Get Paid to Train AI: What Data-Annotation Gigs Pay in 2026

Priya signed up for DataAnnotation in April 2026, hoping to pull an extra $500 a month out of her kitchen table by training AI in the evenings. Her first week she cleared $310. Then the tasks dried up for nine days straight, and she started wondering whether the whole thing was a scam.

It wasn't. She'd just run into the two things nobody tells you about getting paid to train AI: the money can be real, and the flow of work is nothing like a paycheck. If you've seen the "$40 an hour to teach AI, work whenever you want" ads and felt the pull, this is the honest version of what you're signing up for.

What "training AI" really means

The models behind ChatGPT, Gemini, and Claude don't learn good judgment on their own. Companies pay people to supply it. That's the whole job. You might read two AI-written answers to the same question and rank which one is more helpful. You might write a tricky prompt designed to trip the model up, then grade how it responds. If you have a specialty, you might check whether the AI's Python code compiles, whether its Spanish translation reads naturally, or whether its explanation of a chemistry reaction is correct.

The industry calls this data annotation, or reinforcement learning from human feedback. To you, it feels like a mix of grading papers, fact-checking, and writing. Some tasks take two minutes. Some take an hour. The common thread is that a human has to decide what "good" looks like, and the AI companies can't build their products without a lot of humans doing exactly that.

As TIME reported in its look at the field, this has quietly become real remote work for tens of thousands of people, sitting in an odd space between a tech job and a gig-economy hustle.

What it really pays

The ads get slippery here, so it helps to use the numbers workers report rather than the headline promises.

On DataAnnotation, generalist tasks tend to run $20 to $25 an hour, and specialized work like reviewing code pays roughly $40 to $45 an hour, paid out weekly through PayPal. Outlier, run by Scale AI, sits in a similar range: most standard tasks land around $15 to $30 an hour, while niche projects in areas like software, law, or chemistry can reach $35 to $60.

You'll see wilder figures floating around. Glassdoor, for instance, lists an estimated average of about $52 an hour for an Outlier annotator, which sounds fantastic until you remember those numbers are self-reported by the people who landed the best-paying projects. They're the ceiling, and most people work well below it.

The rate on the task is only half the story. Three things quietly eat into your real hourly pay:

You don't get paid for the qualification tests. Before you can touch a project, you sit assessments, and DataAnnotation is explicit that annotators aren't compensated for training assessments. The starter test can't be retaken, and the company says it accepts only about 2.6% of applicants. So your first few hours might earn nothing.

You don't get paid to hunt for tasks. Work appears on a dashboard and you claim it, which means slow days look like Priya's nine-day gap. When a big project ends, your income can drop to near zero until the next one opens.

You don't get benefits or withholding. Every one of these platforms treats you as an independent contractor, which brings us to the part that surprises people every April.

The tax bite is bigger than you think

When you earn money as a contractor, no one takes taxes out for you. That $25-an-hour task is not the same as $25 an hour from an employer, because you now owe both halves of Social Security and Medicare, the self-employment tax, which runs 15.3% on top of regular income tax.

If you earn $600 or more from a platform in a year, you'll get a 1099-NEC form. Even if you somehow don't receive one, the income is still taxable and the IRS still expects it. A reasonable rule of thumb: set aside 25% to 30% of every payout in a separate account and don't touch it. If you end up owing more than about $1,000 for the year, you may also need to make quarterly estimated tax payments instead of settling up once in April.

None of this makes the work a bad deal. It just means the take-home is lower than the sticker rate, and the person who mentally spends the full $310 is the person who gets an unpleasant surprise at tax time.

The scam version is everywhere now

For every real platform, there are a dozen fakes riding on the hype, and the numbers on this are alarming. The Federal Trade Commission tracks "task scams" as their own category, and reports of them exploded from essentially zero in 2020 to about 5,000 in 2023, then roughly 20,000 in just the first half of 2024. The FTC estimated task scams made up nearly 40% of all job-scam reports that year, and job scams overall cost people more than $220 million in the first six months of 2024 alone.

The tell is simple, and it's worth tattooing on your brain: a real AI-training platform will never ask you to pay to work. The FTC puts it plainly, that anyone telling you to send money in order to get the money you supposedly earned is running a scam. The classic version starts with an out-of-the-blue text or WhatsApp message about "online tasks," pays you a small amount to build trust, then asks you to deposit your own money (often in cryptocurrency) to "unlock" bigger earnings. That deposit is gone the second you send it.

There's a quieter version too. Some fake "assessments" exist only to harvest your personal details, your full name, birth date, address, and Social Security number, which then feed identity theft. If a first-day application asks for your SSN before you've done a minute of work, close the tab.

Identity checks confuse people, so it's worth clearing up: a legitimate platform asking you to verify with a photo ID and a selfie is normal and expected. That's them confirming you're a real human rather than a bot farming tasks. A scammer asking for your bank login or Social Security number on day one is something else entirely. The direction of trust is the giveaway: real companies verify that you exist, while scammers try to pry loose details they can sell.

How to try it without wasting your time

If you want to test the water, go straight to the source. Type the platform name into your browser yourself rather than clicking any link someone sent you. DataAnnotation and Outlier are the two most commonly cited legitimate players, and both are free to join. Expect to pass an unpaid assessment first, and expect it to be hard, because a low acceptance rate is a feature, not a red flag.

Treat the first month as an experiment, not a second job. The pay is real but the volume is unpredictable, so it suits someone who wants flexible income to fill gaps rather than someone who needs a reliable amount every Friday. If you have a marketable specialty like coding, translation, or a science background, you'll qualify for the higher-paying projects and your effective rate will climb well above the generalist floor.

Track your actual hours, including the unpaid assessment time and the task-hunting, then divide your total earnings by that number. That real hourly figure, rather than the advertised one, tells you whether this is worth your evening.

The Bottom Line

Getting paid to train AI is a real, if lumpy, way to earn extra money in 2026, and it's crawling with imitators. This week:

1. Pick one platform and go to it directly. Search "DataAnnotation" or "Outlier" yourself and apply on the company site. Never start from a text or DM.
2. Bank the tax money now. Open a separate savings account and move 25% to 30% of every payout into it the day it lands, so April doesn't bite.
3. Run one scam check before you share anything. If a "job" asks you to pay, deposit crypto, or hand over your SSN before you've done real work, walk away and report it at reportfraud.ftc.gov.
4. Judge it on your real hourly rate. After your first two weeks, divide total earnings by total hours, including unpaid tests. Let that real number decide whether you keep going, rather than the ad.