Hundreds of human reviewers across the globe, from Romania to Venezuela, listen to audio clips recorded from Amazon Echo speakers, usually without owners’ knowledge, Bloomberg reported last week. We knew Alexa was listening; now we know someone else is, too.
This global review team fine-tunes the Amazon Echo’s software by listening to clips of users asking Alexa questions or issuing commands, and then verifying whether Alexa responded appropriately. The team also annotates specific words the device struggles with when it’s addressed in different accents.
According to Amazon, users can opt out of the service, but they seem to be enrolled automatically. Amazon says these recordings are anonymized, with any identifying information removed, and that each of these recorded exchanges came only after users engaged with the device by uttering the “wake word.” But in the examples in Bloomberg’s report—a woman overheard singing in the shower, a child screaming for help—the users seem unaware of the device.
“We have strict technical and operational safeguards, and have a zero tolerance policy for the abuse of our system,” Amazon said in an emailed statement to Bloomberg, noting that only an “extremely small” sample of records are annotated.
The revelation tells us a lot about how “smart” devices really work. Alexa-enabled speakers can and do interpret speech, but Amazon relies on human guidance to make Alexa, well, more human—to help the software understand different accents, recognize celebrity names, and respond to more complex commands. This is true of many artificial intelligence–enabled products. They’re prototypes. They can only approximate their promised functions while humans help with what Harvard researchers have called “the paradox of automation’s last mile.” Advancements in AI, the researchers write, create temporary jobs such as tagging images or annotating clips, even as the technology is meant to supplant human labor. In the case of the Echo, gig workers are paid to improve its voice-recognition software—but then, when it’s advanced enough, it will be used to replace the hostess in a hotel lobby.
A 2016 paper by researchers at Stanford University used a computer vision system to infer, with 88 percent accuracy, the political affiliation of 22 million people based on what car they drive and where they live. Traditional polling would require a full staff, a hefty budget, and months of work. The system completed the task in two weeks. But first, it had to know what a car was. The researchers paid workers through Amazon’s Mechanical Turk platform to manually tag thousands of images of cars, so the system would learn to differentiate between shapes, styles, and colors.
It may be a rude awakening for Amazon Echo owners, but AI systems require enormous amounts of categorized data, before, during, and after product launch. The ideal state is a feedback loop: The Echo performs decently, voice data from customers are collected and used to improve the service, and then more people buy it as it improves and more data are collected, further refining it. This is true for very different types of AI products.
For example, human workers watch and manually tag footage uploaded from Amazon’s Ring products, surveillance cameras that homeowners can install on their doorbells and front porches. Facebook’s content-moderation AI relies on thousands of people across the globe teaching software what counts as objectionable in different contexts. Another loop: Humans flag content, the AI gets better at detecting it, more people use Facebook as it gets safer, and the AI gets smarter as more content is flagged. In all cases, Silicon Valley would have us believe that AI is smart enough to replace humans, when in reality it only works because of the role of hidden human labor in creating and maintaining these loops. AI is always a human-machine collaboration. It can accomplish incredible feats, but rarely alone.
Astra Taylor, a writer and documentary director who studies what she calls “the automation charade,” notes that companies have a vested interest in obscuring the role of humans in automation’s last mile. “If ‘friction’ is actually knowing it’s a human being on the other side of your smart speaker,” she says, “then that’s a discomfort we have to face.”
The Bloomberg report also featured a harrowing account of Amazon’s global team overhearing what might have been a sexual assault or violent encounter. Workers reported to Bloomberg that Amazon superiors told them to ignore the sounds and not file a police report. (When reached for comment, Amazon referred The Atlantic to statements emphasizing encryption and the security of its cloud storage.) Taylor notes this is another of automation’s paradoxes: To make machines smarter, the humans that train them have to act dumber.
“What is so tragic about that piece is, it’s also forcing the humans to be less intelligent, to be less empathic,” she says. “They have to be trained basically to say, ‘Oh, that’s not relevant. I’m going to be less human, less affected by that.’”
Taylor argues that hiding human labor means people act more like machines, without personal ethics or empathy. In a more extreme case, obscuring humans means removing choice entirely. In February, The Intercept reported that Google contracted with the Department of Defense to augment government drones’ computer vision systems, outraging employees. Drones with computer vision systems are capable of “seeing” their targets—but they still need refinement and precision, so Google contracted hundreds of human workers to manually tag items and help create its data set. According to The Intercept’s report, they had no idea what their labor was being used for.
Humans are part of what makes any AI system work, and they bring all their dumb, human baggage with them: privacy and eavesdropping concerns, potential biases, individual preferences and judgments. That doesn’t always serve the interest of making society frictionless, smart, or convenient. But hiding their role—essentially trying to make the human-machine systems that engineer AI less human—only serves Silicon Valley’s fiction that AI is powerful, rational, and inscrutable; the same fiction the industry tells about itself. Exposing the human labor of advanced technology uncovers both.