Google uses external collaborators as linguistic experts to improve the performance of Google Assistant, allowing them to listen to the audio files recorded by the devices equipped with the virtual assistant. All this should be done in full respect of the privacy and security of users.
However, the Belgian broadcaster VRT News has published a report in which it claims to have over 1000 user registrations, of which 153 should never have been registered as the "Ok Google" command would never have been clearly given.
Furthermore, the issuer could have traced back to the identity of the people to whom the records belonged by putting together some information. Google, however, claims that scripts sent to reviewers lack any personal information or data allowing for user identification.
In a post on the official blog, the company officially declared that " one of these linguistic reviewers violated data security regulations by filtering confidential Dutch audio data " and to have started an investigation in order to take the right steps.
With regard to the processes adopted to improve the assistant's understanding of certain user requests, the Mountain View giant emphasizes that a series of actions are applied to protect privacy during the review process, such as the elimination of sensitive data.
Language experts review only 0.2% of all audio fragments. Audio snippets are not associated with user accounts as part of the review process and reviewers are not required to transcribe background conversations or other noises and only to transcribe snippets addressed to Google
Google Assistant should send the audio only when it recognizes that the user is trying to interact via - for example - the "Ok Google" command. Apparently, however, some of the widespread recordings were recorded following what we could call "false commands", that is, after sentences similar to the command were uttered or when the button was accidentally pressed to activate it on smartphones.