Four of the biggest tech companies are using human contractors to listen in on samples of voice requests. This means requests between you and your voice assistant may have another set of ears listening in.
One of the most popular voice assistants, Google Assistant, is included with most Android devices, devices running the Chrome browser, and iOS devices running the Google app. Google also has a dedicated smart speaker for voice assistance, as well as a range of smart TVs with similar functionality.
Its strength comes from off-device processing, meaning that voice requests are sent to Google’s cloud to be transcribed, understood, and sent back for assisting a user. This means that almost any smartphone with an Internet connection can run the assistant as smoothly as the latest Android flagships.
Google has had a somewhat-hushed scandal with human voice contractors in Belgium, where Time Verheyden, a journalist with Belgian broadcaster VRT, was given access to over 1000 voice files containing requests and some random snippets of audio. 150 of these requests were snippets from smartphones that thought they heard the “OK Google” wake word. These included a child and a grandfather speaking and a woman in distress.
The voice assistant included with most Apple devices, Siri, is also not as innocent as Apple claims it is. Most new iPhones, iPads, and Apple Watches are consistently listening, and some “false positives” – when the device mistakes sounds for the wake word – are sent back to contractors. Apple says this done to analyse why the device thought it was requested and what can be done to prevent it in the next version update. Contractors told The Guardian that Siri can be summoned by the sound of a zip being zipped up, and that they have heard potential drug deals, as well as the sound of sexual acts.
Apple has since suspended this project with its contractors, presumably to continue at a later date, and has planned an opt-out feature for users to have their recordings sent to Apple for improving the assistant.
Some contractors came forward to Bloomberg to say that, while most requests were mundane, some were “possibly criminal”, including potential sexual assault. Amazon told CNN Business that an “extremely small number of interactions from a random set of customers” were sent to these listening centres. While this quote may be reassuring at face value, 6.3-million Amazon Echo devices were sold worldwide in Q3 of 2018. This means that an “extremely small” amount of questions taken from an extremely large request base is less non-trivial than it sounds.
Microsoft’s innovations are vast, but come at the cost of privacy. Its voice assistant included with Windows 10 and Xbox, Cortana, is most likely on and listening if users chose the “Express Setup” option when setting up their devices for the first time. This means that Microsoft may have far more false positives than the other voice assistants, because users either don’t know it’s there, or have forgotten about it since setup. Like Google, Amazon and Apple, contractors were given access to random voice requests, some of which were false positives, to improve its voice assistant.
Skype conversations are also sent back to contractors to assess the effectiveness of translations. Some contractors heard very intimate conversations, and some clips even included bank account information, home addresses, and bedroom conversations.
The events mentioned above all happened in 2019, which shows that artificial intelligence, especially natural language processing, is not ready to do all the work for big tech companies. This is why a little more human training is required to teach these systems about what to listen out for and what can be ignored.
However, it comes at a cost to privacy. Often, it is a hidden cost.
Read more about how voice has become the new oil.