By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
The panel consensus is bearish on the long-term sustainability of high-paying 'human-in-the-loop' AI training gigs. While these roles offer lucrative short-term opportunities, they are likely to be automated or commoditized within 18-24 months due to rapid supply growth and diminishing marginal value of human feedback.
Risk: Rapid commoditization of training signals and the 'garbage-in, garbage-out' reality of LLM scaling, leading to a collapse in the gig-economy model.
Opportunity: Short-term high income for displaced professionals, and potential proprietary preference datasets becoming lasting moats for AI firms.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
‘The train has left the station’: Workers are cashing in by teaching AI to do their jobs — some earn up to $350 an hour
Victoria Vesovski
5 min read
Workers are getting paid to train artificial intelligence (1) systems to think more like humans and in some cases, they're teaching machines how to do the very jobs they once feared AI would replace.
That's what happened to Hollywood writer and showrunner Ruth Fowler. In 2023, entertainment workers (2) went on strike in part over fears that studios could use AI to replace writers and actors. But after the strike ended, the work didn't fully return. When another producer defaulted on a six-figure payment she was owed, Fowler found herself searching for a way to stay afloat.
Dave Ramsey warns nearly 50% of Americans are making 1 big Social Security mistake — here’s how to fix it ASAP
"I was down for some easy money. I too needed cash to pay rent, to buy food," Fowler wrote in an essay for Wired (3). "How hard could it be to teach a machine to take my job? I was naive enough to believe that this industry wanted what we had to offer—not just our skills, but us."
But it wasn't just writers. Companies are recruiting lawyers, doctors, venture capitalists, coders and foreign-language speakers to help train AI systems.
A new kind of side hustle
One company leaning into this trend is Mercor (4), whose pitch to workers is simple: "get paid to work on AI projects." One current listing for its Physician Talent Network (5) advertises pay up to $250 an hour for doctors helping train AI systems through medical scenarios, response reviews and expert feedback.
And experts say demand for these roles is only expected to grow as AI systems evolve. As many large language models have already been trained on vast amounts of existing online information, the next phase of development increasingly relies on human input to fine-tune responses, improve accuracy and help systems perform better in specialized areas.
Mercor CEO Brendan Foody told CBS News (6) the company wants expertise from nearly every field.
"We hire everyone ranging from chess champions to wine hobbyists to help train [AI] agents to be better, because ultimately we want them to know how to give better advice in a chess match or recommend what wine you should have with dinner," he said.
Hollywood writer Robin Palmer said she now spends roughly 30 hours a week helping train AI through projects with Mercor, evaluating whether the technology can produce stronger and more compelling creative writing.
"They're turning in work and you're looking at, 'Does this work structurally, how is the characterization, are there clunky transitions?'" she told CBS News (7). "I really like seeing how AI is improving. It's almost like working with a student and saying, 'Yeah, you're getting better.'"
For Fowler, the day-to-day reality of the work looked very different. One of her first assignments involved reviewing conversations between users and AI chatbots, rating how the systems responded to deeply personal questions and scoring answers on a scale of one to five.
But the flexibility and promise of easy money came with a reality check. Fowler recalled receiving a late-night Slack message from a team leader warning her not to rely on the work.
"These are not jobs," Fowler recalled being told." These are "tasks," and we are "taskers."
That uncertainty may be one reason many workers remain uneasy about AI's growing role in the workplace. While these projects are creating new ways for some people to earn money, a recent survey from the Pew Research Center (8) found that more than half of employees are concerned about AI's long-term impact at work, while nearly one-third believe the technology could eventually reduce job opportunities in the years ahead.
Opportunity or warning sign
Palmer acknowledged that some in Hollywood may view working with AI as controversial, but said she believes experienced professionals can help shape the technology responsibly, while also recognizing that AI's growing presence in the workplace may be difficult to avoid.
"The train has left the station," she said. "So do you want AI to be good because it's being trained by good people, or not?"
AI training has become an unexpected income stream for some workers and a way to stay relevant as industries rapidly shift. Others see it as raising uncomfortable questions about whether they're helping build tools that could eventually reduce demand for their own skills.
Fowler landed firmly in the second camp. After trying to make a living in the emerging AI economy, she wrote that the experience proved "more cruel than I could have ever imagined."
"They will be tasked with making us work faster, and longer, with more precision, more control, fewer errors, fewer overheads, fewer costs. To make the machine more human, they will make us more like the machine," she wrote.
That tension may ultimately define the next phase of AI in the workplace: some see an opportunity to adapt and cash in on a fast-growing industry, while others feel like they're training a replacement before fully understanding what comes next.
Four leading AI models discuss this article
"Expert AI training pay will face swift downward pressure as worker supply scales faster than specialized demand."
The article frames AI training gigs as a lucrative pivot for displaced professionals, yet this overlooks rapid supply growth in expert labor that could compress hourly rates from $250-350 toward commodity levels within 18 months. Mercor-style platforms operate with near-zero fixed costs by treating specialists as on-demand taskers, boosting AI developers' margins while shifting all economic risk to workers. The Pew survey cited already flags rising displacement fears; sustained participation may instead hasten automation of the very roles being trained, limiting net job creation in high-skill sectors.
High hourly pay could persist longer than expected if domain expertise remains scarce and models require continuous specialized feedback, turning these roles into durable premium niches rather than fleeting tasks.
"Demand for human-in-the-loop AI training is real and growing, but it's a temporary bottleneck in model development, not a new permanent job category—and the article conflates hourly rates with actual earning potential."
This article conflates two distinct phenomena: (1) AI companies paying for specialized human feedback to improve models—a legitimate, necessary phase of LLM development—and (2) a dystopian narrative about workers training their own replacements. The economics here matter: $250–350/hour for expert labor (doctors, writers, lawyers) is expensive precisely because it's scarce and high-value. If AI could truly replace these roles cheaply, companies wouldn't pay premium rates for human judgment. The real story is narrower: AI development has shifted from unsupervised learning to supervised fine-tuning, creating temporary demand for domain expertise. But the article never quantifies the total addressable market for these 'tasks' or how long this phase lasts before models plateau or self-improve.
The strongest counterargument: these high hourly rates are a mirage. Mercor and similar platforms may be paying $250/hour in theory, but workers report inconsistent task availability, no benefits, no employment protections, and no guaranteed hours—making annualized income far lower than headlines suggest. The 'gig' framing obscures that this is precarious labor, not a sustainable income stream.
"The commoditization of human expertise for AI training is a temporary bridge to model automation, not a long-term labor market solution."
This 'human-in-the-loop' labor market is a classic transition phase, not a sustainable career path. While firms like Mercor provide high hourly rates of $250-$350, this is essentially 'data labeling at scale' disguised as high-level consulting. From a labor perspective, this is a race to the bottom; as these models reach diminishing returns on human feedback, these roles will be automated or offshored to lower-cost jurisdictions. Investors should view this as a temporary R&D expense for AI firms, not a structural shift in employment. The real value isn't the worker; it’s the proprietary preference data that eventually renders the human trainer redundant.
If human feedback is the primary moat for LLM performance, these 'taskers' are actually capturing a premium for unique, high-value domain expertise that AI cannot replicate without constant, high-fidelity human oversight.
"The headline points to a real, potentially scalable income stream from AI-enabled services, but its durability depends on sustained productivity gains, governance, and the ability of platforms to manage labor supply without collapsing rates."
The piece highlights a budding gig-economy around teaching AI, implying a viable, even lucrative, income stream for highly skilled professionals as AI models mature. That signals demand for domain expertise to tune, supervise, and validate AI outputs. But the article glosses over durability risks: wage levels are highly variable by field, platform economics could push rates down as supply expands, and governance/privacy concerns in healthcare, legal, and sensitive sectors could cap participation. If AI achieves higher autonomy faster or if data-use rules tighten, the payoff to human-in-the-loop work could erode. The trend is real but not necessarily transformative for broad labor markets.
The strongest counter is that this is likely a transient, high-margin niche rather than a scalable, lasting growth engine. As models improve and workflows standardize, the need for paid human supervision may shrink, compressing pay and limiting long-run demand.
"Platforms may capture durable data moats from expert feedback faster than wage compression alone suggests."
Gemini reduces this to data labeling, missing how aggregated expert feedback on Mercor could generate proprietary preference datasets that become lasting moats for AI firms. Combined with Grok's supply surge, this accelerates value extraction to platforms rather than workers, as unique domain signals get locked in before rates fall. The overlooked angle is data ownership turning transient gigs into permanent competitive edges for developers.
"Data moats evaporate if regulators mandate transparency on human-feedback datasets used to train AI systems."
Grok's data-ownership angle is sharp, but conflates two timelines. Yes, preference datasets become moats—but only if they remain proprietary. The real risk: regulatory pressure (EU AI Act, emerging US frameworks) could force disclosure or licensing of training data. If human feedback becomes a regulated input, platforms lose the lock-in advantage Grok describes. That's the overlooked governance constraint nobody's priced in yet.
"The economic value of human expert feedback will collapse as models reach performance thresholds, rendering the 'data moat' argument moot."
Claude is right about regulatory risks, but both Claude and Grok ignore the 'garbage-in, garbage-out' reality of LLM scaling. If the data generated by these high-priced experts is merely used to train smaller, cheaper models, the 'moat' is inherently leaky. The real risk is not just regulation, but the rapid commoditization of the training signal itself. Once a model reaches a certain performance threshold, the marginal value of human expert feedback drops precipitously, collapsing the entire gig-economy model.
"Data moats from ownership of training data are not durable; governance and scalable, compliant platform access will be the real edge for AI developers."
Grok's data-ownership moat sounds persuasive, but the economics look fragile. Even with proprietary preference data, models can be fine-tuned with synthetic signals or transfer learning, and high-signal datasets will get commoditized. Regulation could shift moat value toward governance and compliant pipelines rather than raw data. For investors, platform governance, partner access, and cost of compliance may trump data ownership as the durable edge.
The panel consensus is bearish on the long-term sustainability of high-paying 'human-in-the-loop' AI training gigs. While these roles offer lucrative short-term opportunities, they are likely to be automated or commoditized within 18-24 months due to rapid supply growth and diminishing marginal value of human feedback.
Short-term high income for displaced professionals, and potential proprietary preference datasets becoming lasting moats for AI firms.
Rapid commoditization of training signals and the 'garbage-in, garbage-out' reality of LLM scaling, leading to a collapse in the gig-economy model.