Detect AI Text — Self-Check

A pre-flight audit that scores a draft for AI tells and tells the user exactly what to fix. No external tool required — this is a structured read, not a detector.

When to use

• The user asks whether a draft "sounds like AI" or "will get flagged".
• Before submitting or publishing anything that was drafted with AI help.
• As the final gate after the humanize-ai-text skill, to confirm the fixes landed.

What real detectors key on

AI detectors (GPTZero, Turnitin, Pangram, Originality.ai, and others) mostly react to two statistical properties plus learned patterns:

• Low perplexity — the text is too predictable; every next word is the obvious one.
• Low burstiness — sentence length and complexity barely vary; humans spike and dip.
• Learned fingerprints — modern detectors (e.g. Pangram) are neural nets trained on millions of AI documents, so they also catch structural habits, not just statistics.

You can't compute perplexity by hand, but the symptoms are readable.