Google Home, the smart speaker that can you can control with only your voice, and Google Assistant, the search giant’s answer to Siri, record some of your interactions with them for the company to review. We technically already knew about that—it says so in the terms & conditions for those services—and we all read those, right? What’s shocking, then, is a new leak from a Google employee of around a thousand recorded audio files that contain more than a hundred and fifty examples of the company eavesdropping on conversations where the “OK Google” command was not given.
The leak comes via the Belgian news outlet VRT. According to their report, Google has been employing subcontractors to work on its speech technology. Whenever your Google device has trouble understanding a speech command, it uploads a recording of that command for humans to review, so that they can teach the speech-comprehension algorithm how to not make the same mistake again. Although Google concedes that Google Home and Google Assistant record your interactions with them in their terms & conditions, it doesn’t disclose that there are humans who listen to some of this data.
Google employees explained to VRT that part of their job involves transcribing these audio files. If they came across words they didn’t know how to spell—often names of locations, people and workplaces—they would search them up, and in the process accidentally discover the identities of those who were recorded.
The privacy concerns this practice raises are obvious. Although names are anonymized in Google’s system, incriminating details contained within the audio files themselves are not censored. Captured in the leaks include medical questions, addresses, private phone calls, bedroom talk, and porn searches. While Google doesn’t exploit this data for anything other than teaching its voice recognition algorithm, it only takes another security breach like these leaks ending up in the wrong hands for things to go south.
In an official statement in response to the VRT report, Google condemned the leaks as a violation: “Our Security and Privacy Response teams have been activated on this issue, are investigating, and we will take action. We are conducting a full review of our safeguards in this space to prevent misconduct like this from happening again.” They also claimed that “Language experts only review around 0.2 percent of all audio snippets;” for context, there are one billion devices out there that can use Google Assistant. Google also dismissed “false accepts,” their term for audio recorded without an “OK Google” command, as a rare occurrence.
Google also told Ars Technica that all voice recording for Google Assistant is optional, and that storing your recordings is by default turned off. Users can also delete their previous recordings or schedule them for regular deletions in the settings of their Google Account.
Bloomberg reported in April that Amazon and Apple employ thousands of voice reviewers as well to improve their respective digital assistants, Alexa and Siri.