[Ukv]

> it's used to train their neural networks

Is it? I thought the screenshots were stored and analyzed locally. This seems like something that can be verified with Wireshark.

If you mean they could sneaikily update Windows in the future to start sending screenshots to their server - I feel they could do that regardless of whether or not this local search tool exists, and it'd still get caught almost immediately. If anything, it'd seem counter-intuitive to draw lots of attention/scrutiny through marketing this feature.

[torginus]

Boxer the horse also wasn't sent to the glue factory the first day Napoleon got in charge.

First, they store it locally - then they merge it to your Microsoft account so you can have the Copilot experience anywhere you go, but they are deeply concerned about your privacy.

Then they start processing it, and voila - they have the AI to replace you.

I'm sure they have plans to undertake these steps, each one innocuous enough to not warrant reaction, but they'll nickel and dime you down, and they'll do it too. If they overreach and there's pushback they'll split the steps into two. Hide it, schmooze the regulators etc.

The plan is in motion the only question is the timetable. I bet those GPU farms will be churning using those screenshots come a year or so.

[Ukv]

> I bet those GPU farms will be churning using those screenshots come a year or so.

Would you be willing to make a monetary bet with me, to be resolved 2026-08-23?

[sillyfluke]

not the parent, but I'm curious what you place the odds at. It's a 50/50 coin toss from where I'm standing, and your bet should include you losing if those screenshots are sent one-shot "by mistake" or because of some random minor update. Given the Microsoft related shit-show that happened last month, it's a weird approach to risk if the odds of it happening were even less than 50%, as it would be a low probability but high impact event for those who cared.

I'm curious if you yourself would view the event as a big deal if your data had been sent or if you would simply take the "life is short, who gives a shit?" scenic route.

If you read the article, you would see that the earliest release date for standard Windows versions is planned for early 2025, so you're even kinda baiting the parent from a position of cowardice -- a good faith opening bet would suggest Feb 2026 for the date at the least:)

[Ukv]

> not the parent, but I'm curious what you place the odds at. It's a 50/50 coin toss from where I'm standing, and your bet should include you losing if those screenshots are sent one-shot "by mistake" or because of some random minor update. Given the Microsoft related shit-show that happened last month

Depends a lot on the criteria that torginus and I agree on (if we do). I believe the given scenario itself, Microsoft issuing an update that breaks their guarantee by exfiltrating your snapshots for training their LLMs/etc., is very unlikely. But torginus may argue it's something Microsoft are likely to do in secret and successfully lie about such that lack of admission/evidence is not sufficient to determine it hasn't happened, so the criteria may need to be something weaker about Microsoft having made changes that make it in theory possible for them to be secretly training LLMs on the snapshots (e.g: setting them to store unencrypted in OneDrive).

> I'm curious if you yourself would view the event as a big deal if your data had been sent or if you would simply take the "life is short, who gives a shit?" scenic route.

I think training generative AI on private data would be a huge violation and a big deal. There's the chance of exact regurgitation (bank account details, passwords, API keys), but even without that it's pretty much inherently teaching the model things it should not know and would now be able to talk about.

> If you read the article, you would see that the earliest release date for standard Windows versions is planned for early 2025, so you're even kinda baiting the parent from a position of cowardice -- a good faith opening bet would suggest Feb 2026 for the date at the least:)

Not entirely sure what you mean - the date I proposed (2026-08-23) is a full two years from now. Even from the launch of Windows Recall on non-Copilot+ PCs, if that's what we're measuring from, it should give more than "a year or so".

[sillyfluke]

>Not entirely sure what you mean - the date I proposed (2026-08-23) is a full two years from now.

My bad, I could have sworn I read 2025-08-23.

>I think training generative AI on private data would be a huge violation and a big deal.

Just to be clear, I think a local LLM user input leak is by itself a big enough deal before getting into using it as training data for a public MS LLM. The former is getting hit by a car, the latter is getting hit by a train depending on how bad a "mixer" the public LLM being trained is.

I would take a $100 bet that has me winning if there is a data leak or shown to accessible by a third party or a case where it has been used as training data by 2026-08-23 provided it's released by Jan 2025.

[Ukv]

I think I'm probably more interested in the concerns of novel/systematic abuse around this feature (like a decision to send these snapshots to OpenAI for training), less so in the scenario where there's no change from Microsoft (so files are still stored encrypted locally on-disk) but in some one-off event (malware, 0-day exploits, choosing to sync to Google Drive) a user's files are exposed in the same way their browser's password DB could have been.