aiisms

A lightweight tool that scans text for likely "AIisms" and reports what it found.

Goal

Given a block of text, the tool should:

1. Detect patterns that feel overtly AI-generated.
2. Count how many AIisms appear.
3. Return a readable report with matched phrases and definitions.

Example summary:

• "Detected 6 AIisms"

What we mean by AIisms

AIisms are language patterns that are unusually common in model-generated writing, including:

• overused transition phrases
• verbose boilerplate framing
• hedging or certainty disclaimers used unnaturally
• repetitive sentence structures
• generic, non-committal wording

Initial detection model (v1)

Start with a hybrid scoring approach:

1. Phrase lexicon: weighted list of common AI-ish phrases.
2. Structural patterns: regex checks for repeated templates.
3. Stylistic signals: counts of hedge words, stacked qualifiers, and repetitive transitions.
4. Score + threshold: classify text as low/moderate/high AIism density.

This keeps v1 fast and explainable. Later versions can add an ML classifier.

Proposed output format

• total_ai_isms: integer
• matches: list of matched phrases/patterns
• definitions: brief explanation for each match
• density: low | moderate | high
• aiism_score: non-negative number (length-normalized weighted AIism intensity)
• aiism_ratio: detected AIisms per 100 words

Next steps

1. Build a small baseline lexicon and weighted rules.
2. Implement a simple CLI (stdin or file input).
3. Add fixture texts and expected reports.
4. Tune thresholds against real examples.

Validate taxonomy

Run:

npm run validate:taxonomy

Or validate a custom file:

npm run validate:taxonomy -- path/to/taxonomy.yml