The Rise and Future of AI’s White-Collar Gig Economy, ETTelecom
By Parmy Olson
A common proclamation made by tech leaders is that while artificial intelligence will destroy jobs, it will also create many new ones. But what kinds of new careers will AI spark? And, more importantly, will they last?
Few spell out what will replace the swathes of white-collar roles that could be automated out of existence in the coming years. New forms of employment such as AI researchers and prompt engineers are few and far between. And data from the World Economic Forum suggests that although there will be a net increase of 78 million jobs from around now till 2030, the fastest growth will be in roles including farmworkers, delivery drivers and construction workers — all driven by demographic shifts rather than a productivity boom from AI.
In an ideal world, AI would trigger something similar to the automation of automobile manufacturing in the early 20th Century, when the knock-on effect of more cars led to booms in other parts of the economy like transportation and retailing, even as skilled craftspeople were replaced. Daron Acemoglu, who won the Nobel Prize for economics in 2024, says in the book Power and Progress that current AI trends skew more heavily towards cost-cutting; with some AI tools only doing a nominally better job than humans — like customer-service bots — the resulting productivity gains are minor. The creation of new jobs also looks less likely. Yet there is one clear example of “new work” being triggered by AI: professional trainers for the software models. Several startups with names like Surge AI (valued last year at $25 billion) and Turing (valued at $2.2 billion, according to Pitchbook) populate a new market hiring white-collar professionals to train AI systems, usually as contractors, to form a new kind of office-worker gig economy. Among the biggest is Mercor.io Corp., founded in 2023 in San Francisco and valued at $10 billion, and frequently named as one of the fastest growing startups in AI.
Around 30,000 people from a wide range of disciplines — lawyers, doctors, financial consultants, cooks, osteopaths — are paid hourly by Mercor to train AI models to become more proficient in those fields, often for clients like OpenAI, Anthropic and Google. Mercor plans to grow its contractor base “by many orders of magnitude,” founder and Chief Executive Officer Brendan Foody tells me.
Foody, who is 22 and a graduate of the coveted Thiel Fellowship program, says he’s been fascinated by labor markets since high school. To chase rocket-ship growth he limits some meetings, like our interview, to an efficient 15 minutes and requires customer-facing staff to work six days a week, with the option to work from home on Saturdays. The company automates its recruitment process for contractors, screening new recruits through AI-powered video interviews.
Their work is done in secret, protected by hefty non-disclosure agreements, and is technically challenging. Contractors sometimes create scoring guides known as “evals” or “rubrics,” which are then used to teach AI models and evaluate their responses, a bit like designing the bar exam to test human lawyers. Grim and paradoxical as it seems to train an AI that could replace you, Mercor’s contractors say the pay is excellent. One of them, who declined to be identified, told me they’ll never make as much money as they are right now.
But a question looms: When does the music stop? While tech leaders including Sam Altman and Anthropic’s Dario Amodei say super-intelligent AI is imminent, Foody says that threshold is a decade or more away. AI models will be a lot like software, he argues, which has evolved over decades with lots of human input.
Of course, it’s in Foody’s interests to claim ample runway for his startup, but he might be right. Large language models were initially trained on huge amounts of text and images on the internet, and then fine-tuned with the help of low-paid data labelers in emerging markets like Kenya, where contractors famously helped filter harmful content from early ChatGPT. But tech firms increasingly want to hire subject-matter experts for that work to make AI models smarter.
Scale AI, which once paid workers in the Philippines and India as little as one cent per task, now recruits contractors with master’s degrees and PhDs to do more complex training. (Meta Platforms Inc. bought a 49% stake in Scale and hired its CEO for $14.3 billion last summer.) Surge AI offers as much as $1,000 an hour for expertise from startup founders and venture capitalists, and has contracted more than 20,000 professionals with doctoral degrees. The market for expert AI trainers appears to be growing: Foody says his “frontier customers” are increasing their budgets for hiring humans by five to ten times.
But let’s presume there are roughly 100,000 professionals doing expert AI training work today, on top of the millions more doing the traditional low-paid data labeling. It is hard to tell if this will be the new form of work that fills the hole AI could leave in labor markets, while many other jobs morph. Those changes will be amorphous and difficult to track. “In the past you might have a marketer paired up with an engineer and a designer to work on a project together,” says Wade Foster, the CEO of workflow automation company Zapier Inc., based in San Francisco. “Now that might just be one person. You’ll see a lot of hybrid jobs and squishing of the job titles.” Mark Zuckerberg recently said a single good engineer at Meta could now do the work previously carried out by a large team, thanks to AI.
It’s unclear if that increased efficiency will mean less hiring overall. British software company Sage Group plc saw its revenue rise 10% last quarter thanks in part to the use of AI tools internally, says CEO Steve Hare, who went on to hire more people. Headcount rose 2%.
But for this new generation of professional gig workers, the future looks murky when you consider what happened to the older market of lower-paid data labelers. Many are now being squeezed out of their roles, with software handling simpler tasks and the new, expert trainers taking over complex ones. The lawyers, doctors and former journalists who work for companies like Mercor might follow a similar arc. One tells me they expect the work to last another two years, maybe five if they’re lucky. In the end, expert AI trainers will likely meet the fate of automation too.
