Some of the conversations we have with Google Assistant or Google Home devices can be heard by Google employees or subcontracted from the company . 0.2% of them according to Google. As part of the improvement and training of their voice recognition service, audio fragments are sent to third parties to be analyzed and learned from them.
In the last few hours, the process by which Google sends audio fragments (anonymized) to subcontracted employees for them to analyze them has come to light . The Belgian radio VRT NWS has shown how third parties can through this process listen to audios with sensitive information. They are not the only ones, Amazon also recently admitted that it retains data from some users even if the user deletes the voice recording.
Why a Google out sourcer can listen to your conversations
To clarify the situation Google has published an article explaining why these conversations are sent and what precautions are taken to protect the identity of users. As they indicate, it is a key process to improve voice recognition technology . They have employees and outsourced linguists all over the world who analyze audios to improve voice technology. In summary, it is one more aspect to improve Google Assistant and Google Home.
To protect the privacy of users, the company ensures that all audio clips sent to these employees are completely anonymized clips, so they are not associated with any user account with which they can be identified. But of course, if the audio that is sent contains some name, email address, phone number or postal address … it is something that Google can not do anonymous so that it does not associate with anyone. Some of the conversations obtained by VRT NWS have this type of data.
Google Home and Google Home mini.
The incident has happened with Dutch audios that have leaked to VRT NWS, this is something that violates the privacy and security policy that Google has with the companies with which it collaborates. They comment in the official statements that they will investigate the matter to see what has happened and will take measures so that it does not happen again.
A final detail has to do with when and when these audios are not recorded. According to the VRT NWS report, of the more than 1,000 audios that received a total of 153 were recordings that users were not aware of. This is because both Google Home or Google Assistant can accidentally record conversations. It happens when the system thinks it has been invoked but it really is not, which has been a false positive.
Google’s solution to involuntary eavesdropping is to provide clear indicators of when you are listening . Both with audio tones and with LED illuminations or animations in the devices that indicate that the user is actively listening. Do not avoid this that accidental recordings are made by pressing a button without wanting to or hear an “OK Google” in a video for example.
From Xataka we have contacted Google Spain to see what their position is. They have indicated the following:
“We are now partner with expert linguists around the world to improve voice technology by transcribing a small set of words.This work is key to developing the technology that makes possible products such as the Google Assistant. Expert linguists review only about 0.2% of all the small audio fragments and those fragments are not associated with the user accounts in the review process.
“We have learned that one of the reviewers has violated our data security policies by filtering out Dutch audio data, our security and privacy teams have been activated, are investigating and taking action, and we are conducting a complete review of our security devices for prevent behaviors like these from happening again “