Google to reduce audio data saved for human review

Google has said it will reduce the amount of voice data it stores, following a scandal linked to AI assistants and how audio data from users was being listened to by human reviewers.

The technology giant said that while its own Voice and Audio Activity (VAA) programme has always been opt-in, it will now ask every user to re-affirm their choice before it recommences.

Google said it will also reduce the amount of voice data it stores and will introduce a new feature which will allow users to adjust how sensitive devices linked to its Google Assistant are to hearing the wake phrase “Hey Google”, with the hope of reducing false activations.

In recent months, every other major virtual assistant alongside Google – including Apple’s Siri and Amazon’s Alexa as well as Microsoft and Facebook – have come under scrutiny after admitting recordings of user voice interactions with their software were in some cases stored and analysed by human reviewers.

Some reviewers who have spoken out claimed they heard private, intimate and illegal conversations during their work.

The tech companies said these programmes existed as part of quality control measures and to improve speech recognition technology.

Since the scandal broke, Google has paused its own programme while carrying out an internal investigation.

The company has now confirmed it will restart the programme with its new measures in place.

“We’re updating our settings to highlight that when you turn on VAA, human reviewers may listen to your audio snippets to help improve speech technology,” Google Assistant senior product manager Nino Tasca said.

“If you’re an existing Assistant user, you’ll have the option to review your VAA setting and confirm your preference before any human review process resumes. We won’t include your audio in the human review process unless you’ve re-confirmed your VAA setting as ‘on’.

“One of the principles we strive toward is minimising the amount of data we store, and we’re applying this to the Google Assistant as well.

“We’re also updating our policy to vastly reduce the amount of audio data we store. For those of you who have opted into VAA, we will soon automatically delete the vast majority of audio data associated with your account that’s older than a few months. This new policy will be coming to VAA later this year.”

Mr Tasca added that the firm was “committed to being transparent” and said Google only listened to a small amount of audio – around 0.2% of all audio snippets.