Amazon Employees Might Be Listening to Your Alexa Recordings
Amazon wants us to think of Alexa as a computer that does our bidding with the help of magical AI technology, but there’s still a significant human component, according to an investigation by Bloomberg. The company has allegedly used an army of human beings to listen to audio clips from Alexa devices and grade the interactions. Amazon believes this is an essential way to improve the service, but it raises numerous privacy concerns.
Some improvements to Alexa are on the technical side via machine learning techniques, but the human-driven data analysis is increasingly seen as the best way to make Alexa smarter. According to Amazon workers interviewed by Bloomberg, the Alexa review process starts when the company pulls a small random sample of audio recordings. It sends those files to employees and contractors around the world to listen to the interaction.
The workers pull nine-hour shifts in Boston, Romania, India, and other locales. Reviewers might process 1,000 audio clips each day. Some workers transcribe the user commands and compare that with what Alexa heard. Others add annotations to the recording that informs future user interactions. For example, one worker described scanning mountains of Alexa recordings to utterances of “Taylor Swift” and adding annotations that marked them to indicate the user was asking about music.
Recordings forwarded to Amazon employees and contractors include all the audio sucked in by the company’s smart speakers. Workers describe numerous instances of Alexa capturing personal information and background noises when it was incorrectly triggered. Two employees even described hearing what they believed was a sexual assault, but Amazon told them to ignore it as it was not their job to interfere. Alexa users can opt out of some audio sharing in their account settings, but the company says their snippets might still end up in the review queue.
Amazon claims it treats user data with the utmost care. Recordings are encrypted in transit, and employees adhere to a strict code of conduct. While it doesn’t include your full name or email with the data, the recordings do have an account ID and first name attached. Apple uses a similar system of human workers to improve Siri, but it doesn’t include any identifying data with the recordings. Google has humans working on Assistant, but audio recordings are intentionally distorted, and there’s no account data associated with them.
Amazon is still largely trusted by consumers, but that could change with too many privacy snafus. Several months ago, The Intercept discovered that Amazon-owned Ring gave employees access to all user camera recordings and live feeds. The Alexa situation isn’t as egregious, but there’s definitely a creepy factor.