Facebook says it has stopped feeding users’ audio chats to human reviewers for transcription, similar to what other Big Tech were recently caught doing. Facebook’s contractors were reportedly kept in the dark about their job.
The company “paused human review more than a week ago,” Facebook said in a statement on Tuesday, adding that the audio being reviewed was from Messenger users who opted in to having their voice chats transcribed to text by AI. The human transcribers were merely checking whether the AI transcriptions were correct, the company said.
Contractors often caught snippets of intimate or “vulgar” chats, though the messages were anonymized, and they were not told why they were transcribing the audio, where it came from or how it was obtained, contractors told Bloomberg. Some said the uncertainty and apparent duplicity – Facebook doesn’t tell users third parties may review their audio – troubled them.
While the company admits it collects “content, communications and other information you provide [when you] message or communicate with others,” there is no mention in its data use policy of humans processing the information. Instead, it’s Facebook’s “systems” that are supposed to “automatically process content and communications you and others provide to analyze context and what’s in them.” Even the list of third parties who might potentially receive users’ information doesn’t explicitly include human transcription teams, only “vendors and service providers who support our business” by “analyzing how our products are used.”
Further adding to contractors’ concerns about eavesdropping on users’ private messages is the fact that Facebook forbids at least one outsourcing firm from referring to them by name. Santa Monica, California-based TaskUs instead refers to one of its largest and most important clients only by the code-name “Prism” – which is, ironically, also the name of the NSA initiative exposed by Edward Snowden that installed backdoors in popular online platforms like Google, Skype, Twitter, and…Facebook.
TaskUs contractors also moderate content for policy violations, screen political ads, and work on “election preparation,” an ominous term Bloomberg opts not to define but which has previously coincided with mass deplatforming of political accounts in the months preceding elections. Facebook’s content-moderation contracting practices were exposed earlier this year after ex-contractors from outsourcing firm Cognizant complained that the work was giving them PTSD, leading them to drown their traumas in drug use and casual sex, and shaking their beliefs in the “official” versions of historical events like 9/11 and the Holocaust. Those employees, too, were forbidden from referring to Facebook by its given name.
Facebook added the AI transcription function to Messenger in 2015 and claims it is switched off by default. Its documentation, however, reveals that only one participant in a chat needs to opt in for the audio to be transcribed.
After similar practices were exposed by Google Assistant, Amazon Alexa, and Apple Siri, all three have promised to end the human eavesdropping – or at least give users the option of avoiding it. Apple has reportedly ended human “grading” of Siri recordings until a future software update gives users the option of participating, while Amazon has added an opt-out setting and Google has promised to suspend human review…in Europe, for at least three months.