AI data labeling demand jumped 154%. Read what it pays.

Upwork says data-labeling demand rose 154% in a year. The nominal wage sits near $25 an hour, and after unpaid training many labelers clear $4 to $7.

The Editors · 6 min read · Jul 7, 2026

People sitting on chair in front of computer

Data labeling is the fastest-growing way to earn money with AI, and for most people doing it, the pay gets worse. Demand for AI data annotation and labeling rose 154% year over year on Upwork, one of the sharpest jumps in the platform's In-Demand Skills 2026 report. The wage went the other way. A generalist labeler in the US clears about $25 an hour on paper, and once you count the unpaid training, screening, and idle time, workers' own accounts put the real rate closer to $4 to $7. The money that held its value moved to a small tier of credentialed specialists. If you are weighing data labeling as a side income, the honest read is that the demand headline and the wage describe two different jobs.

What the 154% actually measures

Upwork's number is real, and it does not mean what the listicles imply. The report, published February 4, 2026, tracks freelancer earnings across its marketplace from 2024 to 2025. Data annotation and labeling grew 154%, sitting alongside AI video generation at 329% and AI integration at 178%. That figure measures total money flowing into the category, not the rate any one worker gets.

A category can double in total spend while the per-hour rate falls. It happens when the number of workers grows faster than the budget. Entry-level labeling needs no degree and no portfolio, so when the demand headline hits, the supply of people willing to do it climbs faster than the pay. The growth number is what a crowded auction looks like from the top.

What the work is

Data labeling is the human work behind model training. You tag images, rank two chatbot answers against each other, flag unsafe text, transcribe audio, or check whether a model's output is correct. The main platforms are Outlier (owned by Scale AI), DataAnnotation, Remotasks, Mercor, and Surge AI. The pitch is remote, flexible, and open to anyone.

That open door is the whole problem. When a task needs no credential, the only thing setting your pay is how many other people will take the same task for less. For the simplest labeling, that number is very large.

The wage after the unpaid parts

The nominal rate hides the structure of the job. A 2022 study of platform workers, cited by AlgorithmWatch, found that about 34% of all time spent on the platform was unpaid: 7.8 unpaid hours against 14.8 paid ones in an average week. The unpaid time is onboarding, qualification tests you can fail, project training, and meetings.

Count that time and the hourly wage collapses. Workers interviewed by AlgorithmWatch estimated their real earnings, including everything unpaid, at roughly $4 to $7 an hour. At the bottom end, a Canadian Affairs investigation found Remotasks paying around one US cent for tasks that could take several minutes. The same report noted Anthropic requires its labeling contractors to be paid $16 an hour or higher. A mandated floor is a tell: you set one when the market rate drifts below it on its own.

This is the gap the median AI side hustle already shows. The advertised number is a ceiling a few people touch. The number most people live on sits well underneath it.

The lawsuits are part of the data

When a wage complaint becomes a legal filing, that says something about the category, not one bad week. In December 2024, a former Scale AI contractor sued the company, alleging he was promised $25 an hour and paid only a portion of it. In March 2025, the US Department of Labor opened an investigation into Scale AI over alleged wage violations and worker misclassification.

Misclassification is the mechanism worth understanding. Treat a worker as an independent contractor and the unpaid onboarding, the failed qualification tasks, and the idle time become the worker's cost instead of the platform's. That is legal to argue over, and it is exactly what the filings are arguing over.

Where the pay is real

There is a tier of this work that pays well, and it is the opposite of the entry-level pitch. Reinforcement learning from human feedback, the work of ranking and correcting model outputs, pays $40 to $55 an hour on Mercor and $30 to $42 on Outlier, with coding and STEM specialists reaching $50 to $65 and expert contractors averaging around $95.

The barrier there is a credential. These platforms want a computer science degree, a nursing license, a law background, a graduate physics student, or a fluent second language. The AI part is not the moat. The domain expertise is. This is the same split that runs through skilled AI freelance rates, where the people charging $80 to $150 an hour are selling a skill that predates the AI label.

The honest verdict

If you already hold a real credential, specialist AI training is a legitimate, well-paid contract lane, and the demand is genuine. Sign up, and treat the rate cards above as the number to hold out for.

If you are looking for easy AI money with no barrier to entry, the fastest-growing corner of it is a race to the bottom. The 154% is real, the $25 headline is real, and neither one is the wage you will clear after the unpaid hours. The tell is the entry requirement: any income that anyone can start today is priced for a crowd, because the crowd is already there.

The thing to watch is whether platforms get forced to pay for onboarding and qualification time. That single change would reset the effective wage more than any demand figure, and the Department of Labor investigation is the first place it might come from.