DetectAI

What It Does

Answers two questions — how much does this read as machine-generated? and does it carry an embedded mark? — with four independent measures:

• Heuristic audit. Flags known AI tells (inflated vocabulary, the "not X, it's Y" tic, recycled transitions, uniform rhythm) against a severity-tiered pattern catalog. Free, instant, and it explains why each flag fired.
• Statistical signals. A deterministic pass (LIFEOS/TOOLS/StatSignals.ts) measuring the keyless distributional tells the research literature rates real: n-gram entropy, type-token ratio, repetition structure — plus the weak-alone folklore tier (burstiness, paragraph uniformity, function-word stylometry), each labeled with its reliability. Features, never verdicts (arXiv:2310.15264: paraphrase degrades every keyless statistic). Free, no key.
• Empirical score. Runs the text through the Pangram detection model and returns a real probability — AI% / AI-assisted% / human% — plus per-segment counts. Costs money, needs an API key, and doesn't care what your word list says.
• Watermark scan. A keyless, deterministic pass for character-level covert channels — invisible chars, variation-selector/Tags-block steganography, homoglyphs, bidi, odd whitespace. Catches embedded marks that live in the bytes; by design it cannot read sampling-time statistical watermarks (SynthID, Kirchenbauer, Anthropic's announced mark), which are key-gated. Free, no key.

The measures disagree often, and that disagreement is the useful part. Text can clear every pattern on the list and still score 100% AI, which tells you the tells are structural, not lexical — and a watermark hit is bytes-level proof regardless of what the other two say.

The Problem

"Does this sound like AI?" gets answered by vibes, and vibes are wrong in both directions. Heuristic word-lists flag legitimate writing and miss AI text that avoided the obvious words. Detector scores look authoritative but saturate — Pangram will confidently call a short human paragraph 100% AI. Neither measure alone is trustworthy, and a raw number with no baseline is close to meaningless.

This skill runs both, and anchors the empirical score against known-human writing so the number has something to be read against.