20.07.2018 Views

Fortune

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

MULLIGAN: BENJAMIN TICE SMITH; EX MACHINA: A24/COURTESY EVERETT COLLECTION; BATTLESTAR GALACTICA: FRANK OKENFELS—SCI-FI/PHOTOFEST;<br />

THE MATRIX: WARNER BROS./PHOTOFEST; METROPOLIS: UFA/PHOTOFEST; ULTRON: WALT DISNEY STUDIOS MOTION PICTURES/COURTESY EVERETT COLLECTION<br />

say, ‘Oh, now I can do accounting at home,’ ”<br />

says Kate Crawford, principal researcher<br />

at Microsoft and codirector of the AI Now<br />

Institute at New York University. “These are<br />

very advanced systems that are going to be<br />

influencing our core social institutions.”<br />

HOUGH THEY MAY not think of it as<br />

such, most people are familiar<br />

T with at least one A.I. breakdown:<br />

the spread of fake news<br />

on Facebook’s ubiquitous<br />

News Feed in the run-up to the 2016 U.S.<br />

presidential election.<br />

The social media giant and its data scientists<br />

didn’t create flat-out false stories. But the<br />

algorithms powering the News Feed weren’t<br />

designed to filter “false” from “true”; they were<br />

intended to promote content personalized to<br />

a user’s individual taste. While the company<br />

doesn’t disclose much about its algorithms<br />

(again, they’re proprietary), it has acknowledged<br />

that the calculus involves identifying<br />

stories that other users of similar tastes are<br />

reading and sharing. The result: Thanks to an<br />

endless series of what were essentially popularity<br />

contests, millions of people’s personal<br />

News Feeds were populated with fake news<br />

primarily because their peers liked it.<br />

While Facebook ofers an example of how<br />

individual choices can interact toxically with<br />

A.I., researchers worry more about how deep<br />

learning could read, and misread, collective<br />

data. Timnit Gebru, a postdoctoral researcher<br />

who has studied the ethics of algorithms at Mi-<br />

A.I. SPECIAL REPORT<br />

machine-learning<br />

algorithms “haven’t<br />

been optimized for any<br />

definition of fairness.<br />

They have been<br />

optimized to do a task.”<br />

> deirdre mulligan<br />

associate professor, uc berkeley school of information<br />

crosoft and elsewhere, says she’s concerned about how deep learning<br />

might afect the insurance market—a place where the interaction of<br />

A.I. and data could put minority groups at a disadvantage.<br />

Imagine, for example, a data set about auto accident claims. The<br />

data shows that accidents are more likely to take place in inner<br />

cities, where densely packed populations create more opportunities<br />

for fender benders. Inner cities also tend to have disproportionately<br />

high numbers of minorities among their residents.<br />

A deep-learning program, sifting through data in which these<br />

correlations were embedded, could “learn” that there was a relationship<br />

between belonging to a minority and having car accidents, and<br />

could build that lesson into its assumptions about all drivers of color.<br />

In essence, that insurance A.I. would develop a racial bias. And that<br />

bias could get stronger if, for example, the system were to be further<br />

“trained” by reviewing photos and video from accidents in inner-city<br />

neighborhoods. In theory, the A.I. would become more likely to conclude<br />

that a minority driver is at fault in a crash involving multiple<br />

drivers. And it’s more likely to recommend charging a minority<br />

driver higher premiums, regardless of her record.<br />

... to malevolent<br />

Or will they develop their own agendas, in pursuit of a colder logic that sees us as a distraction or a threat? It’s the “intelligence” of A.I., the facet<br />

that makes a computer almost human, that compels us to ask these questions again and again. —Matt Heimer and Armin Harris<br />

www.t.me/velarch_official<br />

AVA<br />

Ex Machina<br />

NUMBER SIX<br />

Battlestar<br />

Galactica<br />

AGENT SMITH<br />

The Matrix<br />

FALSE MARIA<br />

Metropolis<br />

ULTRON<br />

Avengers:<br />

Age of Ultron<br />

pure evil<br />

59<br />

FORTUNE.COM // JULY.1.18

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!