Privacy-first AI — runs on your device, not in the cloud

Most AI tools upload your text to a server, run a model there, and send the answer back. Web Highlights does the opposite: the model sits on your computer, inside your browser. Your highlights, pages, and questions never leave your device.

Get Web Highlights Try it free — no signup
Your device Your page Gemini Nano No cloud, no servers Your content never leaves your device

Every AI feature works this way

AI Summary

The page you're reading is handed to the model in your browser, and the summary is written right there. Nothing to upload — the text was already on your screen.

AI Chat

Ask questions about any page or PDF. The conversation happens between you and the model on your machine — not with an API in someone's data center.

AI Flashcards

Flashcard decks are drafted from your highlights on-device. Your study material — often the most personal content of all — stays yours.

How on-device AI works

  1. 1

    Your browser downloads the model — once

    Chrome ships a small language model called Gemini Nano (Microsoft Edge uses Phi-3). The first time you use an AI feature, the browser downloads it — roughly 3–4 GB, one time.

  2. 2

    The model runs inside your browser

    From then on, every summary, chat answer, and flashcard is computed by your own CPU/GPU. There is no Web Highlights AI server — we couldn't read your content if we wanted to.

  3. 3

    Nothing is uploaded

    No accounts, no API keys, no per-request costs, no prompt logs. You can verify this yourself: open your browser's DevTools network tab while generating a summary — no request carries your page anywhere.

  4. 4

    It even works offline

    Because the model lives on your device, AI features keep working without an internet connection once the model is downloaded.

The honest fine print

On-device AI is a trade: you get privacy, but the browser sets the rules. Here is what that means in practice:

  • You need Chrome 138+ or Microsoft Edge on a desktop computer — mobile browsers don't ship the model yet.
  • The one-time model download is about 3–4 GB. Chrome only starts it if you have about 22 GB of free disk space — headroom it wants on the volume, not the size of the model.
  • Browsers only expose their built-in AI on secure (https://) pages. On a plain http:// page the AI features pause — your browser is fine, the page is the blocker.
  • A small model on your laptop is not GPT-sized. For summaries, chat about a page, and flashcards it does the job — that's exactly what we use it for.

Something not working? Read the Chrome AI troubleshooting guide

Frequently Asked Questions

Does my data ever leave my device?
Not for AI. The page content, your highlights, and your prompts are processed by the model in your browser. What DOES sync (if you create an account) are your saved highlights and bookmarks — that's the normal sync feature, separate from AI, and described in our privacy policy.
Which AI model does Web Highlights use?
Chrome's built-in Gemini Nano. On Microsoft Edge, the built-in model is Phi-3. We use whichever your browser provides — we don't run any model on our own servers.
Which browsers support it?
Chrome 138+ and recent Microsoft Edge, on desktop (Windows, macOS 13+, Linux). Mobile browsers don't ship an on-device model yet — the AI buttons stay hidden there until they do.
Why do I need to download a model first?
That download IS the privacy feature: instead of sending your text to a server that already has a model, the model comes to you. It's about 3–4 GB, happens once, and the browser manages it.
Does it work offline?
Yes. Once the model is downloaded, summaries, chat, and flashcard generation work without an internet connection.
Why doesn't the AI work on http:// pages?
Browsers only expose their built-in AI on secure (https) pages — a platform rule we can't override. Open the page over https, or open your highlights in the Web Highlights dashboard, and the AI features work again.
Is it free?
The free plan includes a daily allowance of AI actions (for example one generated flashcard deck per day); paid plans lift those limits. Since the model runs on your device, there are no per-token costs hiding anywhere.
View all frequently asked questions →