Facebook has been paying hundreds of outside contractors to transcribe clips of audio from users of its services, according to people with knowledge of the work.
The work has rattled the contract employees, who are not told where the audio was recorded or how it was obtained – only to transcribe it, said the people, who requested anonymity for fear of losing their jobs.
They’re hearing Facebook users’ conversations, sometimes with vulgar content, but do not know why Facebook needs them transcribed, the people said.
Facebook confirmed that it had been transcribing users’ audio and said it will no longer do so, following scrutiny into other companies. “Much like Apple and Google, we paused human review of audio more than a week ago,” the company said on Tuesday.
The company said the users who were affected chose the option in Facebook’s Messenger app to have their voice chats transcribed. The contractors were checking whether Facebook’s artificial intelligence correctly interpreted the messages, which were anonymised.
Big tech companies including Amazon. com Inc. and Apple Inc. have come under fire for collecting audio snippets from consumer computing devices and subjecting those clips to human review, a practice that critics say invades privacy. Bloomberg first reported in April that Amazon had a team of thousands of workers around the world listening to Alexa audio requests with the goal of improving the software, and that similar human review was used for Apple’s Siri and Alphabet Inc.’s Google Assistant. Apple and Google have since said they no longer engage in the practice and Amazon said it will let users opt out of human review.
The social networking giant, which just completed a $5 billion (€4.4 billion) settlement with the US federal trade commission after an inquiry of its privacy practices, has long denied that it collects audio from users to inform ads or help determine what people see in their news feeds. Chief executive Mark Zuckerberg denied the idea directly in congressional testimony.
“You’re talking about this conspiracy theory that gets passed around that we listen to what’s going on on your microphone and use that for ads,” Mr Zuckerberg told US senator Gary Peters in April 2018. “We don’t do that.”
In follow-up answers for Congress, the company said it “only accesses users’ microphone if the user has given our app permission and if they are actively using a specific feature that requires audio (like voice messaging features.)” The Menlo Park, California-based company doesn’t address what happens to the audio afterward.
Facebook hasn’t disclosed to users that third parties may review their audio. That’s led some contractors to feel their work is unethical, according to the people with knowledge of the matter.
At least one firm reviewing user conversations is TaskUs, a California-based outsourcing firm with outposts around the world, the people said. Facebook is one of TaskUs’s largest and most important clients, but employees aren’t allowed to mention publicly who their work is for. They call the client by the code name “Prism”.
Facebook also uses TaskUs to review content that’s in possible violation of policies. There are also TaskUs teams working on election preparation and screening political ads, though some of those employees were recently moved to the new transcription team.
The Facebook data-use policy, revised last year to make it more understandable for the public, includes no mention of audio. It does, however, say Facebook will collect “content, communications and other information you provide” when users “message or communicate with others”.
Facebook says its “systems automatically process content and communications you and others provide to analyse context and what’s in them”. It includes no mention of other human beings screening the content. In a list of “types of third parties we share information with,” Facebook doesn’t mention a transcription team, but vaguely refers to “vendors and service providers who support our business” by “analysing how our products are used”.
The role of humans in analysing recordings underscores the limits of artificial intelligence in its ability to recognise words and speech patterns. Machines are getting better at the task but sometimes still struggle with the unfamiliar. That some of the contractors have found the recorded content disturbing is further reminder of the human toll of moderating content on Facebook, the world’s biggest social network.
Facebook first started allowing Messenger users to have their audio transcribed in 2015.