Google hired microworkers to train its controversial Project Maven AI

By Makena Kelly

Illustration by Alex Castro / The Verge

According to a new report from The Intercept, Google hired gig economy workers to help build out a controversial artificial intelligence program that the company had paired with the Pentagon to build.

The workers were hired through a crowdsourcing gig company outfit called Figure Eight, which pays as little at $1 an hour for people to perform short, seemingly mindless tasks. Whether the individuals were identifying objects in CAPTCHA-like images, or other simple tasks, the workers were helping to train Google’s AI that was created as part of a Defense Department initiative known as Project Maven.

Project Maven is a Pentagon project intended to use machine learning and artificial intelligence in order to differentiate people and objects in thousands of hours of drone footage. By employing these crowdsourced microworkers, Google was able to use them to teach the algorithms it was running how to distinguish between human targets and surrounding objects.

According to The Intercept, these workers had no idea who their work was benefitting or what they were building.

Last June, Google said that it had decided not to renew its contract with the Defense Department as it involved Project Maven after over 3,000 employees signed a petition in protest of the company’s involvement in the initiative. The deal is set to end in March 2019.

Figure Eight, which was previously known as Crowdflower, is one of the largest platforms that employs microworkers. On its website, Figure Eight says its platform “combines human intelligence at scale with cutting-edge models to create the highest quality training data for your machine learning (ML) projects.” By partnering with these microworker outfits, Google could quickly and cheaply build out its AI.

“You upload your data to our platform and we provide the annotations, judgments, and labels you need to create accurate ground truth for your models,” the website reads.