Get paid from your couch to make AI smarter. Real pay by platform, who gets accepted, the scams to dodge, and what you keep after tax.
This in-depth guide covers everything you need to know about ai training jobs: get paid to train ai (2026 pay rates). Based on verified income data and real-world case studies from our database of 138 side hustle tactics.
AI training jobs pay you to rate chatbot answers, write prompts, check code and catch a model when it gets something wrong. Published rates run from about $8 an hour on research studies to $100 or more for credentialed experts, and entry level work sits around $12 to $25 an hour. The work is real and it pays, but hours are never promised, assessments are often unpaid, and accounts can vanish overnight, so treat it as a flexible side income until it proves otherwise.
Why this work is suddenly everywhere
You have probably seen the ads. "Get paid to train AI." "Work from home, set your own hours." Maybe a friend says they made rent last month by grading chatbot answers in their pajamas.
It sounds too good. Is it real, or another scam dressed up in AI words? Some of it is a scam. And some of it is exactly what it looks like.
Every big AI model is shaped by people. Someone reads two answers and picks the better one. Someone writes a hard maths question the model fails. Someone checks whether the code runs. That human judgement is what the industry calls RLHF (reinforcement learning from human feedback), and the companies building models pay platforms like Outlier, DataAnnotation and Mercor to supply it.
So yes, there is real money here. This guide shows you where it is, as of October 2026, with the pay each platform publishes, who gets accepted, how you actually get paid, and the problems workers report that the ads leave out.
If you already hold a degree, licence or deep professional skill, read this first, then go straight to our expert guide on AI evaluation, red-teaming and expert data work. That is where the higher rates live.
What the work actually looks like
"AI training" covers five kinds of task. Most platforms mix them.
1. Rating responses. You see a prompt and two or more AI answers. You score them for accuracy, helpfulness and safety, and explain your choice in a few sentences. This is the most common entry task, often listed as "AI rater" or "AI evaluator".
2. Writing prompts. You write questions designed to make the model fail, then write the ideal answer. Harder than it sounds. A prompt the model gets right usually earns nothing.
3. RLHF and ranking. You compare answers, rewrite the weak one, and leave feedback the model learns from. Most "RLHF jobs" are this.
4. Coding and STEM tasks. You check AI written code, debug it, or write graduate level maths, physics or chemistry problems with worked solutions. These pay more because fewer people can do them.
5. Red-teaming. You try to get the model to say something harmful, biased or false, then document how. This is careful, detailed work, and it sits mostly in the expert tier.
Plus a sixth, older kind: classic data annotation, labelling images, drawing boxes around cars, transcribing audio. It still exists, it pays less, and it is shrinking as models learn to label for themselves.
Which of these sounds like you? A careful reader, a sharp writer, a coder, a specialist? Your answer decides which platform to try first.
Rates below are what each platform or its job listings publish, checked in October 2026. Published pay is a ceiling for most people. Your rate depends on your country, your track and the project.
| Platform | Published pay | Who gets in | How you get paid |
|---|
| Outlier (Scale AI) | $10 to $50 an hour, by skill, task, location and project | Typically undergraduate level expertise or above; valid ID and a phone number from your country | Weekly via PayPal, AirTM or ACH |
| DataAnnotation | $20+ multilingual, $25 to $50+ general, $40 to $150+ coding, $40 to $125+ STEM and professional | Fluent English, reliable internet; higher tracks need degrees or licences | PayPal, delivered within a few days of requesting |
| Mercor | Listings we saw ran $61 to $77 and $100 to $150 an hour; the site claims a $109 average contracted rate | Experts in medicine, law, engineering, data science and similar; set per listing | Weekly for hourly work via Stripe or Wise, first payout held 7 days |
| Alignerr (Labelbox) | Listings from $20 to $50 up to $40 to $120 an hour; the site claims an $80 average | PhDs, researchers, professionals, native speakers for language work | Weekly, by the following Friday, typically via Stripe or PayPal |
| Handshake AI Fellowship | $30 to $125 an hour, by education and project | Mostly master's, PhD or postdoc level; needs an SSN or ITIN and US work authorization | Paid directly by Handshake as a 1099 contractor |
| Remotasks (Scale AI) | No hourly rate published; pay per task | Says 90+ countries, but cut several countries in 2024 | Weekly via PayPal or AirTM |
| Prolific (studies) | Minimum £6 / $8 an hour; researchers are advised to pay £9 / $12 | 18+, most OECD countries, waitlist | PayPal, cash only |
| Appen / CrowdGen | "$100+ an hour" top rate, depending on location and skills; most projects far lower | Global; per project | Local currency via bank transfer, PayPal, Payoneer, Airtm |
| TELUS Digital AI Community | Set per project and country; one contract listing showed $15 to $25 an hour | Global, by country and language; qualification exam | Payment terms set per project |
| Invisible | No public rate card | Per listing, often by location | Per contract; we could not confirm the method |
| Turing | No published contributor rates | Developers, data scientists, domain experts | Independent contractor; method not stated publicly |
Two things jump out. The expert platforms (Mercor, Alignerr, Handshake) publish big numbers because they want doctors, lawyers and PhDs. And the platforms most open to everyone publish the least about pay.
Where does that leave you if you are a smart generalist with a bachelor's degree? Probably with Outlier and DataAnnotation as your first two tries, Prolific running in the background, and the expert tier as the goal.
Our full reviews go deeper: Outlier AI review, DataAnnotation review, Mercor review and Alignerr review.
A few notes the table cannot hold
Remotasks. The site is still live and says it runs in 90+ countries. But in March 2024, Rest of World reported that it shut down entirely in Kenya, Nigeria and Pakistan and stopped new sign ups in Thailand, Vietnam and Poland. Scale AI also runs Outlier. Check that your country is accepted before you spend time on onboarding.
Appen. Google ended its contract with Appen in March 2024, as Search Engine Land reported. That contract was a big source of steady rater work. Appen still runs projects through CrowdGen, but expect gaps.
Prolific. This one is research studies and surveys, some of them AI evaluation tasks. Pay is a fixed reward per study, and the hourly minimum is worked out from the study's estimated completion time. It is the easiest place to get a first payout, and the hardest to turn into real hours. Our paid user testing guide covers this world in full.
Handshake. The Fellowship is mostly for graduate level experts. Handshake says fellows typically work 5 to 20 hours a week, on projects that last 2 weeks to 3+ months.
Who actually gets accepted
The honest answer: fewer people than apply, and the bar keeps rising.
As models improve, the easy mistakes disappear, so labs want people who can catch the subtle ones. That means:
- Strong English writing. Nearly every platform requires it, even for maths and code. You will explain your reasoning in writing all day.
- A degree or a provable skill. Outlier says opportunities typically need undergraduate level expertise. Handshake says most roles need a master's, PhD or postdoc. DataAnnotation's STEM track asks for advanced degrees or equivalent experience.
- A subject the labs are short on. Coding, maths, physics, chemistry, medicine, law, finance and non-English languages are the busy lanes.
- The right country. Many projects are limited by location, and pay is often adjusted by country. Handshake requires an SSN or ITIN and valid US work authorization. Prolific covers most OECD countries.
- Patience for strict reviewers. Your work gets checked. A few low scores can take you off a project.
Do you have a subject you know better than most people around you? That is the thing to lead with on your application.
How to pass the assessments
Almost every platform starts with a test. DataAnnotation's Starter Assessment takes about an hour. Its FAQ says you can take it only once, with no retakes, and does not say it is paid. Independent reviews describe the core qualification as unpaid.
Can you explain why an answer is wrong in two clear sentences? That is most of the test. Here is what helps:
- Read the instructions twice. Missing a guideline sinks more tests than missing a fact. If the task says "rate for helpfulness first", do that.
- Write your reasoning like you are teaching. Specific beats long. "Answer B dates the Peace of Westphalia to 1658 when it was signed in 1648, and skips the second question" scores better than "A is better overall".
- Fact check everything. Assume the AI answer contains an error until you have checked it. Usually it does.
- Do not use AI to answer the test. Platforms look for it. Handshake's current agreement, Business Insider reported in March 2026, bans LLM use unless authorised in writing.
- Take your time. Rushing is the most common reason people fail. But do not take forever either, because some platforms flag very slow task times.
- Take every qualification offered. On DataAnnotation, workers describe qualifications as the gate to steady work.
- Use a quiet hour. Treat the test like a job interview, because it is one.
If you fail, you probably missed that platform's style. Try another one.
The downsides workers report
This is the part the ads skip. Read it before you give any platform your evenings.
Empty queues and droughts. You pass, you get a few great weeks, then nothing. On Trustpilot in October 2026, recent reviews of DataAnnotation describe long stretches with no tasks after onboarding, alongside unanswered support emails. A great hourly rate times zero hours is zero.
Sudden removal. Recent Outlier reviews on Trustpilot describe workers removed from projects without a clear reason, including right after passing a project's onboarding exam. When a former worker sued Scale AI in January 2025, the company told TechCrunch it had removed her from Outlier for violating unspecified community guidelines.
Withheld pay. In March 2026, Business Insider reported that Handshake suspended several contractors on OpenAI projects between late December and January and withheld their pay. One support email told a worker: "This decision is final. There is no appeal process, and any work associated with this violation is not eligible for payment." Handshake declined to comment.
Pay cuts between projects. In November 2025, PEOPLE reported, citing Forbes and Business Insider, that Mercor ended a project involving thousands of contractors and offered some of them a new project at $16 an hour, down from $21. Mercor said its materials state the work is temporary and project based.
Unpaid time. Assessments, onboarding and reading long guidelines are often unpaid. A former Outlier worker alleged in her January 2025 lawsuit that she worked about 10 hours a day and was paid for about five, partly because instruction reading and training were unpaid. That is an allegation, and Scale said it paid her fairly.
Contractor status. You are almost always an independent contractor. No sick pay, no overtime, no employer paying half your Social Security. More on that below.
How would your month look if your main platform went quiet for three weeks? If that question makes your stomach drop, keep this as a side income for now.
How to avoid AI training job scams
The real platforms have a fake twin for every ad. The scammers know you want this to be real.
Never pay to get paid. The FTC is blunt: honest employers will never ask you to pay to get a job. CrowdGen says it will never ask you to pay anything to join its projects. Any "registration fee", "training fee" or "ID verification fee" is a scam.
Watch for task scams. The FTC says reports of these jumped to about 20,000 in the first half of 2024, up from about 5,000 in all of 2023. It starts with a text or WhatsApp message offering online work. You rate products or "boost" apps, see fake earnings pile up, then get told to deposit money (often crypto) to release them. That money is gone.
Check the recruiter. Real platforms do recruit through LinkedIn and job boards, which is why fake recruiters copy them. Apply only through the platform's own website. TELUS has warned in its own community knowledge base about people posing as its recruiters and asking for a fee for a "Premium ID", and says it never requests deposits for any role.
Red flags in one list:
- A message out of the blue offering easy, high pay
- Any request for money, gift cards or crypto
- A Gmail or lookalike email address
- An interview held only by text chat
- A cheque sent to you to "buy equipment"
- Pressure to start today
Would a real company text you first, then ask you for crypto? Never. If someone has already taken money from you, report it at ReportFraud.ftc.gov.
The money math: what you really take home
What is your real hourly rate once the empty weeks are counted? The headline rate is what you earn in an hour you actually work. Your real income is that rate times the hours the platform gives you, minus tax.
Our own research for the AI evaluation guide puts trainer pay at about $31 an hour on average, across three very different tiers: about $12 to $25 for entry work, $25 to $53 for the middle, and $75 to $200+ for credentialed experts.
Now a worked example. These are illustrative numbers, built to show the shape of the problem.
Say you earn $20 an hour and aim for 10 hours a week. Over 52 weeks that would be $10,400. Now assume one week in three is empty, a guess based on the droughts workers describe. You work about 35 weeks, so $7,000.
Then tax. In the US, self-employment tax is 15.3% for Social Security and Medicare, and you owe it once your net self-employment earnings reach $400. It applies to 92.35% of your net earnings, so on $7,000 that is roughly $990 before any income tax.
So the $10,400 job is closer to $6,000 after self-employment tax, and less once income tax is counted. Still real money, and smaller than the ad.
Want your own number? The AI training pay calculator works out your real take-home after taxes and empty weeks. Put in your rate, your hours, and how many weeks you expect to sit idle.
Taxes: the 1099 part nobody explains
As a contractor, nobody withholds tax from your pay. You are the business. Have you set anything aside from your last payout yet?
- The 1099 form. For payments made in 2026, a US company sends a Form 1099-NEC once it pays you $2,000 or more in the year, up from $600 before. Below that, you still owe tax on every dollar.
- Quarterly payments. If you will owe a meaningful amount, the IRS expects estimated tax payments during the year.
- Deductions. A share of your internet, a computer used for the work, and other business costs may be deductible. Keep records.
- Outside the US. Rules differ, and payouts in US dollars can still be taxable where you live.
Our side hustle tax guide covers thresholds, deductions and deadlines for the US, UK, Canada and Australia, and the side hustle tax calculator shows what to set aside from each payout.
Quick answers
Is AI training work legit? The big platforms in the table are real companies that pay real people, and Outlier says it has paid out more than $500 million. The risks are inconsistent hours, removals and pay disputes, plus the scams that copy their names.
How much can you make? It depends on your tier and on how many hours the platform gives you. Entry work sits around $12 to $25 an hour in our research, and expert work can run past $100. Monthly income swings with the queue.
Do you need a degree? For the better paid tracks, usually yes, or provable professional experience. Prolific studies and some rating projects ask for less.
Can you do it from outside the US? Often, yes. Outlier, DataAnnotation, CrowdGen and Prolific work across many countries, but pay is often adjusted by location and some projects are limited to certain countries. Handshake's Fellowship requires an SSN or ITIN and US work authorization.
Is it a full time job? For most people, no. Treat it as flexible side income until you have six months of steady records.
What it feels like when it works
Picture a Tuesday night. You are on the couch, laptop on a cushion, tea going cold. You read a chatbot's answer to a chemistry question and spot that it mixed up two compounds. You write three clear sentences explaining why. That is one task. You do twenty more before bed.
On Friday, a notification: your first PayPal payout has landed. You earned it by knowing something, from your own living room.
That first payout matters more than the amount. It proves your knowledge has a price. What would it mean to you to see it land?
Your first steps
Do this over the next seven days.
- Pick your lane. Generalist writer, coder, STEM, professional, or language. Write it down.
- Apply to two platforms that match. Generalists: Outlier and DataAnnotation. Experts: Mercor and Alignerr, plus Handshake if you are a graduate with US work authorization. Add Prolific for small, quick studies.
- Apply only on the official sites. Never through a link in a DM.
- Set up PayPal or your payout account in your own name before you start.
- Block a quiet hour for each assessment. Read the rules first.
- Open a simple spreadsheet. Track hours, pay, and empty weeks from day one. It makes tax time easy and tells you if a platform is worth keeping.
- Run the pay calculator after your first month with your real numbers.
If you want more ways to earn from home alongside this, see our guide to making money from home, the honest side hustle reality check, and how much side hustles actually pay. If you have a professional licence, the expert tier is where this gets serious: AI evaluation, red-teaming and expert data work.
The honest summary
AI training jobs are real. Thousands of people get paid every week to make AI models smarter, and your first assessment could happen this week.
The catch is the shape of the income. High hourly rates, no promised hours, strict reviewers, and a contractor tax bill. The people who do well treat it like a business: two or three platforms at once, a record of every hour, money set aside for tax, and a skill that keeps them in the higher tier.
You already know things the models get wrong. Which subject would you bet they fail at first?