A cookie banner asks permission. Fingerprinting does not ask, because it does not store anything on your machine — it measures what your machine already is, and no consent dialog governs a measurement.

Glassbox, by David Dale, is a tool that runs those measurements on you and shows you the result. It runs locally: everything happens in the browser, with a single public geolocation call as the only exception.

What it measures

More than 30 fingerprinting techniques, including canvas rendering, WebGL, installed font libraries, WebAssembly feature support, API matrices, cross-site login states, and silent audio sawtooth waves — a technique Dale found in use on AliExpress.

None of these are stored identifiers. They are properties: how your specific combination of GPU, driver, fonts, operating system and browser build renders a shape or processes a tone. Clearing cookies does nothing to any of them, and neither does private browsing.

The numbers

Dale's identifiability estimates run from 56% for Tor Browser to 99% for Chrome.

One work Chrome session came out unique among 7.6 billion browsers.

He caps his calculations at roughly 33 bits of entropy, which is the number that makes the whole subject intuitive. Two to the power of 33 is about 8.6 billion — near enough the world's population. Thirty-three bits of distinguishing information is therefore, in principle, enough to pick one human out of everyone. Each independent measurable property contributes some of those bits, and there are more than thirty of them on offer.

The advice is counterintuitive, and it is the important part

Most privacy guidance tells you to harden your browser: install extensions, block APIs, spoof the user agent, tweak settings.

Dale's position is that this is backwards. In his words, the single highest-impact move is to use a browser that puts you in a large, identical crowd.

The reason is arithmetic. Fingerprinting does not measure how protected you are, it measures how unusual you are. A stock Chrome install on a common laptop resembles millions of others. The same machine with six privacy extensions, a spoofed user agent, a blocked canvas API and a custom font set resembles nothing else in the world — and every one of those modifications is itself measurable.

A hardened browser can be more identifiable than the one it replaced. The user gets a stronger sense of privacy and a weaker version of the thing itself.

Dale advises against customised hardened setups for that reason, recommending Tor or a VPN instead, and separately fixing WebRTC leaks.

Why we are writing this one carefully

Root Notes runs a consent banner. So it is worth being straight about the limits of that arrangement.

A consent banner governs cookies and similar storage. It does not govern fingerprinting, because fingerprinting does not need anything the banner controls. Any site that wanted to could measure the properties Glassbox measures without asking, and a reader who clicked Decline would be no less identifiable.

That is not a defence of the practice. It is the reason a tool like this matters: the compliance layer most people believe protects them addresses a different mechanism entirely.

Built with AI, and said so

Dale notes that tools like Claude Code have made it much easier to polish ideas and offer the useful ones to a wider audience.

A privacy diagnostic built with AI assistance, published for free, running entirely on the user's own machine — set against Veracode's finding that AI-written code carries a vulnerability in roughly 44% of generation tasks. Both are true. The tooling lowers the barrier for the person with a good idea and no time, and it lowers the floor on what ships.

What to do

  • Run it and look at your own number. It is local; that is the point of it being local.
  • Use a browser with a large identical population, in its default configuration. Tor Browser is designed around exactly this idea.
  • Resist stacking privacy extensions. Each one is a distinguishing property. Two or three well-chosen ones are not the problem; a bespoke configuration is.
  • Fix WebRTC leaks separately. They disclose your real address regardless of fingerprint surface.
  • Do not read a cookie banner as tracking protection. Different mechanism.

What is not established

  • How Dale's identifiability percentages are calculated, in detail — they are his estimates, on his methodology.
  • How widely the audio-sawtooth technique is deployed beyond the site he found it on.
  • Whether browsers will respond. Some already randomise canvas output; there is no announced change tied to this tool.
  • How the estimates behave on mobile, where hardware and font diversity are much lower.