AI Writing Detector + Humanizer

Commercial detectors (Grammarly, GPTZero, Originality) do not read for "AI words." They score two statistical properties: perplexity (how predictable each next word is) and burstiness (variance in sentence length). The pattern list below catches surface tells, which is a real first pass, but it does not predict a commercial detector's number.

The counterintuitive part: polish is a tell. A clean, confident, tonally consistent voice is the modern LLM default, so smoothing your writing can push the score up. The durable fix is real human substance (your own example, a real number, your phrasing), not another restyle.

Detectors are noisy: false-positive rates run 14 to 34 percent on genuine human writing, and the same text can read 60 percent on one tool and 10 percent on another. Read the spread across 2 or 3 tools, not one number. A detector score is not a Google ranking factor.