Perfection Paradox
Detects suspiciously error-free text. Real human writing almost always contains minor imperfections — their complete absence can paradoxically indicate AI generation.
Technical description
Analyzes text for the complete absence of natural writing imperfections: no spelling near-misses, no grammatical hesitations, no sentence fragments, no self-corrections, no informal register shifts, and perfectly consistent formatting. Measures imperfection density per 1000 words and flags documents where it approaches zero across extended passages.
How it works
Layer 1 (deterministic): Scans for natural imperfection markers: spelling variations, grammatical self-corrections, sentence fragments, register shifts, formatting inconsistencies. Counts imperfection density per section. Flags documents with near-zero imperfection rates across all sections. Higher score = more suspiciously perfect.
Why this matters
Human writers, even experienced academics, produce text with occasional imperfections — a slightly awkward phrase, a sentence that could be tighter, minor inconsistencies in formatting. AI-generated text is often paradoxically 'too perfect' because it has been optimized for fluency. This unnatural perfection is itself a detectable signal.
Score thresholds
- 0-1
- Natural level of minor imperfections present
- 2-3
- Unusually clean but not perfectly polished
- 4-5
- Suspiciously flawless text with zero imperfections
Limitations
Professionally edited manuscripts may be legitimately near-perfect. Short texts have fewer opportunities for imperfections. Multiple rounds of human revision can produce very clean text. This indicator is more useful for draft-stage documents than final publications.