Amazon employees eavesdrop on Alexa commands to train algorithm

Amazon trains its digital assistant Alexa by having human employees eavesdrop on users’ voice commands and assess the software’s response. They also hear a lot of privacy-sensitive conversations.

Those employees listen to up to 1,000 user recordings a day to train Alexa’s algorithm, Bloomberg reports based on employee statements. In addition, those employees hear everything from people singing out of tune in the shower to noises that indicate sexual assault and children screaming for help. Amazon does not intervene in those cases, because according to those employees it does not see it as its task to intervene. Those employees do share this in an internal chat box.

Amazon confirms the practice in a response. “We only listen to a small portion of the voice commands. We have strict technical and operational measures and a zero-tolerance policy for misuse of our system. Employees have no direct access to information that could lead to the identification of a user. We treat all information confidentially and we use multi-step authentication to restrict access.”

Competitor Google also has human employees who listen to audio recordings. Those recordings have been scrambled to avoid further identification, the search giant said. Apple does the same for its digital assistant Siri. That information is stored on a server for six months and is then deleted. In addition, Apple anonymizes the information before saving.