Hacker News Reader

Comment by benmmurphy | Hacker News Reader

why isn't alexa scrubbing the data or anonymising it? presumably amazon is fighting this hard because they don't want it known that the government/amazon effectively has a listening device in your home

View on HN

  why isn't alexa scrubbing the data or anonymising it?

So that if you're a parent and Amazon ships you a doll house, you can figure out whether it was ordered by you, your partner, your child, a TV presenter, or an Amazon mistake.

Shit, it'd be so hilarious if someone made a doll named Alexa, and the TV ad was cleverly crafted to order one for you from Amazon.

This has already happened: http://www.theverge.com/2017/1/7/14200210/amazon-alexa-tech-...

Hahaha. That's great.

apparently they give the ability to access your communication with Alexa with a phone app. So they put themselves in a box by having features in an app that prevent anonymization of the data. Going forward they may have to drop the app or limit its functionality.

I don't have the app in question and am only extrapolating based on the text in the article. Does anyone here have the Alexa app? What are its features?

So is law enforcement going to try to go with voluntarily allowing recording of oneself allows them access? Push a claim that Amazon has no standing in the case and cannot object?

When it's triggered, Amazon stores recordings of every command. You can then play that back, and tell Amazon whether it heard you correctly.

I've used this several times to figure out items on my shopping list where it misheard me.

I have the app, it allows you to play back any prompt that Alexa was triggered with. For instance the maintenance people at my apartment complex decided to play with it and I can listen to them ask stupid stuff like: "how many apples are in a bushel"

Well how many are there?

~120.

Couldn't they encrypt the recordings so that a user could access it but the encrypted data couldn't be read server-side?

Encrypt it with what? The echo? But then how can you listen on your phone? They would need to compromise the UX or store the key themselves, which defeats the whole purpose.

Recognizing spoken word can be improved when you have historical data related to a specific user (or a smaller group of users sharing the device).

Sure, anonymizing it and aggregating much larger numbers helps too. But modeling specific speech of one person, as well as the type of requests they make (to better understand/predict future recognition), can offer better improvements.

Reply on news.ycombinator.com