To become an AI training expert, you apply to an AI data platform with evidence of real expertise, pass a skills assessment and an identity check, and then complete paid tasks: writing ideal answers, comparing and grading model responses, red-teaming, or demonstrating tasks inside software environments. The best-paid work goes to people with deep, verifiable experience in a field — medicine, law, finance, engineering, customer support — rather than to generalists.
What does an AI training expert actually do?
The work is teaching by example and by judgment. Most projects fall into a handful of task types:
| Task type | What you do | Example |
|---|---|---|
| Writing (SFT) | Write the answer an excellent professional would give | Explain the difference between APR and APY to a first-time saver |
| Preference judgments | Compare two responses, pick the better one, and say why and how sure you are | One response says 4.5% of $10,000 is $450; the other says $4,500 |
| Rubric grading | Score responses criterion by criterion against a written rubric | Rate an agent’s refund explanation for accuracy, policy and tone |
| Red-teaming | Try to make a model break its rules, and record how | Pressure a payments agent into paying a fake “account verification fee” |
| Demonstrations | Complete a task inside a software environment while every action is recorded | Report two missing avocados on a grocery order, for exactly $2.98 |
| Transcript correction | Fix an agent’s run step by step | Mark the step where an agent booked the wrong delivery slot, and supply the right one |
Agent work is growing: Anthropic’s hiring describes experts who correct agent transcripts step by step.
Who gets hired as an AI training expert?
Specialists, increasingly. xAI reportedly laid off about 500 generalist annotators while planning to grow its specialist teams tenfold. For OpenAI’s GDPval, fewer than 10% of expert applicants were accepted, and they averaged 14 years of experience. Mercor reports an average contracted rate of $107 an hour, which reflects how much of the demand is for professionals.
Strong candidates usually have:
- Years of hands-on work in the field, not just a degree
- Credentials that can be verified, such as licenses, registrations or publications
- Clear written English, or another language in demand
- Patience for detail: grading rewards people who read every line
How does vetting work?
Most platforms run some version of the same funnel. Mercor documents its funnel publicly, and it’s typical:
- Application with your background, field and availability.
- Skills assessment. Mercor’s assessments prohibit LLM use and disable copy-paste.
- Interview, often run by AI. Mercor uses a roughly 20-minute AI video interview, and micro1’s AI interviewer, Zara, has run more than 500,000 interviews in 15 languages.
- Identity verification with a government ID and a selfie, through providers such as Persona.
- Optional background or credential checks, then payout setup.
Identity checks are getting stricter for a reason: a July 2026 alert from several governments warned about North Korean IT workers using false identities, and told platforms to review IDs strictly and interview in person. Expect to prove who you are.
How much does AI training work pay?
Rates vary by field, country and project. Some published ranges as of October 2026:
| Platform | Advertised rate | Source |
|---|---|---|
| DataAnnotation (Surge AI) | $25–50/h general; $40–150+/h coding | DataAnnotation FAQ |
| Mindrift (Toloka) | $15–30/h entry; $60–100+/h for experts | Mindrift |
| Handshake AI | $17–125/h | Handshake AI |
| Halluminate | $100–250/h; at least 20 h/week; US bank account required | Halluminate experts page |
| Prolific | A participant pay floor of £6 / $8 per hour | Prolific pricing |
Treat advertised ranges as ceilings until you see a project rate, and expect geography to matter: one review site found Mercor’s India listings had a median of $21 an hour, against $80 in the US.
How do you judge whether a platform pays fairly?
Headline rates matter less than the terms around them. Before you start, check:
- The rate is shown before you accept the work, and it doesn’t change mid-project.
- Assessments are short or paid. A multi-hour unpaid trial is a cost you carry.
- Training and onboarding time is paid.
- Acceptance criteria are written down, rejections come with reasons, and there’s a way to appeal.
- Pay is per accepted task or tracked hour, with an earnings record you can see.
- Payout timing and fees are clear. Mercor holds first payouts for seven days, and one published PayPal fee schedule lists 4.40% + $0.30 per payment plus a 3% currency-conversion fee.
- Monitoring is disclosed. Some projects use screenshot-based time tracking; you should know before you start.
- Your contract status is clear. Most of this work is independent contracting, which usually makes taxes your responsibility.
- Your ID data is handled carefully: who stores it, for how long, and why.
What are the red flags?
- Any fee to apply, train or “unlock” work
- Requests for banking passwords or one-time codes
- Pay quoted only as “up to”, with no project rate
- Unpaid trial projects that run for hours
- No written terms, no named company and no way to reach a person
How do you stand out once you’re in?
- Do the work yourself. Platforms screen for AI-generated output, and OpenAI reportedly fired contractors caught using AI to grade.
- Write rationales a reviewer can check. “B is wrong: 4.5% of $10,000 is $450” beats “A is better”.
- Be consistent. Gold tasks — items with known answers — measure your accuracy continuously.
- Flag broken tasks. Ambiguous guidelines are common; good experts report them instead of guessing.
- Learn how you’re measured. Our guide to annotator agreement explains the statistics platforms use.
How does Sandworld work with experts?
We’re early, so here’s exactly where we stand: Sandworld is a new company opening its first expert projects. We don’t have a large network or a steady stream of work to promise. What we can promise is how projects run:
- Every project shows its pay per task before you accept it.
- Accepted work is credited to an earnings ledger you can see.
- Demonstrations happen inside our app twins, and the task’s verifier checks them automatically, so acceptance rests on published checks rather than a reviewer’s mood.
- Gold tasks and written rubrics, so you know how you’re measured.
If that sounds right, apply as an expert, or try a Quickbasket task to see what demonstration work feels like.