HumanMark

Flags AI-generated text on any webpage with confidence scores.

HumanMark spots likely AI-generated text on any webpage and shows you a confidence badge so you can read with context. How it works • Scans visible text blocks on the page (posts, comments, articles, search snippets, summaries). • Scores each block with a fast offline heuristic — sentence-length variance ("burstiness"), vocabulary uniformity, punctuation patterns, and platform-specific tells (LinkedIn 3-bullet-plus-CTA cadence, Strava boilerplate motivational phrases, Twitter low-emoji contraction-free patterns). • Optional API mode: bring your own GPTZero or Sapling key for higher accuracy on longer text. Opt-in, gated behind an explicit consent checkbox. What you see • Red outline + 🤖 badge for high-confidence AI (≥ 75% by default). • Amber outline + ⚠️ badge for uncertain (≥ 50%). • A floating HM pill (bottom-right) with a per-page count and prev/next arrows to jump between flagged items. • Click any badge for a details popover: confidence, source, timestamp, and a "Dismiss" button. What you control • Global on/off and per-site toggle in the popup. • Deny-list any domains in Settings. • AI / uncertain thresholds, minimum text length, badge colors (with presets). Privacy • Heuristic mode is fully offline — zero outbound network calls. • API mode only sends text to the provider you choose, only after you paste a key and accept the consent notice. • No analytics, no telemetry, no cookies, no remote code, no phoning home. • Settings, cache, and stats live in chrome.storage.local, sandboxed to the extension. The cache stores hashes and scores only — never your text. https://github.com/ssoward/ideas/tree/main/humanmark