What is AI training work?
AI training work is paid, remote contract work where you write prompts, rate and rank chatbot answers, and fix their mistakes, so AI labs can make their models more accurate and more useful. You work through a platform such as Mercor, Outlier, micro1 or DataAnnotation, get paid by the hour or by the task, and pick your own hours. On open listings today the typical top rate is $70/hr; general tasks pay less, and doctors, lawyers and senior engineers more.
On this page 8 sections
What do AI trainers actually do all day?
Most tasks fall into five buckets. The names change by platform (tasker, contributor, fellow, expert), but the work is the same.
| Task | What you do | Who it suits |
|---|---|---|
| Response rating | Read two or more AI answers to the same prompt, score them against a rubric, pick the better one and explain why. | Careful readers, generalists |
| Prompt writing | Write hard questions that make the model fail, then write or judge the ideal answer. | Writers, subject experts |
| Rubric building | Write the checklist a good answer must meet, so other raters and automated graders can score consistently. | Experts with teaching or review experience |
| Code evaluation | Run, debug and grade code the model wrote; write tests; fix broken solutions. | Software engineers |
| Safety and red-teaming | Try to get the model to produce harmful output, then label what went wrong. | People comfortable with sensitive content |
Language work sits across all five: the same tasks in Hindi, Spanish, Japanese and dozens of other languages. See jobs for non-English speakers.
Why do AI companies pay people for this?
A language model learns to write from huge amounts of text, but that alone doesn't teach it which answer a person would actually prefer, or when it is confidently wrong. Labs fix that with human feedback: people compare answers and the model is tuned toward the ones humans rated higher. That process is called RLHF, and it was central to how OpenAI built InstructGPT, the predecessor of ChatGPT (Ouyang et al., 2022). We explain it without jargon in RLHF explained simply.
As models get better, easy mistakes get rare. Labs now need people who can catch subtle errors in a tax question, a drug interaction or a race condition. That is why expert rates have climbed while generic tasks have gotten harder to find.
Which platforms hire AI trainers?
These are the platforms we track, sorted from easiest to hardest to get into. Pay is the platform's advertised range, not a promise. Click any name for the full grade sheet.
| Platform | Typical top rate | Getting in | Payout |
|---|---|---|---|
| No public job listPlatform says $2–25/hr | Easy to get in | Bank, monthly | |
| No public job listPlatform says $10–50/hr | Fairly easy | PayPal, weekly | |
| No public job listPlatform says $8–30/hr | Fairly easy | PayPal, on request | |
| No public job listPlatform says $3–25/hr | Fairly easy | Payment portal, weekly | |
| $90/hrPlatform says $20–200/hr | Some screening | Stripe, weekly | |
| $90/hrPlatform says $15–135/hr | Some screening | Micro1 platform wallet, twice a month | |
| No public job listPlatform says $12–90/hr | Some screening | USD bank | |
| Too few pricedPlatform says $30–125/hr | Some screening | Stripe, weekly | |
| $65/hrPlatform says $6–65/hr | Some screening | Varies | |
| $31/hrPlatform says $15–90/hr | Some screening | Payoneer, twice a month |
Not sure where you fit? Find my platform asks four questions and narrows the list. To see who is hiring today, go to open jobs, or only the newest in new this week.
How much does AI training work pay?
Advertised ranges are wide on purpose. On DataAnnotation, for example, the public site lists generalist work at up to $50/hr, language specialists up to $40/hr, and coding and data roles at up to $150+/hr (DataAnnotation). Expert marketplaces like Mercor advertise $20–200/hr depending on the role.
The catch is hours, not rates. Most people don't get 40 hours a week; they get bursts of work while a project runs and quiet weeks after it ends. Our pay report shows the hourly rates on jobs we're tracking right now, and this guide does the monthly math. To try your own rate and hours, use the pay calculator.
How do you get started?
- Pick one or two platforms that match your background. Generalists usually start with the easier platforms in the table. Engineers, clinicians and lawyers should go straight to expert marketplaces.
- Get your resume in order. AI interviewers (Mercor's, micro1's Zara) ask about what's on it. Be specific about degrees, licenses and years in the field.
- Expect an unpaid test. Nearly every platform has an assessment or AI interview. Do it in one sitting, somewhere quiet, without AI tools.
- Read the project guidelines twice. Most people who get removed early broke a guideline they skimmed.
- Apply to several platforms. Work comes and goes per project. Having two or three accounts smooths out the dry spells.
What are the downsides?
- No guaranteed work. Being accepted is not the same as having tasks. "Empty queue" posts are the most common complaint on every platform subreddit.
- Unpaid screening. Assessments and onboarding courses are usually unpaid and can take hours.
- Opaque removals. Accounts or project access can be removed with little explanation, often after a quality review.
- You're a contractor. No tax withheld, no benefits. See how payment and taxes work.
Treat it as flexible side income that can occasionally become full-time, not as a stable job.
New AI training jobs
Everything posted this week
up to $100/hr

$250–280/hr

$40–90/hr

$139–140/hr

$70–120/hr
Apply opens the platform's own page. Rubric never charges you and never sits between you and the platform.
Other questions people ask
Is AI training work a real job?
Yes. AI labs pay contractors through platforms like Mercor, Outlier, micro1 and DataAnnotation to write prompts, rate answers and correct model output. It is contract work, not employment: no benefits, no guaranteed hours, and projects can end without notice.
Do I need a degree to do AI training work?
Not always. Generalist writing and rating projects often ask for strong English and careful reading rather than a specific degree. The better-paid domain projects (coding, medicine, law, math) usually want a degree or professional experience in that field.
Is AI training work the same as data entry?
No. Old-style annotation was drawing boxes on images. Most current work is judging reasoning: is this answer correct, safe, well-written, and better than the other one? It needs attention and subject knowledge, which is why it pays more than data entry.
Can I do AI training work part-time?
Yes, almost all of it is part-time and remote. Some projects set a minimum of 10 to 20 hours a week; others let you log in whenever tasks are available. See realistic hours and earnings.
Do I have to pay to start?
No. Real platforms never charge an application fee, a training fee or a deposit. Anyone asking you to pay is running a scam. See how to spot fake AI training jobs.
Sources
Put this guide to work
- Top picks
This week's best-scoring jobs, and why.
- Resume match
Paste your resume, see the open jobs it fits.
- Pay calculator
Your rate and real hours as monthly pay.