Google admitted on Thursday that “language experts” hired by the firm listen to approximately 0.2% of the conversations that users have with their virtual assistant, implying that some of those interactions are not completely private.
The usual assumption, and what is often reiterated by companies that manage virtual assistants such as Amazon, Samsung, Apple, or Google, is that conversations between a user and his assistant are completely private and that the interaction occurs exclusively through artificial intelligence, that is, the only ones that ‘listen’ to the user are robots.
However, Google’s admission this Thursday that 0.2% of these conversations are listened to by human beings to, they say, improve the quality of service sheds light on a practice that companies generally avoid advertising, although it is known within the industry that to a lesser or greater extent, is commonplace.
The revelation came from the hand of the search product manager of the Californian company David Monsees, who published an entry on the official Google blog in response to information appeared on Wednesday on the Belgian television VRT NWS, which could access about a thousand recordings of anonymous individuals.
The recordings were made available to Dutch-language Belgian television by one of the “experts” that Google had hired in that country to listen to segments of the conversations and “thus understand the particularities and accents of each specific language”. The firm, which has already announced that it will “take action” for the leak as a “violation” of its data security policies, admitted having “experts around the world” whose function is to listen to and transcribe “a small part of the dialogues to help us better understand these languages. “Specifically, the Mountain View (California, USA) firm encrypted the percentage of interactions analyzed by humans at 0.2% and ensured that these fragments are not associated with user accounts and that experts are told not to transcribe background sounds or conversations that are not directed at Google.
However, Belgian television was able to identify “postal addresses and other sensitive information” in the recordings, which allowed them to contact the people whose voice had been recorded and confirm that it was indeed them. “A couple from Waasmunster (Belgium) immediately recognized the voice of their son and grandson,” they gave as an example from VRT NWS.
Google indicated that the virtual assistant only sends them the audio recordings once he has detected that the user is interacting with him after having said, for example, “Hey, Google” and that he has several tools to avoid the “false activations”, that is to say, that the software interprets a sound erroneously as the key word to activate. Despite this, VRT NWS reported that of the 1,000 or so voice fragments it had access to (all in Dutch), 153 were conversations in which no one gave the activation order to the virtual assistant, but the virtual assistant misinterpreted a sound. Among others, the Belgian media claimed to have listened to conversations in bed, between parents and their children, professional calls, discussions and scenes of both sex and violence.