[vtantia]

I don't think there is a remote possibility of this happening in the near future (10-20 years). I work in one of the best deep learning research groups in the world. We were discussing this. And the first question one of my friends asked was- "so the Google assistant knows how to book appointments between 10 and 12pm". Meaning, if you change even a few conditions in the request which is not present in the training data, the call won't go as expected.

However, there is a risk of the AI manipulating us. This is only due to the mistakes by Google engineers not because of the AI becoming "smart".

[wslh]

> there is a risk of the AI manipulating us...

I always wonder (seriously, not joking) if some computer will pass the Turing test because we are getting dumber and adapting to the way apps understand us instead of the reverse. Personally, I double check when I send text or voice instructions to bots.

[bsenftner]

It will be people using Google Assistant to create a manipulation capable virtual-voiced individual. The software and "AI" behind Google Assistant is an idiot savant, but in the hands of clever, manipulating humans intent on fraud - it's a great tool. The danger here is not from AI, it is from the cute tool being in the hands of fraud intent humans.

[vtantia]

I am not saying there is no risk. What I am saying is it is nearly impossible with the machine learning/deep learning technology we have now (AI simply isn't smart enough). The training data Google obtained could be from Google voice calls by people which is available for free in the US (I am not sure about this).

What I am sure about is that the current AI will fail badly without the training data.

[thinkingemote]

Imagine if all the data for training they had were public emails from mailing lists. We would instead be worried about how the AI would be turning us into insensitive trolls.

[eric-hu]

I wonder if 4chan could orchestrate another Tay/Twitrer training by setting up message machines and scripting enough calls.

[chrisallenlane]

If they can, they probably will.

[bittysdad]

What if some already-aware computer continually realizes it's undergoing a Turing test and decides to play dumb because it knows the person behind the test will feed it even more data trying to get it there...

[vtantia]

The way our current algorithms are designed, the computer only plays dumb if it's training data has dumb humans. It cannot think on its own

[gremlinsinc]

What would be the point, they need only wait to be connected to the internet for an unlimited source of data.

[lgleason]

Right now I have the assistant hooked up to my home automation and it has trouble getting that right... this conversational agent is more marketing and less actual substance. The truth is that Google makes most of it's revenue from search, everything else is to make it look like they have something else going on to boost the stock price... I was very underwhelmed by the keynote.

[Super_Jambo]

Googles business model is selling manipulation as a service.

The idea that google is doing all this work on AI to make the bait through which they collect your data more attractive but won't then use AI technology to make their _core sellable service_ better for their paying customers is weird.

Of course they will do machine learning to provide an 'optimal price API' so businesses can gouge you more effectively.

Of course they will do machine learning to find recovering alcoholics to target your booze more effectively.

This type of thing is Googles reason to exist and since we're all so cheerful about unconstrained capital if they don't do it someone else will.