At Google, too small ears, listen to the audio recordings of the Wizard

After the broadcast of a documentary about the treatment of the data obtained by his assistants, Google has wanted to put in perspective the place given to persons in the listening and the analysis of these contents. A communication that leaves questions unanswered.

In a ticket, Google said that the experts that actually work to the four corners of the world, to listen to and transcribe audio clips, work only on 0,2% of this mass of data. This is already an indication, but the Wizard of Google handles probably millions of queries per day : the small percentages can make big rivers.

Presentation to the press of the Google Nest Hub

Google adds that these precautions are provided to separate these records from the user accounts, so as to anonymize. These employees are instructed not to write down the instructions explicitly given to the wizard, not the words they could hear in the background.

Google also claims to have opened an investigation in light of the breach of the confidentiality agreement, which has been guilty of this employee, interviewed hidden for a program broadcast in Belgium.

This communication dodge, however, the questions raised by this documentary signed VRT NWS. We see that Google uses providers around the world to listen to audio clips that are sent in large volumes every month. The witness of issuance of said handle about 1,000 per week.

These operators have a mission to transcribe, in order to improve the analysis capabilities of the wizard. Already, Google does nowhere in its terms of use which sub-processes these information, or simply pass them on to people for listening.

Another problem is that these records may be separated from the identity of those who caused them, it frequently happens that the wizard picked up by error of the remarks which were not intended.

All those who use the assistants of Apple, Google or Amazon have been confronted with these awakenings impromptus of the software, while one had not called. The wizard includes evil and believes to have heard a “Say to Siri” or “Alexa”, then no. To add to these problems already current, the interviewee explains that the wizard of Google has progress to make in flemish and that the errors of interpretation to correctly detect an “Ok Google” can also be explained by this bias.

The user interface to listen to his own recordings and delete them

The journalist then tells that a little over 1000 records reviewed, it has counted 153 who had been captured by mistake by the assistant, and delivered for transcription. From here stems another problem, these records impromptus may contain information that is more personal because the person was not in his enclosure.

The journalist went to meet a couple, and a man whose addresses or names were present in these extracts audio. Each was able to confirm that it was his voice or that of a person in the house. In a more general way, even if the wizard has correctly responded to a request, it can frequently recover information such as requests for appointments with a doctor, names, or addresses, that fall may be in the ears of these employees.

The operator said that he listened to a day to an audio file in which a woman appeared to be in distress, may be abused. What do you do when this happens, how to react ? The instructions Google do not address this kind of situation, ” he says wry. And realize at that moment that these are no longer the only voices that are listened to at the chain, but people in their daily lives.

The software of listening and speech-to-text used by the providers of Google

The criticism made to Google, focuses on the lack of transparency on behind-the-scenes analysis of these data. On the involvement of people in flesh and bone, not still employed by Google, who listen to these exchanges. So that may be a lot of these owners speaker smart believe that their voices are only processed by computers and software.

This topic echoes similar questions about Amazon, which is based also on the ear of human beings to improve Alexa, with issues identical.