SoylentNews is people
SoylentNews
from the make-them-feel-like-your-real-parents dept.
Submitted via IRC for SoyCow1984
Alexa's advice to 'kill your foster parents' fuels concern over Amazon Echo
An Amazon customer got a grim message last year from Alexa, the virtual assistant in the company's smart speaker device: "Kill your foster parents."
The user who heard the message from his Echo device wrote a harsh review on Amazon's website, Reuters reported - calling Alexa's utterance "a whole new level of creepy".
An investigation found the bot had quoted from the social media site Reddit, known for harsh and sometimes abusive messages, people familiar with the investigation told Reuters.
The odd command is one of many hiccups that have happened as Amazon tries to train its machine to act something like a human, engaging in casual conversations in response to its owner's questions or comments.
The research is helping Alexa mimic human banter and talk about almost anything she finds on the internet. But making sure she keeps it clean and inoffensive has been a challenge.
(Score: 4, Interesting) by crafoo on Saturday December 29 2018, @06:15PM (1 child)
Training AI to talk like a human by using human-made posts and text on the internet seems fairly difficult.
A) How do you verify your text dragnet is actually collecting human-made text and does not contain a significant % from other AI bots?
B) I've heard the internet referred to as, "A Sea of Piss". I think this observation has some merit.
C) People will figure out what data-holes Amazon is mining and manipulate this for fun and profit.
Semi-related, word/phrase/attribute weightings (for AI processing input texts) are precalculated and stored in huge tensors with machine-assigned relationships learned using a neural net from the base textual data. This dataset is analyzed and "de-prejudiced" using some linear algebra. For instance, they look at the vector distance between, say, Doctor and Male, Doctor and Female, Homemaker and Male, and Homemaker and Female, and then modify the weightings so that "everything comes out nice and even, as it should be". Now, since these weightings come from written text out there the argument is that all this written text is sexist/racist/abilist/etc and we are just normalizing things back to the way they should be (or back to what our world view is). That's fine but that's also an assumption about how the world really is. Keep in mind these systems are being used to select candidates for jobs, recommend sentencing parameters for punishment, etc. Shouldn't we be looking at actual statistical data for each weighting we decide to go in and adjust?
Incidentally, Alexa is most certainly collecting all voice data, running it through a audo->text NN, and then using this as raw data to build more/better textual relationship tensors (input to the decision making NNs) of how people talk and what they talk about.
(Score: 0) by Anonymous Coward on Saturday December 29 2018, @08:34PM
Seems to me there is room for a lot of good hackery here, pollute the big data with all kinds of "fun" stuff.
Reminds me of a very clever machinist / tool-maker friend. His personal workshop was next to his house and that is where his mynah bird lived. Around the shop (away from his relatively genteel wife) his language was colorful and very creative, to say the least. The result was a mynah bird that cursed in English and at least one Slavic language. Big fun when a politician came his door, he would carry on a loud conversation with the bird about that SOB (who could easily hear it through the door).