Likewise Facebook: The group has hired hundreds of employees from external service providers who automatically check transcribed voice records on Facebook Messenger. This was announced by Bloomberg.
Facebook reported that the rating "more than a week ago" was suspended. These were completely anonymous records. They only used the data to train systems, not to personalize ads. Moreover, Facebook never includes microphones in the background. It was always necessary for users to consciously activate them. It is unclear in which countries the users are affected and in which service providers the employees worked.
A few weeks ago, Facebook's practice would provoke outrage worldwide, now many are shrugging. Facebook CEO Mark Zuckerberg can thank Amazon, Apple, Google and Microsoft.
Alexa instead of Alexandra
The language assistants respond to the activation words and confuse, for example, «Alexandra» with «Alexa». Or hold the sound of a zip for a command to Siri. Then the microphones pick up, the data goes to servers in the US and possibly other companies that work as third party service providers for corporations.
Technology companies talk about individual cases, but do not give specific numbers. Media reports suggest that false alarms are more common. The Flemish VRT operator has leaked about 1,000 Google records. Of these, about 150 have been recorded without users having heard of it. Bloomberg calls 10 percent of unintentional Alexa activations.
This newspaper had contact with five people who copied records from their own account to Amazon, Apple and Google. They also report having heard private records repeatedly. These include sensitive professional phone calls and intimate moments – apparently also in many sleeping smart speakers. Language assistants sometimes take on disputes, threats and physical attacks. This is a burden for employees because they are witnesses to alleged crimes, but they can do nothing.
Although sometimes records allow users to make inferences about their identities, for example, if names or addresses can be heard. But they are not tightly linked to profiles, but pseudonymized, that is, they are assigned randomly generated IDs. Whether the corporations themselves may subsequently outsource them is unclear. In any case, employees of external service providers do not have this capability.
Discoveries are also outraged by the term "artificial intelligence". No system works completely autonomously, all language assistants make regular mistakes. That is why people need to train algorithms. You listen to a small portion of the recordings and check that the machine has heard correctly.