Information & Ethics Committee on May 29th, 2019

[Erik Neuenschwander]

Yes.

[Damian Collins]

—or is the data kept by the company to understand what sorts of things people ask? Do you have metadata profiles of people based on how they use Siri?

[Erik Neuenschwander]

The only metadata profile that we have is one that is used to tailor your actual interactions with Siri. For example, we are training our voice models to do natural language recognition on the sound profile. This is really just within the Siri business or the Siri experience. If your question is whether it informs some broader profile used by the company to market products and services, the answer is no.

[Damian Collins]

Mark Ryland, does Amazon do that?

I'd be interested to know the difference between how Amazon uses data gathered from voice compared to how Apple does. There was a recent case of a user who actually put in a request for data that Amazon held. That included a series of voice recordings from their home that the company was apparently using for training purposes. I'm interested in how Amazon gathers data from voice and how it uses it.

[Mark Ryland]

Similarly, the device listens for a wake-up word. It doesn't store any of that ambient data. Once it's awakened, it will begin to stream data to the cloud to do analysis of what the user is actually requesting. That data is stored; it's explicit in the user's profile, and they can see all the utterances. They can see what Alexa thought they said; they can actually see the text that it was translated into. It also gives them some understanding of where there may be some communication issues, and so forth.

They have the ability to delete that data, either individually or collectively. We use the data just as we would use data with other interactions with that person's account. It's part of their Amazon account. It's part of how they interact with our overall platform.

[Damian Collins]

The representative from Apple has said that the device is constantly listening, but only for the Siri command. It would appear that if you have an Alexa device in your home, that is different. The device is always listening and it is actually retaining in the cloud the things it has heard.

[Mark Ryland]

No. It's very similar. We're retaining the utterances after the wake word. It is just like Siri in that regard.

[Damian Collins]

I know from my own personal experience that Alexa responds to commands other than the wake word. It might be triggered by something it has heard in the room that's not necessarily the wake command.

[Mark Ryland]

That sounds like a malfunction to me. It's not supposed to respond randomly to ambient sounds.

[Damian Collins]

Roger McNamee, who gave evidence to us yesterday, discussed how he put his Alexa in a box after the first day he got it because Alexa starting interacting with an Amazon TV commercial. I think most people who have these devices know that all sorts of things can set them off. It's not just the Alexa command or the wake word.

[Mark Ryland]

Well, we're certainly constantly working to refine the technology and make sure the wake word is the way by which people interact with the device.

[Damian Collins]

If you were retaining data from the device that is based on things that it's heard and is then retained in the cloud—which seems to be different from what Apple does—are you saying that it's only sound data that is based on commands that Alexa has been given?

[Mark Ryland]

Yes. It's only the data that is in response to the user's attempt to interact with Alexa, which is based on the wake word.

[Damian Collins]

Would Amazon be in a position to respond to a police request for data or information about a crime that may have been committed in a domestic setting based on sound picked up from Alexa?

[Mark Ryland]

We happily obey the laws of all the countries in which we operate. If there is a binding legal order that's reasonable in scope and so forth, then we will respond to that appropriately.

[Damian Collins]

That would suggest you're retaining more data than just simply commands to Alexa.

[Mark Ryland]

No, the only thing we could respond with is the information that I just described, which are the responses that come from the user once they've awakened the device. There's no storage of ambient sound in the environment.

[Damian Collins]

You're saying that when someone gives the wake word to the device, then the command they've given—their dialogue with Alexa, if you like—is retained?

[Mark Ryland]

That is correct.

[Damian Collins]

You're saying that unless the wake word is given, the device isn't triggered and it doesn't collect ambient data.

[Mark Ryland]

That is correct.

[Damian Collins]

Okay.

I'm interested in the data case I referenced earlier. The concern there seemed to be that ambient sound was being kept and recorded and the company was using it for training purposes.

[Mark Ryland]

No. The reference to training is simply that we improve our natural language processing models using the data that the customers give to us through their interaction with the device. It's not at all based on ambient sound.

[Damian Collins]

It would seem that all of their commands to Alexa that are triggered by the wake word are being retained by the company in the cloud. Do you think your users are aware of that?

[Mark Ryland]

I think so. In my experience with using a mobile device to set up the device at home, I immediately noticed that there is a history icon, essentially, where I can go and see all my interaction with the system.