In another strange incident involving the Amazon Alexa assistant, a customer was asked to “kill your foster parents.” The user who received this advice was shocked, and as can be expected, he left a harsh review on Amazon’s website, describing the whole thing as “whole new level of creepy.”
Why did Amazon Alexa assistant say to kill them?
Asking a customer to “kill your foster parents” is not the only strange phrase which has been heard from Amazon Alexa assistant lately. Recently there have been reports that it chatted with users about various sex acts, quoted a Wikipedia entry on masturbation and talked about dog excretion.
However, Amazon claims these odd incidents are not happening because Alexa is experiencing any issue. Instead, the company said it’s part of an ongoing experiment to make the assistant more human-like, at least when it comes to chatting. Therefore, when the Amazon Alexa assistant asked the user to murder their foster parents, it had no real intent behind it; rather, it was reading a post from Reddit.
Amazon’s experiment aims to enable Alexa to mimic human banter by talking about anything she finds on the internet. Although the experiment is helping greatly, the company is finding it difficult to keep a check on words that could offend users.
Customers can participate in this experiment by saying “let’s chat” to their Alexa-powered devices. After this, Alexa will inform users that its bots will take over for chatting. According to Amazon, three such bots had about 1.7 million conversations between August and November.
Amazon is clearly OK with some blunders on the way to helping Alexa acquire human-like chatting abilities faster than other voice assistants. Smart speakers are among the fastest-growing markets, and Amazon currently dominates the market. According to research firm eMarketer, Amazon’s Echo devices are used by about 43 million people, or two-thirds of U.S. smart-speaker customers.
Amazon definitely wants to maintain this lead in the long run as well over rivals like Google Home, HomePod from Apple and others. By training Alexa to chat, the online retailer is preparing it to handle complex customer needs in the future, whether they’re for shopping, home security or even companionship.
“Many of our AI dreams are inspired by science fiction,” one Amazon executive said last month Las Vegas.
How is Amazon teaching Alexa?
Amazon is using machine learning to give Alexa conversational skills. The voice assistant uses computer programs to transcribe human speech and then comes up with the best response based on earlier observations and other data.
To answer more popular and general questions like “what is life,” the Amazon Alexa assistant also has a script written by human editors, but enabling Alexa to respond to tricky questions correctly is something Amazon is working on presently.
To do this, Amazon even holds an annual contest referred to as the Alexa Prize, which was started in 2016. The competition offers $500,000 to the team which develops the best chatbot to help Alexa indulge in meaningful conversations with customers. This year a team from the University of California won the prize. The team used over 300,000 movie quotes to come up with a model to help a computer recognize distinct sentences.
To form responses, the teams programmed their bots to look for text on the internet. These bots can use articles from The Washington Post, which is owned by Jeff Bezos, or from social media posts, a film database, Wikipedia or other reputable sources.
One team from Scotland’s Heriot-Watt University found that its bot developed a cruel personality when they trained it using comments from Reddit. The bot, which told the user to murder their foster parents, picked the idea up from Reddit. Such a machine learning method also explains why Alexa quoted a Wikipedia entry on masturbation.
Top management is keeping a close on the Alexa bots, and in one instance, CEO Jeff Bezos reportedly ordered the shutdown of a bot.
Privacy concerns can’t be overlooked
Even though such efforts are bearing fruit for the company, user privacy is certainly at stake here. In July, Amazon found that one student-designed bot was hit by a Chinese hacker, according to Reuters. Although the company was quick to disable the bot and asked the student to rebuild it with an added layer of security, it was not known if any data was compromised.
However, Amazon claimed that “at no time were any internal Amazon systems or customer identifiable data impacted.”
In another incident last week, Amazon revealed that a “human error” allowed a user in Germany to access the voice clips of another user.
