A generalist labeling images for an AI model earns somewhere between $12 and $25 an hour. A board-certified psychiatrist designing clinical scenarios for the same class of model can earn up to $350. That gap is the whole story of where AI-related online income is heading, and most of the “get paid to train AI” advice still circulating describes the half that is quietly disappearing.
The company making that gap visible is Mercor. It pays out more than $1.5 million a day to a network of roughly 30,000 contractors, and its advertised average sits north of $85 an hour (herohunt). Its annualized gross revenue crossed $2 billion by June 2026, up from $760 million at the end of 2025 (Dealroom). It closed a $350 million round at a $10 billion valuation in October 2025 (CNBC), and by July 2026 was reportedly in talks to raise again at $20 billion (TechCrunch). Alongside it, Handshake has built a competing marketplace connecting skilled professionals to AI labs (CBS News).
Those are the headline numbers. The number that actually matters for anyone trying to earn from this is the one hiding inside that $85 average: it is a blend of a floor that is falling and a ceiling that is climbing, and the middle is thinning out.
The two rates that tell the whole story
Mercor’s published pay structure sorts contractors into tiers, and the tiers are not close together. Entry-level generalist labeling and basic reinforcement work pays roughly $12 to $25 an hour. Specialized evaluation, coding review, and translation land around $25 to $53. Credentialed professionals with years of experience sit at $75 to $200 and up, and senior domain experts, defined as people with at least four years in their field, clear $200 an hour (herohunt). When Time profiled the company, it put those senior roles at roughly $400,000 a year on a full-time basis, with contractors averaging 7.25 years of professional experience drawn from firms like Goldman Sachs, McKinsey, and Latham & Watkins (Time).
Read those two ends of the range again. The same platform pays one person $15 an hour and another $200 for work that carries the same job title on a resume: “AI trainer.” The difference is not effort or hours. It is what the person walks in already knowing. That is a very different labor market from the one the side-hustle blogs described two years ago, when the pitch was that anyone with a laptop and a free evening could log in and start earning.
The bottom rung is being automated by the thing it trains
Here is the part that should reframe how you think about the entry tier: the cheap version of this work is being eaten by the models it exists to build.
AI pre-labeling now handles an estimated 70 to 80 percent of straightforward cases, cutting the manual effort on basic datasets by more than half (herohunt). Deciding whether an image shows a cat or a dog no longer needs a human in the loop, and the platforms know it. In September 2025, xAI cut roughly 500 general annotators, about a third of that team, then announced plans to grow its roster of specialist AI tutors tenfold (herohunt). That is the bifurcation in a single move: fire the generalists, multiply the specialists.
Even the better general pools stay low. Surge AI, which reportedly cleared over $1 billion in labeling revenue, pays much of its contractor base in the $18 to $24 range (herohunt). If your plan for making money online in 2026 is to sign up for undifferentiated annotation work, you are competing directly with automation for a wage that automation is already winning. The couch-and-laptop version of AI training jobs is the version with an expiration date on it.
What the expert tier pays, and who it pays
The top of the market looks nothing like gig work. It looks like consulting.
CBS News reported average pay around $105 an hour on Mercor, with generalists reviewing model output near $50 and psychiatry specialists reaching up to $350 (CBS News). Primary-care physicians are paid roughly $130 to $170 an hour to evaluate datasets, and attorneys $110 to $130 to write and grade legal questions (herohunt). The people doing it are not hobbyists. CBS profiled Robin Palmer, a screenwriter who spends about 30 hours a week training chatbots on creative writing; Dr. Mike Prokop, a Sacramento anesthesiologist teaching models clinical reasoning; and Brett Brosseit, a lawyer picking up AI training roles through Handshake (CBS News).
The through-line is simple. AI labs are not buying labor. They are buying judgment that took a decade of practice to build, the kind a model cannot yet reproduce because there is no clean dataset for it anywhere. A litigator’s sense of which argument a judge will find weak, a physician’s read on when a textbook answer is dangerous in a real case: that is the raw material, and it is scarce, which is exactly why it commands consulting rates.
The part nobody markets: you are encoding your own expertise
There is a discomfort in this work that the marketplaces do not put on the landing page. When you teach a model to reason like a good lawyer or a careful doctor, you are helping build the thing that will do a slice of that job next year.
Palmer named it directly. She acknowledged the “crossing the picket line” perception and then gave the reason she does it anyway: “The train has left the station. So do you want AI to be good because it’s being trained by good people, or not?” (CBS News). Prokop, for his part, was untroubled, pointing out that hands-on medicine still needs human hands (CBS News). Both reactions can be right at once, and the split between them is the useful signal.
Some of what you would train is durable: physical, relational, high-liability work where a person still has to be accountable in the room. Some of it is not, and you are being paid a premium precisely because the supply of credentialed people willing to sit down and encode their reasoning is still thin. That thinness is temporary. The rate is real and fair today, but treating a $200 hour as a permanent income stream misreads why it exists. You are being paid well because the work is early, not because it is protected.
How I would treat this income if I took the work
Twenty-plus years in IT operations and a stack of fractional COO engagements taught me one habit that applies cleanly here: before I count money coming in, I ask what happens when the source of it changes its mind. This income fails that test in a specific way, so I would build around the failure rather than pretend it away.
Start with the economics. On these platforms the contractor keeps roughly 60 to 70 percent of what the AI lab pays, and the marketplace keeps the rest (Dealroom). That is a fine split for work you would not otherwise have. It is a bad foundation for a budget, because project volume is not yours to control. Datasets get completed, contracts end, and a lab that needed 200 psychiatrists this quarter may need 20 next quarter. This is contract revenue with a single-platform dependency, not a salary.
So I would treat it as three things and never as one. First, a rate benchmark: if a lab will pay a working attorney $120 an hour to grade legal reasoning, that is a floor to quote from in your own consulting, not a number to be grateful for. Second, a supplement layered on top of income you already have, not a replacement you quit something to chase. Third, a relationship and reputation you try to carry off the platform where the terms allow it, because the durable asset was never the login. It is the credential and the judgment behind it, and those belong to you.
I would also keep the honesty from the last section in the budget. If part of the task is teaching a model to do a thing you currently charge for, price the work knowing the window is finite, and put the proceeds toward the part of your practice a model cannot copy.
What this says to everyone else earning online
You do not need a medical license for the lesson here, because the pattern is showing up everywhere online income is made. AI is driving the price of generic, reproducible output toward zero, and paying a rising premium for verifiable expertise and accountability. The same divide already split freelancing into two markets, one that AI pay is lifting and one it is gutting, which I wrote about earlier this year in how AI is paying one half of freelancing and gutting the other. It is the same reason the real skill for solopreneurs is knowing where AI stops rather than where it starts.
If you want the practical mechanics of signing up, qualifying, and getting your first tasks, I covered the platforms and the honest pay rates in the full breakdown of getting paid to train AI. But keep the bigger read in front of it: the money in training AI concentrated at the top because that is where a machine still needs a human to vouch for the answer. Wherever that remains true, in medicine, law, engineering, or any field where being wrong is expensive, there is durable income. Wherever it stops being true, the rate falls to whatever automation charges. That is not a warning about AI training jobs specifically. It is the shape of the whole online economy in 2026, drawn in unusually sharp numbers.