Google's smart speakers are recording users when they least expect it, according to temp worker language experts hired by the company to listen to the snippets - which include some of users' most private moments.
The transcribers heard just everything: personal information, addresses, conversations between parents and children, bedroom talks, business calls, domestic violence and what not about Google Assistant users in Belgium and the Netherlands. Google justifies this practicing by arguing that "audio snippets are not associated with user accounts as part of the review process, and reviewers are directed not to transcribe background conversations or other noises".
"This is a critical part of the process of building speech technology, and is necessary to creating products like the Google Assistant", said Monsees. The recordings are disconnected from the users' information with the user name deleted and replaced with an anonymous serial number but the person can be relatively easily identified by listening to the recordings provided. This means that everything recorded in the background (including intimate or professional conversations), could be available to Google's employees.
Belgian broadcaster VRT NWS has reported that Google hires contractors to listen and transcribe audio recordings of users talking to Google Assistant, either via a smart speaker or their phones. Some of these recordings included sensitive information such as home addresses, pornography and medical questions - most of which you would expect people want to remain private. Anything that sounds remotely like the hotword triggers the beginning of recording. "This work is essential if we want to develop technologies that enables us to make products like the Google Assistant".
Temporary workers, supplied by contractors, are sent audio recordings to transcribe so that Google can improve its services for people who speak in different accents and languages.
When setting up a Google account, the option for voice and audio activity is set to off by default. But if it happens when you're having a conversation, then Google will have also heard a short except of your conversation. Out of the thousand or so recordings reviewed by VRT NWS, 153 were apparently recorded accidentally.
These contractors review almost 0.2 per cent of "audio snippets", as Monsees put it.
Google also said it's just about 0.2 percent of all audio recordings that are presented to the human interpreters. He added that there are protections in place to prevent these sort of "false accepts". We are actively investigating this and when we find a breach of our policy, we will take action quickly, up to and including the termination of our agreement with the partner.