When users ask Alexa about their mysterious rash, or to turn off the lights, they might not expect someone else to be listening.
A.I. needs human input—and human reviewers—to become smarter. This week, a Bloomberg report pulled back the curtain on the team of people around the world who are tasked with listening to the Alexa queries of unsuspecting users. And the A.I. training team’s members number in the thousands.
The employees listen to recordings of people asking for Alexa to turn off the lights or play Taylor Swift. They transcribe the queries and feed them back to the Alexa software, making it smarter and more adept at grasping the way humans speak.
“It is normal to train this way, and a less sexy side of A.I.,” said Nico Acosta, director of product and engineering at Twilio Autopilot, a platform that allows developers to build bots and Alexa apps. “All speech engines need to be trained on real world audio, which implies the need to have a human transcribe it to continuously train the engine.”
There’s a clear privacy trade-off in having these smart speakers in your home. In a statement to Fortune, an Amazon spokesperson said the company uses “an extremely small number of interactions from a random set of customers,” who are not identifiable to the employees who are listening.
“For example, this information helps us train our speech recognition and natural language understanding systems, so Alexa can better understand your requests, and ensure the service works well for everyone,” the spokesperson said. “We have strict technical and operational safeguards, and have a zero tolerance policy for the abuse of our system.”
Raw human training data is “critical” when it comes to keeping the quality of the service, said Richard Ford, chief scientist at cybersecurity firm Forcepoint.
“If you want to do voice recognition for Alexa, the best data to train it on is on actual ‘as used’ scenarios, where there’s background noise, dogs barking, people changing their minds… all the ‘mess’ that you find in the real world,” said Ford.
However, there are other ways Amazon could train Alexa without eavesdropping on tens of millions of queries, he said.
“You could pay people to opt in to share their data willingly, or take part in trials, but at the end of the day, getting truly realistic data in a tractable way probably involves capturing real world data,” he said. “There are mitigations you can potentially put in place to minimize the privacy risks, but they are not infallible. Privacy is a confluence of good governance, good design, and good implementation.”
While the story has added to the concerns of people who are already worried about the privacy issues involved with allowing a tech giant’s smart speaker to live in their home, Amazon said its speaker only records queries and sends them to the cloud after it hears its wake word, such as “Alexa” or “Amazon.” A clear sign that the Echo speaker is recording: the device’s blue ring lights up.
There are ways to get rid of old recordings. Users can manually delete everything they’ve ever asked Amazon Alexa by visiting the Amazon Connect and Devices website. Once there, select “devices,” the Amazon Echo, and then “manage voice recordings.”
To opt out of being an unwitting A.I. trainer altogether, in the Amazon Alexa app, tap the menu button in the upper left corner of the screen. Then select “Alexa Account” and “Alexa Privacy.” Choose “Manage how your data improves Alexa.” Next, click off the buttons next to “Help Develop New Features” and “Use Messages to Improve Transcriptions.” The settings will keep Amazon from using raw recordings to train its software.
Of course, if too many people opt for privacy, the A.I. will take a lot longer to improve its understanding of natural language. “Getting such a corpus is really hard without using real data, which is why there’s often a genuine need to collect data from actual usage,” said Ford. “If you’re going to deliver your product on time and with a high efficacy, it’s a hard problem.”