About the Project — Tell Me Human

Inspiration

We increasingly read AI-generated text in emails, forums, and social feeds, yet verifying and adjusting that content is still clunky: copy/paste to another tab, bounce between tools, and lose context. Tell Me Human fixes this in situ: detect likely AI-written text, check pertinence/risk, and “humanize” it in one click—without leaving the page.

What it does

AI-likelihood score (0–100) with concise reasons (probability, explanation[]).
Pertinence/safety check that flags doubtful claims, tone issues, or policy risks.
One-click humanization that preserves meaning while improving rhythm and naturalness.
Side Panel and Context Menu entry points; English/Spanish support.
Privacy-first: uses Chrome AI on-device when available; optional OpenAI/Gemini fallbacks with user keys.

How we built it

MV3 extension with service worker (background.js), Side Panel (sidepanel.html/js), and settings in chrome.storage.sync.
On-device Chrome AI: analysis/rewriting through LanguageModel (system prompt init) and, when present, window.ai rewriter/proofreader.
Fallbacks: optional OpenAI (/v1/chat/completions) and Gemini (REST) toggled by the user.
UX: SweetAlert2 for feedback, an animated gauge for the score, and i18n bundles (en.json, es.json).
Prompt hygiene: all analysis responses are coerced into single-line JSON to avoid formatting pitfalls.

Challenges

Side Panel messaging + Chrome AI quirk: a chrome.runtime.Port used to track Side Panel state sometimes closed right after invoking Chrome AI; we switched to chrome.runtime.sendMessage + state in chrome.storage to keep the panel in sync.
Environment variability (LanguageModel/window.ai not always available): capability detection + clear guidance to enable flags or use fallbacks.
Detection (\neq) verdict: communicate probability, not certainty, with actionable explanations.
API errors/rate limits: circuit breakers, provider validation, and cleanup of invalid tokens.