Signals of Slop

Slop isn't limited to low-effort AI-generated content. Humans can also create slop. This summary lists signals of slop: traits that make something slop, independent of the level of effort spent or the degree of AI involvement.

Repetition

People often identify artifacts exhibiting popular patterns as slop, regardless of whether they consider those patterns intrinsically good or bad. Repetitiveness is considered a signal of slop for at least three reasons:

1. People experience fatigue from popular characteristics like common verbal tics or clichés.
2. Some creators copy superficial traits to ride the popularity of another work without understanding what made it good or coming up with something more distinctive themselves.

Bruceploitation is a good case study of imitation as a subtype of repetitive slop.

A subgenre of exploitation film, Bruceploitation films sought "commercial success by capitalizing on current trends": often casting Bruce Lee look-alikes with screen names like "Bruce Li" or "Bruce Le."

1. Slop tends to have less internal variance, which leads to it becoming boring faster.

Lack of intentionality

Slop contains details that are chosen arbitrarily: little thought goes into why those particular details were chosen instead of an alternative. You can check for this with the swap test: pick a random detail and ask if the meaning of the work as a whole would change if the detail were different. For instance, in a well-written story, changing the name of a character also changes the meaning of the story.

Low information density

A small amount of novel information is expanded into something much longer. For instance, a long email might be generated from a short prompt. Finding the substance is more difficult due to the amount of filler used.

Increased cognitive effort to understand

Slop often requires additional interpretive effort on the part of the reader. Examples:

• Fiction could rely on the reader's imagination to fill in gaps without directly evoking it
• Pull requests could rely on maintainers to check for bugs and code style
• An email could be unnecessarily long

Pollutive

Many intellectual and creative commons implicitly relied on proof of work to show what work to filter in to be reviewed and rated. Those signals are easy to fake (and increasingly easy to spot) now. Since it's cheap to apply to any given job, every job description gets tens of thousands of applications, even if each applicant doesn't believe themselves a particularly good fit.

Lack of domain understanding

Choices being made without understanding the context within which they're going to be used. People dislike this because it abdicates a responsibility under the social contract between creators and recipients: that the work reflects an understanding of how things are done.

Case study: Nikolaos Balanos' attempt to restore the Parthenon

The ancient builders connected the masonry with iron clamps, waterproofed by sheathing them in lead. Balanos connected the masonry with iron clamps sheathed in cement, which he incorrectly believed would also make them waterproof. Water seeped through the joints of the masonry and the cracks of the cement, which corroded the iron. The rust expanded and split the marble.

Eating your own tail

Slop is often based on repeatedly reprocessing the same inputs that other works - often other slop - use, without gathering new information.

Case study: Did 38 witnesses watch the murder of Kitty Genovese?

In 1964, a police commissioner made a remark to an editor over lunch that 38 witnesses watched on as she was murdered. This inspired social psychologists to start a research agenda on the bystander effect. In 2007, a peer-reviewed paper found that all ten psychology textbooks sampled repeated some variant the claim that 38 witnesses stood by the murder and did nothing and and found no evidence for it: they only found evidence for "half a dozen" people visually identified Kitty and her assailant, 3 of whom didn't recognize the incident as an ongoing murder, 2 of whom called the police, and 2 of whom intervened.

Information that appears well supported by multiple sources may appear accurate while actually being fabricated, which escalates as more credible sources make the same claim without checking, contributing to a notable snowball effect.

Thanks to Max Spero (as CEO of Pangram) and Rika (for her expertise with comparative media analysis) for reviewing various drafts of the linked post and providing feedback that I incorporated! Lyn helped me review the literature and structure the post. The boid simulation depicted in the thumbnail was written by GPT-6 Astra with my guidance. (I authored the thumbnail's layout, crop, and visual identity.)