AI chatbots are the new digital sweatshops

A couple of days ago I got a weird email. Well, I get a lot of weird emails, but this one stuck out. It asked me if I wanted to apply for a job as an "AI English Writer."

The job was remote, had flexible hours, and would allow me to "make an impact on making LLMs [large language models] safer, more accurate, and more reliable." It also paid "up to $25 an hour."

I thought, well, how could I resist? I hadn't made "up to $25 an hour" since I was a stock clerk in my teens. [1]

It got weirder when an old friend and colleague, like me a gracefully aging tech journalist, told me he'd gotten a similar email. And then on Threads another established tech journalist reported getting this email. So I decided to apply, just to see what all this was about.

The link took me to a job description for a company I’d never heard of called Outlier. The requirements for this job were a little steep for something that pays slightly better than fast food wages: It wanted me to have at least a bachelors degree from a "Top 50 University" in a writing related field (though a masters from a "Top 100" university would be better). I had to submit a resume, a cover letter [2], my LinkedIn credentials, and my GPA (!) [3]. I also had to agree to a background check and a brief video interview, then pass an exam, assuming I made it that far.

Something about this wasn't quite right, so I decided to dig a little deeper. And, as it turns out, it was very much not right.

Working for Scale

Outlier is affiliated with a company called Scale AI. Scale AI churns out training data for large language models that form the heart of generative AI tools like ChatGPT and Google Gemini. Its clients include Tesla, Microsoft, Meta, OpenAI, Toyota, Nvidia, and Harvard Medical School. It is kind of a big deal.

Scale AI has received more than $600 million in venture investments and has a valuation in excess of $7 billion. It took in nearly $300 in revenue in 2022. It is rolling in money. But you wouldn't know that based on what it pays its data lackeys.

Here's the thing about large language models. They consume data like a blue whale eats krill, and the companies that operate them are willing to pay good money to keep their whales well fed. Scale AI needs its training data to be clean and accurate in order to sell it to Microsoft and Tesla. And to make sure the data is clean and at least somewhat accurate, it hires humans to label images and generate text in languages that aren't as common on the Internet, like Urdu and Bengali.

It also turns out that "up to $25 an hour" is caviar wages for Scale AI. Most of the time it pays far, far less, when it pays people at all.

Last August, The Washington Post ran an expose on Scale and its other gig economy management partner, Remotask. It focused on the 'digital sweatshops' operated by Scale/Remotask in the Philippines, but the same policies apply across multiple global locations. Wages for 'taskers' (the people hired to tag or correct the training data) were plummeting, and many people didn't get paid for the work they performed. Per that report:

"Initially, taskers said, they could earn as much as $200 in a week. Then in 2021, around the time Remotasks expanded to India and to Venezuela, pay rates plunged, according to workers and screenshots of project assignments. Filipino freelancers went from earning $10 per task on some projects to less than 1 cent..."

On its website, Scale AI declares in large print that it provides "fair and competitive compensation." [4]

Yeah, maybe. But I kinda doubt it.

Fairwork, a University of Oxford project that measures the working conditions for the global gig economy, awarded Scale/Remotasks a score of one out of ten in 2023. It got one point for taking measures to prevent taskers from working too many hours, and zero points for fair pay, fair conditions, fair contracts, fair management, or fair representation.

Don't ask, don't tell

I wanted to know what it was like to work for a company like this, so I reached out to five people on LinkedIn who'd listed Outlier as one of their freelance gigs. I wanted to know what their jobs were like — what kind of tasks Scale/Outlier was asking them to perform, what they got paid to do these things, and how they liked it.

Two people have gotten back to me so far. Both said that their work contracts forbid them from saying anything to anyone about what they did for Outlier. [5]

There are only three kinds of people who cannot under any circumstances talk about what they do for a living: 1) paid assassins, 2) spies, and 3) data training wallahs working for Scale AI.

If you have to swear your underpaid workers to secrecy, that doesn’t speak well about the kind of company you're running, does it?

This is the dark underbelly of the generative AI revolution that nobody talks about, and I was frankly unaware of until now. For chatbots to work their magic, it takes thousands of ordinary human beings to tell them what they're actually seeing. The people who run these companies are eating like kings; the humans who make it possible are lucky to get table scraps.

The Fairwork report put it pretty well:

While automation, bias and other problems and the workplace have attained public attention, the massive work behind AI development is still subject to little scrutiny and public awareness ... our findings and worker evidence suggest that there is a long road ahead to ensure basic standards of fair work for these millions of workers who are essential to the evolution of AI but remain mostly invisible to companies, policymakers, and society.

I'll let you know if I get the job. [6]

Like this post? Don't keep it to yourself — share the wealth with others. The buttons on the upper left make it semi-easy.

[1] Technically, I was paid just over $3 an hour as a stock clerk. In today's dollars, that would be the equivalent of just under $15 an hour. But I think the operative phrase in that description is "up to."

[2] I used ChatGPT to write it, naturally.

[3] How exactly were they planning to check my GPA?

[4] Setting aside for the moment that, as a resident of the San Francisco Bay Area, "up to $25 an hour" would be barely enough to keep my cat in Friskies.

[5] I am guessing that Scale started using Outlier instead of Remotasks as a result of the bad press coverage they were getting. Ditto re their ‘don’t ask/don’t tell’ policies. Total speculation on my part, but…

[6] After I connected my LinkedIn account to the application form, I received a second, slightly more personalized invitation to apply for this job, with a salary bump: $35 an hour, no “up to” included. I feel deeply honored.