WEBVTT

00:00.000 --> 00:01.660
Can a company read your stuff?

00:02.020 --> 00:05.120
If the company holds the keys to your data, it can scan your stuff.

00:05.380 --> 00:09.640
Watch out for terms like encrypted at rest and encrypted in transit.

00:09.840 --> 00:12.640
That just refers to technology that's used to encrypt your data.

00:12.880 --> 00:15.320
This is different from something like a zero-knowledge provider

00:15.320 --> 00:18.820
or an end-to-end encrypted provider that directly speaks to the fact

00:18.820 --> 00:21.720
that they don't control the encryption keys you do.

00:21.980 --> 00:23.200
So that's the simple question.

00:23.580 --> 00:25.780
Can the company read this?

00:25.960 --> 00:26.920
If the answer is yes,

00:26.920 --> 00:31.340
then they might be able to scan your data in ways that you don't consent to doing.

00:31.560 --> 00:33.400
In terms of the photo scanning in particular,

00:33.640 --> 00:35.020
there's a few ways that this can happen.

00:35.180 --> 00:37.680
The first method, which is matching against a known database.

00:37.940 --> 00:42.920
So every photo gets fingerprinted and compared against a list of known illegal images.

00:43.140 --> 00:45.020
It sounds precise and it mostly is,

00:45.100 --> 00:48.240
but it means every single photo you send or back up is getting checked

00:48.240 --> 00:52.020
always forever, which has obvious privacy implications.

00:52.320 --> 00:54.960
The second thing is AI guessing at new images.

00:54.960 --> 00:57.540
This is the one that really freaks people out as it should

00:57.540 --> 00:59.160
is this is still experimental technology.

00:59.160 --> 01:01.320
There's no database to match and guarantee it.

01:01.440 --> 01:05.320
It's just a model that looks at a brand new photo and guesses as a machine,

01:05.480 --> 01:06.820
whether it's abuse or not.

01:06.920 --> 01:08.960
The third type of scanning is a bit more indirect

01:08.960 --> 01:11.540
and it can target actual conversations and messages.

01:11.840 --> 01:13.240
This is things like grooming detection.

01:13.480 --> 01:14.680
It doesn't really match anything.

01:14.860 --> 01:16.960
It just reads words and judges intent.

01:17.300 --> 01:21.080
Things that can obviously trip this up are sarcasm to adults flirting,

01:21.260 --> 01:23.880
apparent texting about picking up their kid from swim practice.

01:23.880 --> 01:25.820
All of it's going to go through a model that decides

01:25.820 --> 01:27.440
whether you sound like a predator or not.

01:27.680 --> 01:29.840
Now, these are the actual technologies that are being implemented

01:29.840 --> 01:32.320
and endorsed by things like chat control.

01:32.580 --> 01:33.240
But here's the thing.

01:33.400 --> 01:35.440
The actual numbers don't paint a positive picture.

01:35.700 --> 01:39.340
Germany's federal police said that 48% of the alerts they received

01:39.340 --> 01:41.800
from these programs aren't criminally relevant.

01:42.040 --> 01:47.960
That means nearly half 99,000 wrong reports in 2024 alone.