the anatomy of AI slop

a word worth knowing

In May 2024, Simon Willison wrote down a definition the industry adopted within weeks: slop is unwanted, unreviewed AI-generated content, the way spam became the word for unwanted email. His personal rule travels well: "I'm happy to use LLMs for all sorts of purposes, but I'm not going to use them to produce slop. I attach my name and stake my credibility on the things that I publish."

Read that back slowly. The word describes a publishing decision, and the decision belongs to you. An agent generates; whether the output becomes slop depends on what stands between generation and shipping. This module builds that layer, surface by surface: code, design, copy, accessibility, speed.

the numbers behind the feeling

Slop in code has been measured at scale. GitClear's research across 623 million changed lines from 2023 to 2026 tracks what happened as AI authorship grew: duplicated code blocks rose 81%, within-commit copy/paste rose 41%, error-masking constructs rose 47%. On the reuse side, refactoring line moves fell 70% and cross-file function calls fell 35%. Their earlier report marked 2024 as the first year on record where copy/pasted lines outnumbered moved lines.

The throughput gain is real, and so is the debt. GitClear's phrasing for the risk lands on the operator, and this curriculum agrees: the danger concentrates "among developers who haven't recognized the failure modes." Recognizing them is a skill. You already hold half of it from module 1's verification habit and module 4's review gate; this module extends the same move to every surface a user touches.

where the gates go

Each lesson in this module installs one gate:

• Tests give the agent a check it can run, so correctness stops depending on your attention.
• Design law replaces taste with written tokens and anti-patterns, so the UI stops regressing to the template mean.
• Copy checks catch the measurable tells of machine prose before a reader does.
• Accessibility and speed budgets gate the two qualities almost every site fails.
• The closing lesson wires all of it into one pipeline and proves it catches a planted bug.

Slop grows in the gap between "the agent said done" and "someone checked." Five lessons from now, that gap on your project is a set of scripts.

This platform runs the same gates on itself. When a sweep of the whole curriculum looked for machine-prose patterns, the findings were concrete enough to grep:

exercise

Audit this repo for shipped-without-review artifacts: find one duplicated code block a review should have merged, and one user-facing text nobody proofread before it went live. Show the evidence, file and line.