Slop-marker convention
- title
- Slop-marker convention
- type
- concept
- summary
- The idea of an explicit, machine-and-human-readable marker declaring code (or text) as slop — defined by its production process, not its surface
- tags
- agentic-coding, llm-skepticism
- created
- 2026-07-21
- updated
- 2026-07-21
The inverted convention — declaring human authorship rather than machine — is human-made-disclosure, argued for in no-ai-statements. Same structural move, opposite claim, and a different incentive problem: a slop marker costs the producer something to attach, a human-made marker is free to fake.
A slop marker is an explicit signal a producer attaches to an artifact to declare that it wasn't made for unwitting human consumption. The canonical proposal is William Woodruff's READMENOT file, but the concept is name-agnostic: any consistent marker (a file, a header, an @generated-style comment) that lets a reader make an informed decision before spending attention on machine output.
The definition that makes it work
The marker is only coherent if "slop" is defined by how the artifact was produced, not by how it reads. Woodruff's two sufficient conditions: an artifact is slop if it either (1) was developed primarily without human supervision, or (2) reflects a fundamental lack of operator understanding. Neither is visible from the surface — which is exactly why a marker is needed. You can't reliably detect slop by reading it (see credibility-as-slop-test: fluent LLM output looks like effort that never happened, and a single em dash can get careful writing dismissed as slop). If detection worked, no marker would be necessary.
Why production-process, not surface
This is the load-bearing move. Surface-based slop detection fails in both directions, so any convention built on "does it look like slop" is doomed. A marker sidesteps detection entirely by moving the declaration to the one party who actually knows the production process — the author. It's the same structural answer as the credibility-as-slop-test (author stakes their name) and the human-sized-artifact countermeasure in agent-principal-agent-problem (force a small thing a human vouches for): when the signal can't be recovered from the output, you have to get it from the producer.
The precedent
Generated code already has this: the @generated marker (and tools like linters/review bots that skip such files). Nobody objects to generated code existing; the friction is reading it unaware. Slop is the same category one level up — machine-produced artifacts where the honest move is to flag provenance so readers opt in knowingly. The marker doesn't say "never read this"; it says "here's what you're getting into." Security research is the standard reason to read flagged slop anyway.
Limits
- It's voluntary and honest-system. It helps people who want to signal honestly; it does nothing about slop dumped without a marker, which is most of it. The value is giving honest producers a way to express intent, not enforcing disclosure.
- The target moves. "2026's slop is not 2025's slop." As models improve, the line drifts, so a marker is a point-in-time claim, not a permanent verdict.
- It presumes the producer can tell. Same objection raised against the credibility-as-slop-test: if LLM output is uniformly average, an author acclimated to that average may not recognize their own artifact as slop.
Related: llm-as-average-democratizer (why the effort-signal collapsed in the first place), cult-of-vibe-coding, reviewing-ai-code (provenance changes how you should review).