Research: AI Slop Patterns

AI Slop is the "synthetic friction" left behind during the rapid iteration cycles of agentic coding. It isn't just "bad code"—it's code that carries the hallmarks of its generation process.

1. The Heuristics Library

The vibe-check engine maintains a library of known patterns.

A. Conversational "Fingerprints"

These are regular expressions used to sweep comments and documentation.

B. Structural "Slop"

Patterns that emerge from LLM prompting styles.

• "The Fragmenter": AI models often extract logic into small, descriptively named functions that are only used once.
  • Heuristic: Detect functions $< 10$ lines with $1$ call-site and $0$ sibling reuse.
• "The Filler": Redundant comments that repeat the function name.
  • Heuristic: Jaccard similarity $> 0.8$ between function name and docstring content.

---

2. Evolution of Slop

We categorize slop into three evolutionary stages:

Stage 1: The Obvious (2023-2024)

• Left-over conversation headers.
• Plain-text apologies from the model in the comments.
• Hallucinated imports.

Stage 2: The Subtle (2025)

• Consistent over-documentation.
• Generic, overly-safe variable naming (processed_data_output).
• Performance-blind abstractions (e.g., unnecessary deep clones).

Stage 3: The Fraudulent (2026+)

• Intentional "human" errors to mimic manual work.
• Masking synthetic patterns with slang or informal comments.
• Detection: Requires Entropy Analysis to find "too perfect" probability distributions.

---

3. The Baseline Vector Store

To detect subtle Stage 2 slop, we use Vector Comparison.

1. Source: We index a "Trusted Baseline" (e.g., the Linux kernel, or a project's manual MVP).
2. Analysis: We compare the "Distance" between a new PR and the baseline.
3. Threshold: If the PR moves into a "Generator-Specific" cluster in the vector space, it flags for review.