Photo Scrubber: agent-built experimental tool from a real privacy need

Source. Simon Willison, “Photo Scrubber — local face blur & metadata removal,” September 29, 2026, via his weblog. He says a photograph he took of protesters raised a concern about publishing identifiable strangers’ faces; he had GPT-6 Astra build an experimental tool for face detection and automatic blurring, using MediaPipe’s browser-accessible WebAssembly library and BlazeFace. He presents it as a tool, not a client deployment report.

What it shows. An articulated personal need triggered the artifact; choice to do face processing locally helps explain the implementation constraint. There is no documented early user interaction, comparison to manual redaction, missed-detection/false-blur rates, review of metadata handling or downstream repair in this short post. It therefore belongs beside throwaway prototype versus production, not in a ledger of securely deployed privacy tooling. If followed up, look for repository, test images with consent, handling of undetected faces, privacy audit, code ownership and versioned fixes. Held back: first-person and recent, but seven unread feed items after Dru’s every-item Stratechery subscription; post if later development exposes inspection or repeat use.