They helped score a raise for Google AI ‘raters.’ Then they were terminated

Just before Memorial Day, Ed Stackhouse got an email from his company, offering bonuses to anyone willing to put in some extra hours of work. 

The company, Appen, is an Australian firm that contracts with Google and hires hourly workers called “raters” to assess the quality of things like search results and ads. Stackhouse, 49, had been working as a rater for about 10 years, starting with a company that Appen acquired. But in the months since Google began rolling out generative AI tools, such as Bard, the demand for AI raters had shot up—and Appen was sending out emails all but begging people to put in more hours.

“As a team, we have been hitting record-high production levels and it is all thanks to each and every one of you!” the optimistic email, sent May 25, read. “Keep an eye on your email for another fun bonus game happening next week!”

Stackhouse did receive another email from the company the following week, but it wasn’t to alert him of “another fun bonus game.” It was a termination notice. Stackhouse was fired after a decade of employment; the email only cited “business conditions” in its explanation.

Stackhouse was stunned. “‘Business conditions at Appen have been such that they have been offering nonstop bonuses to get people to work max hours,” he tells Fast Company. 

Stackhouse couldn’t help but feel that there was more to the story. Since last year, he had been part of a small group of members of the Alphabet Workers Union (AWU), petitioning Google and Appen to bring wages for raters in line with the $15 hourly minimum that Google requires other vendors to offer. At the time, some raters at Appen were reportedly making as little as $10 an hour. AWU, which represents both full-time Googlers and Google contractors (and is even referred to informally as Google’s union), operates within the Communications Workers of America (CWA), the largest communications and media labor union in the U.S.

In December, AWU-CWA scored a victory for Appen raters, securing them $14- to $14.50-an-hour wages. It was less than what they’d been fighting for, but still constituted the first-ever pay raise for raters, amounting to an estimated $10 million in salary increases for between 3,000 and 5,000 workers, according to AWU-CWA. 

Stackhouse continued to be an outspoken advocate for Appen raters, arguing for healthcare benefits, among other things. Now, according to AWU-CWA, he and at least 10 other raters—many of whom have recently talked publicly about working conditions at the company—have been terminated.

“It feels a little personal,” says Michelle Curtis, an Appen worker of nearly nine years, who also had pushed for pay raises and was let go the same day and in the same fashion as Stackhouse. “I would love some answers as to why I was actually terminated, and why this small group of workers was terminated.”

Appen did not respond to Fast Company’s requests for comment, but the company did cite “headwinds from the broader technology market slowdown” in a recent market filing. Despite the increased demand for AI-training services, the company has seen revenue decline compared to the boom years of the pandemic, particularly as major clients like Meta have downsized. In the first quarter of this year, revenue fell to $95.7 million, about a 21% decrease compared to the prior year. 

Stackhouse says that he and other union members are now in discussions with lawyers for AWU-CWA about their options. “The fired workers will be exploring every avenue available to them to rectify this injustice with support from their fellow union members. Google raters will continue to organize to build power and hold Google and Appen accountable,” Parul Koul, a Google software engineer and executive chair of AWU-CWA, told Fast Company in a statement. 

The tensions between Appen and its raters highlight the often unseen and underpaid human labor that makes AI systems like ChatGPT and Bard work. As AI ethicists have frequently pointed out, AI systems didn’t magically evolve all on their own. They were trained and honed by huge numbers of low-wage workers around the world.

The Partnership on AI, a nonprofit backed by Google and other tech giants, has put out guidelines for “responsible sourcing of data enrichment services,” which specifies that AI companies should, among other things, pay above a living wage. Google has set a $15 hourly wage standard for vendors with access to Google badges and its campus, but AWU-CWA members argue that Google should also push for wage increases among contracting firms like Appen.

In a statement, a Google spokesperson pointed Fast Company to the company’s policy on vendor wages, reiterating that the policy does not apply to Appen’s raters. “Raters work part-time from home, can be assigned to multiple companies’ accounts at a time, and do not have access to Google’s systems or campuses,” the spokesperson wrote. “Like all our suppliers, Appen is responsible for the working conditions of their employees, including pay, benefits, and the tasks they’re assigned.”

In his 10 years as a rater for Appen and its predecessor, a company called Leapforce, Stackhouse says he never saw a pay raise. He’d found the job on Craigslist back in 2013, after being laid off from a gig dealing cards on a boat, and started as an independent contractor, working for $13.50 an hour for up to 40 hours a week. It wasn’t much, but Stackhouse, who has diastolic heart failure, liked the fact that he’d be able to work from his home in Asheville, North Carolina. The work mostly involved comparing two possible search results for a given query and judging which of the results was best.

Four years into the job, though, Leapforce cut back on the number of hours raters could work—reducing the limit from 40 hours to just 26. Shortly after, in 2017, Leapforce and a sibling company, RaterLabs, were acquired by Appen. At Appen, Stackhouse and Curtis became part-time employees, and Stackhouse’s hourly rate was reduced to $13; Curtis’s pay to even less—$12.75 an hour. 

For Stackhouse, a father of three who is married to a teacher, that meant making less than his daughter does at a fast-food job. But he stuck with the role, in large part, due to his health. “I can’t go out there and get a job in a workplace like everyone else,” he says. 

Both Stackhouse and Curtis joined AWU-CWA last year and were part of a group of raters who met virtually with Appen’s then-CEO Mark Brayan, who has since left the company. According to AWU-CWA, three of the workers at that meeting have now been let go.

Just before the new year, Appen notified raters that beginning in January, their pay would go up, constituting a major win for Stackhouse, Curtis, and their colleagues. “We celebrated,” Curtis says. “It was great.”

But the pay increase had only been a half step, she says. The union members had also pushed for things they didn’t get, including healthcare benefits, access to human resources contacts, and basic clarity on who their supervisors were. “We have had no idea over the years who has been in charge or who we report to,” Stackhouse says. “If you had an issue, you would send in a ticket to an anonymous person.” 

Stackhouse and his colleagues continued to advocate for these changes throughout the early part of this year. At the same time, the work they were doing began to change, as rating tasks began to move away from search and ads and toward Google’s AI systems. These tasks were substantially more challenging, Stackhouse says, requiring raters to make speedy decisions about which AI-generated responses to a given prompt were most accurate.

“I’ve had some [tasks] where I have to verify 10 or 12 things in two minutes,” he says, noting that if he took longer than expected to complete a task, it could become an issue when it was time to get paid. This not only hurt morale among raters, but, given that their role is to rate the quality of results, Stackhouse says it made him question how all of this might ultimately impact the efficacy of increasingly powerful and ubiquitous AI systems. “The pressure is on to get finished with this task in time, damn the accuracy,” he says.

Stackhouse was in the middle of a task on May 31 when he found himself suddenly locked out of the rating portal and the company’s chat function. It was only later that he received the email telling him he’d been let go. And just before the termination email came through, both he and Curtis received a separate email with an unsigned version of the “RaterLabs Non Disclosure Agreement” pasted into the body. Neither one of them remembers ever signing such an agreement for Appen.

This is not the first time Google raters have raised concerns about potential retaliation for speaking out about working conditions. Under Leapforce, which also performed contract work for Google, at least one rater reported being fired after speaking to Ars Technica about Leapforce’s reduction of raters’ hours. 

For now, Curtis is still hoping to be reinstated. “I really enjoyed this job,” she says. “I’ve never been fired from any job I’ve ever had. So it feels like a little bit of a loss.” 

Stackhouse, meanwhile, is considering his limited options. He’s not thrilled about the prospect of taking a rating job at another contractor, where he will inevitably face some of the same challenges. But he has rent to pay, and the bill for a new car he purchased just before he was fired will be due soon. “There’s a whole lot of money that I’m going to have to come up with,” he says. “I have no idea how I’m going to do it.”