| Hi everyone, it's Sarah Frier. Our most recent Bloomberg Businessweek cover story is about the artificially intelligent assistant devices that collect and parse clips of our voices—with some extra help from humans transcribing the audio. Big tech companies thought the practice would be a completely normal and expected evolution of all their other work in AI. But the contractors listening to the audio were uncomfortable with the intimacy of what they were hearing. They realized that consumers didn't know they were being recorded. The responses to this story, and others Bloomberg has written this year, fall into similar categories. People who understand AI programming say that of course humans are behind the scenes—doesn't everybody know that? Consumers say no, everybody does not know this, and it's a problem. So why is there such a disconnect? If you ask Facebook, it's because people think machines are smarter than they actually are. "Ten years ago it would be shocking to think a human wasn't reviewing it," Andrew Bosworth, the head of hardware at Facebook, tells me. "Now the expectation is that machines are more likely to be reviewing content than people. That's a real shift in consumer expectation." And he says Facebook—just like Amazon, Google, and Apple—was caught off guard by the outrage. If consumers are over-estimating their machines, the companies' marketing is largely to blame. The phrases "artificial intelligence" and "machine learning" certainly imply human-free solutions, which the media hyped. When you buying these devices, you're told Alexa or Siri is responding, not that it's being fact-checked by thousands of real people who can hear your actual voice. Facebook told consumers that the more they used the machine-transcription service in its Facebook Messenger app, the smarter the machine would get. It didn't say that it was getting smarter because people in a contractor hub in Texas would have more voice samples to check for accuracy. Bosworth, like others representing companies in our story, said language understanding is one of the hardest problems in technology. Without all these real examples of people talking, only certain kinds of people—the ones with the most common accents or speaking cadence—would be served well in a future voice-activated world. "You want these to be inclusive tools that work for everybody," Bosworth explained. Computers are looking at sounds, not words. "Subtle differences in tone, in frequency, in modulation, to humans are very similar but to computers are completely different." That sounds like a laudable mission people might have volunteered for, if companies asked nicely and explained clearly what happens behind the scenes. So why didn't companies start to tell consumers what was happening until after transcribers told their stories to Bloomberg News? Because the companies knew what they were doing was controversial, says Justin Brookman, director of privacy and technology policy at Consumer Reports. "They kind of got caught with their hands in the cookie jar."—Sarah Frier | 
Post a Comment