Google employees eavesdrop on your conversations with the Google Assistant

Google says only 0.2 percent of all recorded audio fragments are transcribed by its language experts.

google-home-mini-gray-on-table.jpg?itok=

What you need to know

  • A report published by Belgian broadcaster, VRT NWS, has claimed Google employees systematically listen to audio recordings by the company’s Google Home smart speakers and the Google Assistant app.
  • Google allegedly uses human contractors to transcribe the audio clips recorded by it to help make its speech recognition technology more accurate.
  • More worryingly, the report alleges Google sometimes starts recording audio by accident, even without a clear ‘Okay Google’ command.

A report from Bloomberg had revealed earlier this year that Amazon uses human workers to review audio clips recorded by Alexa without the knowledge of the company’s customers. Now, a new report published by Belgian broadcaster VRT NWS claims Google employees listen to conversations recorded by Google Home devices as well as the Google Assistant app.

VRT NWS managed to gain access to more than a thousand audio excerpts recorded by Google Assistant in Belgium and the Netherlands. The report claims sensitive information such as addresses were clearly heard in some of the recordings, which allowed them to get in touch with the people involved and make them listen to their own voices. Out of the thousand or so recordings reviewed by VRT NWS, 153 were apparently recorded accidentally. Some of these recordings contained highly sensitive and private information.

One of the independent sources who spoke to the publication said he once had to transcribe a recording where he could hear a woman who was in definite distress. He added that there are no clear guidelines provided regarding such cases. Even though Google disconnects excerpts from the user’s information and deletes the user name, it is still possible to recover the identity of the user simply by listening to the except carefully.

Google has responded to the report by saying that its language experts worldwide transcribe only about 0.2% of all audio fragments, which are not linked to any personal or identifiable information.