←back to thread

Today Google bricked my Chromebook by force-installing a hidden extension

(cloudisland.nz)
237 points robin_reala | 2 comments | 29 Mar 25 09:35 UTC | HN request time: 0s | source
Show context
bri3d ◴[29 Mar 25 11:26 UTC] No.43514612[source]▶
The source is available: https://chromium.googlesource.com/chromium/src/+/refs/tags/1...

It’s not “training an AI model on screen contents without consent.”

It is a stupid feature for Google to enable by default: likely what’s making OP’s machine useless is that it’s running an OCR inference model on the OP’s images to index them for search.

Go to chrome://flags and disable “Enable OCR For Local Image Search” and I bet the problem goes away. The AI Service does have a few other features, but that’s the one that’s likely to be cooking the machine.

As for the other comments on this thread, I doubt there’s anything to do with GDPR here. It’s all local.

replies(1): >>43515266 #
xg15 ◴[29 Mar 25 13:13 UTC] No.43515266[source]▶
This also seems crazy to me on a technical level: OP says, this operates on the contents of the screen. So essentially the OS already has the text in memory, renders it to the frame buffer, then OCRs it back - I suppose because it's "cheaper" in dev time to just slap OCR on a screenshot than maybe spend some time looking up what's already accessible in memory and through the UI toolkits.

CPU time is indeed cheaper than dev time, especially if it's your users' CPUs and not yours.

replies(1): >>43515357 #
1. bri3d ◴[29 Mar 25 13:25 UTC] No.43515357[source]▶
Well, kind of, I don’t think it’s that crazy. This service does two things:

* Performs image OCR on images, generically. This is then used for several features: “I type a word in the search box and it can look through my screenshots and photos,” “I’m in one of those horrible scanned image-only PDFs and I want to search,” and so on.

* Performs “main content extraction” on websites by using a screenshot of the website _alongside_ the accessibility tree for that website’s structure. It basically says “given this tree of elements and screenshot, can you prune the tree to just the elements a user would care about.” The fact that this is necessary is more an indictment of the DOM than this feature, IMO :)

replies(1): >>43515384 #
2. xg15 ◴[29 Mar 25 13:30 UTC] No.43515384[source]▶
Ah ok, that seems more sensible. My understanding was it would pretty much literally make screenshots and then run OCR on them. If there is enough smartness to only run OCR on the parts that have no text information, it makes more sense. (Though as we see here, even that approach can be too much)