Has there been ANY hint of a rationale for linking the user’s details to their recordings?
Or why they couldn’t have asked users if they would anonymously share data to help improve things? (Don’t MS and Apple do something similar with how their users use their software?)
Again there is some contradictory information here about what exactly is being recorded and uploaded, the article talks about recordings of someone singing in the shower or a child screaming for help and then later talks about a wake word but it’s listening constantly, right? If it’s listening constantly then that is surely being uploaded to amazon’s servers or these employees wouldn’t know about someone singing in the shower or a child screaming for help.
Most likely the voice recognition software is mischaracterizing some sound for the “wake up” word, then recording whatever noise follows until the algorithm times out due to silence or elapsed time.
Only thing that irked me about that is the likelihood that it contributes to the unnecessary use of overly difficult captchas.
I think it’s best to treat these devices like you should treat every mic - assume they’re always live.
Any device that recognizes a “wake word” or “activation word” has to be listening all the time, because otherwise it can’t recognize the wake word.
However systems that have all the “smarts” elsewhere (say Siri, Echo, or Google Assistant) can do all the recognition of the “wake word” locally, and not send anything unless it hears that word. It can be useful to send some of the sound from somewhat before the wake word (did you say “Hey Siri turn on the kitchen lights”, or “Hey Siri turn on the …dammit! Didn’t beep! Hey Siri…HEY SIRI turn on the kitchen lights”?). I don’t think it is right to send that “extra” audio without making it clear that it will happen.
Singing in a shower may sound enough like the activation word to an AI…especially if the “listening device” is fairly far away. Likewise screaming may be clearly not “Alexa” to a human, but to an AI that is really only trying to recognize 3 words and the training data did NOT include a child screaming (“don’t match this, it is a small child screaming about broccoli”) well stranger things have happened.
It would probably work out better for people if they stop thinking about what TV and movies have shown us for AI, and start thinking of AI as “some modestly smart animal trained with treats to do some stuff”. So you won’t be shocked when Alexia responds to some really odd phrases the same as my dog does if I ask my wife if we have chicken soup (I discovered she knows the word “chicken”, it means “treat” every bit as much as all the words I bought her on purpose). It isn’t a great model because AI/ML as currently deployed has had all the learning pre-baked in, and the learning is not just “turned off” it doesn’t actually even exist in consumer deployments.
I am aware of this but the question is if it’s uploading everything it’s been listening to until the wake word is used and that appears to be yes, though i’m not sure most people realise that. Whether they are storing it all in perpetuity is another question but i would not be surprised if they are.
This is a good point because all the marketing bullshit around the supposed “AI” of these devices is just that and again the punters not realising there’s a whole boatload of low paid workers managing this stuff. It’s like that concept of the automated supermarket back in the 30s/40s - enjoy this futuristic shopping experience but don’t worry yourself about the low paid, unskilled workers toiling away behind the walls. But people have bought in to the marketing spiel yet as you point out, any true AI development certainly isn’t being used in a cheap electronic toy.
I’d even go as far the final scene of The Conversation.
I worked in audio quality for phone systems way back then. IIR, we were allowed to listen in on conversations, but only for a couple of seconds at a time. We had a recording that was made up of hundreds of 2-second snippets of conversations to run through compression algorithms.
"The Register asked Apple, Microsoft and Google, which all have smart search assistants, for a statement on the extent of human involvement in reviewing these recordings – and their retention policies.
None would disclose the information by the time of publication."
“Hi, I’m from Amazon. We’re trying to make our products better, and we need lots of different voices saying various words to improve their voice recognition. Would you be willing to say a few words for me today so I can record them and help improve our Alexa line of products? We’d be happy to give you a small credit on your Amazon account for your troubles.”
See how easy that is? No need to misguide consumers, simply ask and offer them something in return for their troubles.
Note: I also avoid these products like the plague given these issues. So far as I know, Siri on my phone requires me to press a button for a few seconds to activate. If I ever find out its just listening in on me, that’ll be gone, too. I value my privacy too much to trust any corporation to listen in on me for any reason. Corporations have NEVER proven themselves to be trustworthy at ANY time in history.
And this is why I only use Siri to test my timer when I’m cooking. Like honestly, I can’t find any other use for voice activated tech other than maybe hands free driving and what I mentioned before.
Also:
They ask your consent to share crash logs, and they do (at least superficially) anonymise the data, in that it won’t explicitly report your IP or user name. You would be foolish to assume that a dedicated spy couldn’t find things out about you if they got hold of that data, but overall it is a relatively low risk, and really does increase the odds of bugs getting fixed.
They don’t collect usage data (“heat maps” etc), or at least, not under that rubric. There is no way to do that that plausibly protects your privacy, because it pretty much requires recording what you personally typed or said or clicked. If an app is recording your keystrokes as you type an email to your therapist, that greatly increases your vulnerability, no matter how strict the developer’s security regime is. And it’s well established that developers’ internal security is consistently much laxer than it should be.
Hope they’re big fans of that Pewdiepie song banned in India.
The recordings the team reviews are tagged with the customer’s first name, account number and the serial number of their device.
Who is stupid enough to use their real name when joining these services?
actually, I found “compression” to be much more insightful, so that’s intentional, now.
/sorry, I don’t make the rules
