Watch this episode on YouTube.
Reasonably Accurate 馃馃 Transcript
Morning everybody. How are you doing today? On this episode of the show, we're gonna talk about the latest news that Facebook has humans reviewing your audio messages in messenger. Now, I had a great talk this morning um on my CBC radio column about this very issue because it's raising a lot of concerns and there's a lot of misconceptions around what's actually happening, why it's happening and what you should be concerned about.
Now, here's the reality of what's happening. Facebook was the latest to announce that they had humans reviewing audio trans from messenger. So if you send uh an audio file back and forth, there is a chance that a human is gonna review that at some point.
Now, no word on whether they're transcribing uh voice calls made over messenger, but it's probably likely as well. Now, that's disconcerting to a lot of people, you figure when you're sending a message back and forth to messenger that it's private, it's not really the case and we're going to dive into that.
Now, this follows a series of announcements from Amazon, from Microsoft, from Google and from Apple saying that they were doing the same thing for their voice assistants. Now that's a different case and I'll tell you why. So with the voice assistant, you use a trigger word, you say, hey Siri Alexa, ok.
Google. And I know like anybody else. I have no idea what Cortana trigger word is because who uses Cortana. But at least there, there's an explicit interaction. I am requesting that the device listen to my next few words. Now there's a privacy issue there around accidental triggering.
Um And that's a problem but those who are those voice, a percentage of those voice commands are sent for review to look for accuracy, to see if they're actually being interpreted correctly. And you can actually look at that up in Amazon, Alexa's app, you can go to your activity and you'll look and it'll say Alexa Herd and then you can say was this right?
Yes or no, right. And that makes sense. That's an explicit interaction of me saying, hey, uh you know, hey Siri, what's the uh you know, next game for the W NBA? And it will tell me the response, but there's an interaction there that's expected, I think on both sides.
And that's OK where Facebook really crossed the line here is in these audio messages. So if I'm messaging with you on Facebook messenger and I send you an audio message, there's a chance that that could be transcribed similarly with Microsoft and Skype translations, those were being sent for accuracy.
Now, from a technical perspective. I totally understand why this is happening. This is happening because they want to verify the accuracy of the machine transcription. So they have a machine learning model that takes the text and creates or takes audio and creates text from it.
I do that with a show manually through either GCP or through AWS and I get, send an audio file, get the text back. Now, Facebook is doing the same thing in the background on messenger and that should shock nobody. They're looking for behavioral signals so they can better target business services and advertisements towards you.
But the challenge is direct messages. People just assume they're private even though they're not, there are some exceptions, whatsapp, which is under Facebook, but it's still end to end encrypted for now. And iMessage when you're talking to an either iMessage user is also end to end encrypted.
And that's an interesting contrast because iMessage is obviously Apple's big foray into messaging. And the goal for iMessage is to sell more devices to lock people in. I read an interesting article last week where teens were kicking non iMessage users out of iMessage chats because they didn't like the color change that comes with an S MS user and that has an ancillary privacy and security benefit is that that chat goes back to being end to end encrypted amongst the group participants.
And whatsapp is still end to end encrypted by D for everything on the platform, which is great. Um But we'll see what happens when Facebook migrates everything to their unified chat because they've said they're going to do that, combine Facebook messenger, Instagram, messaging and whatsapp together.
And while Facebook messenger technically supports end end encryption, nobody I have ever asked actually knows that or knows how to enable it under the secret conversation mode. And that's what's really important here is defaults, it needs to be default and then encrypted, but also the expectation.
And that's the key issue here is that the expectation for an interaction with uh with a voice assistant is that the computer is listening to your voice. That's how it works. And the fact that somebody might manually review that to improve the model isn't as jarring or shocking.
There's still some concerns but it's not jarring or shocking for messenger and any other direct messaging thing. People assume these are private, they are not, they are private message between you, your intended recipient and the company that's hosting them, never forget that. And people do so when Facebook was revealed that they were doing this, it was extremely jarring even though it was covered under the privacy policy in the terms of service, it was in such broad and flowery language, essentially saying that any content or use of the services may be monitored by ourselves or a third party blah, blah, blah, they were legally covered, they didn't do anything legally wrong here.
And technically you have agreed to this by agreeing to the terms and service. But we all know that doesn't really count for anything. And I think that's key takeaway from today's episode is what's user expectation for direct messages. The expectation is that they're private, they're not mined, they're not monitored, they're not transcribed and they're actually secure.
That's far from the truth on most platforms. But for voice assistance is a little more liberal because you are actually talking to the system, you are having this voice transaction, you're expecting some sort of transcription so that the system can interpret your response.
The fact that a human was reviewing it still a little bit jarring but not nearly as much as the messaging case. Let me know what you think. Hit me up online at Mark NC A in the comments down below. And as always by email me at Mark N dot C A, have a fantastic day, we'll talk to you in the next episode of the show.
