The complete guide to AI training work
What AI training work is, who pays for it, how to get in and stay active, and the downsides worth knowing before you apply.
On this page7 sections
The AI assistants people use every day were shaped by human judgment: people ranked their answers, wrote better ones and tried to make them fail. That work is paid and remote, and some of it is open to applicants well beyond the US and Europe. This guide covers what the work is, who pays for it, how to get in and stay in, and the trade-offs the sign-up pages tend to skip.
What the work is, and who pays for it
Most projects are built from five kinds of task, and many mix several.
| Kind of task | What you do | What it is for |
|---|---|---|
| Preference ranking | Compare two or more responses to the same prompt, choose the better one and explain why | Teaching a model what people judge to be a good answer; the core of RLHF |
| Demonstrations | Write the ideal response to a prompt yourself | Supervised fine-tuning, which shows the model the target directly |
| Evaluations | Grade outputs against a rubric, or write test questions with known answers | Measuring where a model fails, before and after training |
| Red-teaming | Try to draw harmful, false or unsafe output from the model | Finding weaknesses before users do |
| RL environments | Build tasks a computer can score: problems with verifiable answers, code with tests | Giving a model something to practice on, at scale |
The first two are the foundation. In OpenAI's 2022 InstructGPT paper, contractors wrote example answers used to fine-tune the model, then ranked its outputs. Those rankings trained a reward model, which steered further training through reinforcement learning. Human raters preferred the outputs of the resulting 1.3-billion-parameter model to those of the 175-billion-parameter GPT-3. On the prompts OpenAI tested, careful human feedback outweighed a hundredfold difference in size.
Labs automate what they can. Anthropic's 2022 Constitutional AI paper trained a model to be harmless without human labels identifying harmful outputs, relying instead on AI feedback guided by a written list of principles. DeepSeek's 2025 R1 paper describes reinforcement learning with rewards that check whether a math answer is correct and whether code passes predefined test cases.
Neither approach removes people. Someone still writes the principles, designs problems hard enough to matter, writes the tests and checks that an automated grader rewards the right thing. That calls for expertise. A model that already writes fluent prose gains little from another fluent paragraph, and a great deal from a physician who can spot the dangerous error inside one. Safety work draws on specialists too: in 2023, OpenAI opened applications for a Red Teaming Network of outside experts to probe its models.
Who actually pays you
Labs often buy training data from specialist vendors rather than hire contributors directly. Vendors recruit, test and manage the people who produce it. Even InstructGPT's labelers were about 40 contractors hired through Upwork and Scale AI. Scale AI now runs Outlier as its contributor platform.
The brand you apply to is usually the vendor, not the lab. Projects can run under code names, and the client is not always disclosed.
The roles, and what moves the pay
Most platforms sort contributors into tracks. Your track decides which assessments you take, which projects you see and, roughly, what you earn.
- Generalist. Writing, ranking and fact-checking in careful English. The widest door, and the most crowded.
- Coding. Writing and reviewing code, grading model solutions and building tasks with tests. More in coding for AI labs.
- STEM. Hard problems in mathematics, physics, chemistry and biology, with answers that can be checked.
- Domain experts. Law, medicine, finance and other credentialed fields, where your judgment is the product.
- Languages. Native fluency in a language other than English. Demand varies widely by language and over time.
- Creative. Fiction, poetry, screenwriting and other work where taste is the skill.
Pay is set by project, not by platform, and four things move it most.
Expertise. Scarce, verifiable expertise earns the highest rates. Generalist work usually pays least, because the most people can do it.
The project. Rates are set project by project, so one platform can pay very different rates for different work.
Country. Some platforms, Outlier among them, factor location into rates, so similar work can pay differently from one country to another.
Quality. More work, and often better-paid work, tends to go to contributors whose quality stays consistently high.
Current pay ranges, payout methods and country notes live on each platform profile, starting with Outlier and DataAnnotation. Treat any figure in a forum post as a snapshot of one project, in one country, at one moment.
How to start
Start with two or three platforms, not ten. Each asks for unpaid time up front, and a few careful applications are worth more than many rushed ones. Vendors bring people in two ways.
| Task marketplaces | Contract matching | |
|---|---|---|
| Examples | Outlier, DataAnnotation | Mercor, micro1 |
| Getting in | Assessments, then qualification project by project | An application and an AI-led interview, then matching to a role |
| How work arrives | Projects appear on a dashboard; you choose when to work | A contract for one project, often with a weekly hours cap or target |
| How pay is set | By project, usually hourly | A rate set in the listing or offer, usually hourly |
| Main risk | Work can dry up without warning | A long wait for a match, or none at all |
There is a good case for one of each: a marketplace for flexible hours, and a matching platform for contracts that make full use of a professional background. If an AI interview is part of your plan, read how the micro1 and Mercor interviews work.
- Lead with your strongest credential. A nurse, a tax lawyer or a physics graduate should apply on that basis, not as a generalist. Our guide to where to apply first matches profiles to platforms.
- Prepare as you would for a job. Make your résumé, LinkedIn profile and application tell the same story, and be exact about your qualifications.
- Treat every assessment as your only attempt. Some are timed, some cannot be retaken (DataAnnotation's Starter Assessment is one), and they are graded against the instructions you are given, not your personal taste. See how to pass AI training assessments.
- Verify your identity cleanly. Expect to verify with a government ID. Use your legal name exactly as it appears there, from the country where you actually live.
- Set up payouts before you work. Confirm that the platform's payout method works where you live and that the account is in your verified name. Our guide to getting paid from Africa covers the options.
Tip
Keep a short log of every application: platform, date, assessment, outcome. When a platform goes quiet, the log tells you where to follow up and where to stop waiting.
When you are ready, these are sensible first applications if they accept applicants where you live. DataAnnotation, the other marketplace above, reportedly takes applications only from the US, Canada, the UK, Ireland, Australia and New Zealand. Where we have a referral link, Apply buttons on ASV use it; the disclosure page explains how that works.
Scale AI
Scale AI's freelance network for training and evaluating AI models.
$10–$50/hrAssessedSome countries
micro1
An expert network supplying human data to frontier AI labs.
$20–$200/hrSelectiveOpen in Africa
Mercor
Specialist AI-training contracts for frontier labs, paid weekly via Stripe.
$20–$200/hrHighly selectiveSome countries
Staying active and growing
Getting in is the first filter. Staying in is the second, and it runs on quality.
Your submissions are reviewed, often by other contributors, and scored against the project's guidelines. Those scores shape which projects you are offered and whether you keep access to the one you are on. On some platforms, consistently strong work also leads to reviewer roles, grading other people's tasks.
A few habits carry most of the weight.
- Re-read the guidelines whenever they change. They can change mid-project, and an answer that was right last week can be marked wrong today.
- Learn to write a rubric. Rubrics sit at the center of evaluation work, and writing one that two graders apply the same way is a skill in its own right. See rubrics that hold up.
- Expect droughts. Projects fill, pause and end. A dashboard can go empty overnight and stay that way for weeks.
- Diversify. Keep two or three platforms active, take new qualification tests as they appear, and keep income that does not depend on any of them. ASV also runs its own paid projects, from task authoring with rubrics to software tasks.
The honest downsides
The work is real, and for the right profile it can pay well. It is also contract work in a young and volatile market.
- No guaranteed hours. Platforms offer access to work, not a volume of it. Income can swing sharply from one month to the next.
- Abrupt endings. Projects end when a client's needs change, often with little notice. Accounts are deactivated too, sometimes for clear violations and sometimes with little explanation.
- Thin support. Help usually means a ticket queue and a help center, and appeals can be slow.
- Unpaid time. Applications, entry assessments and some onboarding are usually unpaid, and so is time spent waiting for work.
- Your own taxes. You are an independent contractor: no benefits, no pension contributions and, usually, no tax withheld for you. Keep records and declare the income where you live.
Important
Treat AI training income as variable until it has been steady for several months. Do not leave a job or take on fixed costs on the strength of one good project.
If you are applying from Africa
Everything above applies to you. A few points deserve extra weight if you live in Nigeria, Kenya, Ghana, South Africa, Egypt or elsewhere on the continent.
Availability is set by platform and by project. A platform can accept applicants from your country and still restrict individual projects to others, and its country list can change. As of October 2026, Outlier's country list included Egypt and South Africa but reportedly not Nigeria, Kenya or Ghana; Mercor listed Kenya as unsupported; and micro1's exclusion list named none of these five countries. Check the notes on each platform profile before you apply, and do not build your plans around any single platform's presence in your country.
Payouts need planning. Before you put in hours, find out what it costs to move money to your bank and how long it takes.
Location and identity must be real. This rule has no exceptions.
Warning
Never use a VPN to appear to be somewhere else, and never work on an account verified with someone else's identity, whatever you are offered. Outlier's community guidelines, for one, bar misrepresenting your identity or location and list disabled accounts among the consequences.
Be wary of anyone who offers to sell or rent you an account, or asks for a fee to get you onto a platform. Our guide to AI training job scams covers the common patterns.
The work rewards the same things everywhere: real expertise, careful writing and close reading of guidelines. For where demand is heading, see the state of AI training in 2026.
Sources
- Training language models to follow instructions with human feedback (OpenAI, via arXiv)
- Constitutional AI: Harmlessness from AI Feedback (Anthropic, via arXiv)
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (DeepSeek, via arXiv)
- OpenAI Red Teaming Network (OpenAI)
- Your guide to working on Outlier (Outlier)
- Community Guidelines (Outlier)
- Supported Countries for Payment (Mercor)
- Leverage your expertise to train next-generation AI models (micro1)
In this post
- OutlierApply(opens in a new tab)
$10–$50/hr
- DataAnnotationApply(opens in a new tab)
$20–$100/hr
- micro1Apply(opens in a new tab)
$20–$200/hr
- MercorApply(opens in a new tab)
$20–$200/hr