Facebook’s Latest Audio Transcription Scandal Exposes a Key Dilemma of AI

Mark Zuckerberg

Facebook put its reputation at risk just a month after settling a record privacy breach case with the FTC. Artur Widak/NurPhoto via Getty Images

Last month, Facebook settled a yearlong case with the Federal Trade Commission (FTC) by agreeing to pay a $5 billion fine for sharing user information with the now defunct British marketing analytics firms Cambridge Analytica during the 2016 U.S. presidential election. Despite the fact that it was the largest penalty ever issued by the FTC, Facebook still hasn’t learned its lesson.

On Tuesday, Bloomberg reported that the social media giant recently paid hundreds of contractors to transcribe audio messages from some users’ voice chats on Messenger.

SEE ALSO: Apple Contractors Might Have Listened to You Having Sex Via Siri

Facebook said the users affected had agreed to have their voice chats transcribed in their Messenger settings. But Facebook’s data use policy has no explicit mention of audio recordings; it broadly states that “we collect the content, communications and other information you provide when you use our products.”

“It’s really intense right now, because a lot of people have heard Facebook say that they do not listen to their users’ conversations. This just runs afoul of what everyone was expecting to happen,” Bloomberg reporter Sarah Frier, who broke the news on Tuesday, said in a TV interview later that day.

So, why did Facebook insist on handing its users’ private conversations over to strangers knowing that its public trust was already on the brink of collapse?

The answers lies in a fundamental dilemma of Facebook’s development of artificial intelligence tools: robots are ideal replacements for humans to perform privacy-sensitive tasks (such as transcribing a private conversation), but in order to function well, these robots need to be trained by humans in the first place.

Facebook said the collection and transcription of voice chats were used to improve the voice-recognition systems used in the Messenger app’s automated, anonymous transcription service. Messenger offers this service to users who would like to keep a text version of voice memos sent by their friends or other audio-based conversations.

As in the case with all artificial intelligence-powered language detection tools in the market, the more data algorithms have to “learn” from, the better they can perform their tasks. In the Messenger case, having human workers transcribe real users’ voice chats would help robots improve their accuracy in future transcriptions.

Facebook told Bloomberg that it had stopped this human transcription program about a week ago and argued that this data collection practice is no different than what Google, Apple and Amazon have done for their voice assistance products, such as Alexa and Google Home.

But Frier pointed out that Apple, Google and Amazon only collect conversations between users and their devices, which are equivalent to conversations between users and those companies, whereas Facebook was collecting private conversations between users themselves.

Facebook’s Latest Audio Transcription Scandal Exposes a Key Dilemma of AI