ResAIKit
Research Integrity Toolkit
Back to the encyclopedia
C3Text analysisStylisticLayer 1 (Deterministic)

Artificial Structure

Detects formulaic paragraph patterns where every section follows the same rigid template, suggesting machine-generated organization rather than natural idea development.

Technical description

Analyzes paragraph-level structural patterns by extracting sentence templates (opener type, closer type, transition usage). Detects repeated paragraph skeletons where multiple paragraphs follow identical structures (e.g., topic sentence → evidence → elaboration → conclusion). Measures template diversity using the ratio of unique paragraph structures to total paragraphs. A seventh sub-check flags paragraph-initial connective overuse (consecutive paragraphs opening with Moreover/Furthermore and similar transition words).

How it works

Layer 1 (deterministic): Classifies each sentence by its structural role (opener, evidence, transition, conclusion). Builds a structural fingerprint for each paragraph. Compares fingerprints across paragraphs to detect repetitive templates. Flags documents where more than 60% of paragraphs share the same structural skeleton.

Why this matters

AI models produce text with highly regular structural patterns because they learn to replicate the most common organizational templates. Human writing exhibits more structural variety — some paragraphs are short and punchy, others are extended discussions, and transitions vary organically. Uniform structure across many paragraphs is a strong indicator of synthetic text.

Score thresholds

0-1
Varied paragraph structures reflecting natural thought organization
2-3
Some repetitive patterns but with occasional variation
4-5
Highly formulaic structure with identical paragraph templates throughout

Limitations

Certain document types (lab reports, methods sections) inherently follow rigid templates. Short documents with few paragraphs provide insufficient data for pattern detection. Highly structured writing guides (e.g., IMRAD format) may cause false positives in methods sections.

References

  1. Yildiz Durak H. (2025). Human-written versus AI-generated text in educational discussions: ChatGPT, Gemini and BingAI. European Journal of Education
  2. Liang W, et al.. (2025). Quantifying large language model usage in scientific papers. Nature Human Behaviour