• Photo Scrubber detects faces and blurs them in the browser using MediaPipe compiled to WebAssembly plus the BlazeFace model, according to its author.
• The author describes it as experimental and says GPT-6 Astra built it; the post publishes no recall or false-negative figures.
• Google documents BlazeFace as a lightweight detector optimized for mobile GPU inference, with short-range and full-range variants for different camera distances.
• The post gives no technical detail on how metadata removal works, so that half of the tool's claim cannot be checked from the source.
• Treat it as a lead for a reproducible test, not as a guarantee that every face in a crowd photo will be hidden.

Simon Willison has published Photo Scrubber, an experimental tool that blurs faces and removes metadata from photographs. He says he built it after photographing protesters and deciding he did not want to share images of strangers with identifiable faces. It is a useful design for a privacy problem, but the source offers no measured detection rate, so the verdict is Signal for the approach and unproven for reliability.

What the tool is made of

According to the post, Photo Scrubber uses Google's MediaPipe C++ library, compiled to WebAssembly and distributed as @mediapipe/tasks-vision, together with the BlazeFace face detection model. The author says he had GPT-6 Astra generate the tool. Google's own documentation lists Web as a supported platform for its face detector task, so the stack matches a documented path rather than a custom model. The post describes the tool as local, which keeps the photograph out of any server.

Why BlazeFace's documented design matters

Google describes BlazeFace as a lightweight detector optimized for mobile GPU inference, with a short-range variant for selfie-like images and a full-range variant for images like those from a back-facing phone camera. A crowd photo at a protest is a harder case than either description. The documentation does not publish recall for small or partly occluded faces in that setting, and the post does not say which variant the tool loads.

What cannot be verified from the source

The post gives no benchmark, test set or false-negative count, and it offers no technical detail on how metadata is removed. A face blurrer fails silently: one missed face leaves an identifiable person in the published image, and a metadata step that leaves location tags would expose the photographer. Neither failure is demonstrated or ruled out in the source, so any reader relying on the tool is relying on an untested claim.

What to do with it

Treat Photo Scrubber as a good first pass, not a guarantee. Run it on a set of your own images with small, angled and partly hidden faces, count the misses, and inspect the output file's metadata with a separate tool before publishing anything sensitive. Anyone photographing protests should review the blurred result by eye. The architecture is right for privacy; the reliability is something each user has to measure.