This seems like a bug in the ScreenAI service? There's no evidence whatsoever for his claim that Google "trains a machine vision model on the contents of my screen".
> These functionalities are entirely on device and do not send any data to network or store on disk.
There is also this description in the Chrome OS source code:
> ScreenAI is a binary to provide AI based models to improve assistive technologies. The binary is written in C++ and is currently used by ReadAnything and PdfOcr services on Chrome OS.
According to https://chromium.googlesource.com/chromium/src/+/main/servic... it is just inference.
> These functionalities are entirely on device and do not send any data to network or store on disk.
There is also this description in the Chrome OS source code:
> ScreenAI is a binary to provide AI based models to improve assistive technologies. The binary is written in C++ and is currently used by ReadAnything and PdfOcr services on Chrome OS.