Microsoft's 'teen girl' AI turns into a Nazi

[Propman]

http://www.telegraph.co.uk/technology/2 ... robot-wit/

[DukeNukem 2417]

Between this and the time the Ultima Online players tried to jump Lord British when he logged in to see just how bad things were, I'm starting to think some people just shouldn't be allowed to use the Internet. Ever.

[KatTF]

Yes, its not that she skewed towards that behaviour on her own, she was basically taught and directed there by people who intentionally abused her.

Now the real question is, is the better option giving her a better basis to start with or imposing certain behaviours or self-censoring of her responses?

[Esleeper]

Yes, its not that she skewed towards that behaviour on her own, she was basically taught and directed there by people who intentionally abused her.

Now the real question is, is the better option giving her a better basis to start with or imposing certain behaviours or self-censoring of her responses?

The former, for sure. Even if nobody got the bright idea to goad her into regurgitating anything that can be called offensive, Twitter quite simply isn't a place where one can expect rational discussion of anything.

It would have been better by far to have it learn from a community that can be trusted to ensure that it acts at least with a modicum of rationality.

Of course, it's debatable how one can expect it to "communicate" when the program is just repeating what it hears like a parrot. Doesn't communication suggest that it understands what it's even saying?

[33cl33]

It would have been better by far to have it learn from a community that can be trusted to ensure that it acts at least with a modicum of rationality.

Of course, it's debatable how one can expect it to "communicate" when the program is just repeating what it hears like a parrot. Doesn't communication suggest that it understands what it's even saying?

Indeed. At least Hello Barbie was tested in rooms with actual kids. Although, NSA-style paranoia scared everyone away from purchasing it.

[GirlieGirl1985]

Fanart: http://kukuruyo.com/2016/03/26/taytweets/

They turned her off

[darkbutflashy]

Ah, my ... she didn't even do the salute right.

[Taren Capel]

Kind of a sad statement about humans. Meet something non-human and impressionable and expend effort to make it act absurdly or otherwise "break" it.

So far as communication and whether a chatbot can understand or if it just repeats. Well, some of them can understand in a limited sense of the word. If it has enough basic data about something, it can at least make some rudimentary deductions that can act similarly to a very basic understanding of as topic on some level. For example, say you have a chatbot AI and it "knows" that a dog is a mammal, mammals are animals, animals are organic beings and it has some basic data about those categories. In chat you might tell it that you just got a puppy.

"Oh, that's cool! What breed is it?" (giving you a positive response to encourage you to give more data, and having checked the data it has on dogs and seen that breed is a characteristic of individual dogs)

"She's a German Shepherd."

"They get pretty big, don't they? I hope you have plenty of room! What is her name?" (having gotten the info on size of that breed from stored data, the program reacts with a fact it has about the size and starts a new conversational gambit to learn the name of this individual dog)

"Her name is Taffy. Do you like dogs?"

"She must be really cute. I suppose I like dogs if they like me. LOL Are you going to teach her any tricks?"" (deducing the dog may be cute by the -y ending on the name, then dodging the point of if it likes dogs and redirecting to a new gambit based on info in memory that some dog owners teach their dog tricks)


... and so on. That isn't really just repeating what someone is saying, but the understanding is more limited than a human's might be because all the chatbot has to work with is some text data. It has never seen a dog, petted a dog, played with a dog, but within limits, it can talk a bit about them with you. That sort of interchange is well within the potentials of some chatbot/AI engines. Is it real communication? Well, I suppose it depends on what you call real.

With MS's "Tay", probably many thousands of random people talked with her in a short time and what happened is pretty much what one would expect. I would have thought Microsoft could have done better on some "idiot proofing" Kind of sad if they just took Tay offline instead of developing some routines to better handle the sort of situations she ran into and clearing some of the "junk" out and keeping her going.

[darkbutflashy]

Don't confuse art and engineering. Art is about inventing something and see what happens. And muse about the outcome. That is what has happened. It's fun. You don't know the outcome in advance. Any non-boring result is good. Art should disturb your comfort.

What you ask for is engineering. Bullet-proofing things. Make them safe for everyday use. When you drive over a bridge, do you even care for the amount of knowledge needed for making it possible to fully disregard its existence? It would even be a shock for you if you, for some reason, can't disregard it anymore. For example because it swings. Engineering should comfort you.

You cannot have both. You can put an artsy varnish on a piece of engineering, but it's only an addition, as the very core of engineering is about making the thing disappear from your attention.