hands, which help readers tell whether an image is AI-generated, and models pick up subtler ones. An Ultralytics YOLO26 classification model trained on labeled real and generated images can flag likely AI images for human review. Such detectors have to keep pace with new generators: a classifier trained on images from one GAN generalized to other CNN-based generators when trained with blur and JPEG data augmentation, yet GAN-trained detectors failed on images from latent diffusion models, and building the detector on CLIP features raised accuracy on unseen diffusion and autoregressive generators by about 26 percentage points.