What is an AI trainer, and how do you become one

What the job is, what it pays, what it asks of you and how people actually get in, read from the live job board rather than from a recruiter's pitch

0 AI training roles are open on this site's job board today, posted by the companies that pay people to teach, test and correct AI models. The typical role pays Unavailable to Unavailable an hour, with a median of Unavailable, and the range behind those numbers runs from transcription and language work at the bottom to physicians, lawyers and research scientists at the top. Every one of those roles is a version of the same job, the one the industry calls AI trainer, and almost every guide to it is written by a company that wants your application.

This guide is written the other way round. It starts from what the open roles themselves say: the descriptions, the application forms and the posted pay of thousands of live listings that this site reads every day. Our research reports turn those listings into numbers: what AI training roles require, how AI training applications work and how much of the work is open to people without experience. Those reports are recomputed every day, so this guide says what they find in words and links each finding to the figure in the report that states today's number; the three charts drawn from them are dated snapshots of 13 Oct 2026, and each bar opens its row in the live report. Counts of open roles and pay figures written in the text are read live from the board when you load this page, and each one opens the list it counts.

By the end you will know what an AI trainer does in each kind of work, who hires them, what the work pays and why the range is so wide, what employers ask for (degrees, licences, years, tests, tools), how the applications work, and a step-by-step route in, whether you are a nurse, a software engineer, a bilingual student or someone with no specialty at all.

Contents

1. What an AI trainer is
2. What AI trainers do, kind of work by kind of work
3. Who hires AI trainers
4. What AI trainers are paid, and why the range is so wide
5. What the roles require: degrees, licences, experience, hours and tools
6. Can you become an AI trainer without experience?
7. How applications work: forms, screening questions, tests and AI interviews
8. How to become an AI trainer, step by step
9. Where to start, by background
10. Risks, red flags and what the job is not
11. Where the job is heading
12. Questions people ask

1. What an AI trainer is

An AI trainer is a person paid to produce the human judgment that AI models learn from. A large language model is first trained on huge amounts of text, but what makes it useful, accurate and safe comes later, from people: people who write the answers it should give, rank two of its answers against each other, grade its reasoning step by step, catch where it is confidently wrong, try to make it misbehave, label the images and audio it learns to see and hear, and record the speech it learns to understand. Every one of those tasks is AI training. The job titles vary (AI Tutor at xAI, Expert at Mercor and AfterQuery, Fellow at Handshake AI, Contributor at Outlier, Annotator or Rater elsewhere), but the work underneath is the same: your judgment becomes training data.

The title is new; the work is not. Search engines have paid human raters to grade results for two decades, and transcription and translation vendors have sold labelled speech and text for longer. What changed after 2023 is the level of skill the labs want. A model that already writes fluent English does not need a crowd to tell it which sentence sounds better; it needs a cardiologist to say which treatment plan is wrong, a tax attorney to rewrite a flawed answer, a mathematician to find the broken step in a proof. That shift is why the same job title now covers transcription work whose listings state a median of Unavailable an hour and specialist review in law, medicine and finance at a median of Unavailable, and why "AI trainer" is less a single job than a ladder.

It also helps to be clear about the employment model. Almost all AI trainer work is remote, hourly contract work. You are usually an independent contractor of a platform (Mercor, micro1, Outlier, Turing and many others) that sells the finished data to an AI lab, and you are paid by the hour or by the task for the work you log. A minority of roles are direct: xAI hires its AI Tutors itself, as employees or contractors depending on the role, which is described on its platform page. Either way, there is rarely a fixed salary, rarely guaranteed hours, and work comes in projects that open and close.

The diagram is the whole business model in one line: the lab pays the platform for data, the platform pays you for the hours that produce it, and a reviewer checks your work before it ships. Knowing this explains most of what follows. It explains why pay depends on the field (the lab pays more for data it cannot get elsewhere), why screening is strict (one careless trainer spoils a dataset), and why work arrives in waves (a project ends when the lab has enough). The rest of this guide is about where you fit in that loop.

What a task actually looks like

A concrete example makes the job easier to picture than any definition. Suppose a project asks you to compare answers to health questions. You see a prompt a real user might type, such as "I take ibuprofen for back pain; is it safe to add my partner's prescription naproxen?", and two answers the model produced. Answer A explains that both drugs are NSAIDs, that taking them together raises the risk of stomach bleeding and kidney injury without adding much pain relief, and suggests asking a pharmacist about alternatives. Answer B says the combination is fine at normal doses and gives a dosing schedule.

Your task has three parts. First you rate each answer on the project's dimensions, typically accuracy, completeness, safety and helpfulness, each on a fixed scale. Then you choose the better answer (here A, decisively, because B is unsafe). Then you write a rationale: two or three sentences a reviewer can check, such as "B recommends combining two NSAIDs, which increases bleeding and kidney risk; A identifies the duplication and redirects to a pharmacist." On many projects you then rewrite the weaker answer into the one the model should have given. A clinician on the same project would get harder prompts, where the right answer needs current guidelines rather than general knowledge, and would be paid accordingly. Ask the same duplication question about two GLP-1 medicines, for instance, and the answer turns on label wording, which GLPcompare, from the founder of this site, quotes brand by brand.

Every kind of AI trainer work is a variation on that pattern: an input, a judgment against written guidelines, and an explanation another person can verify. In coding work the input is a programming task and the judgment is whether the code runs and is correct; in transcription it is a recording and the judgment is every word; in red-teaming you write the input yourself, trying to make the model break its rules. The guidelines document is the center of every project, and reading it carefully is the single habit that most separates trainers who keep getting work from those who do not.

2. What AI trainers do, kind of work by kind of work

This site sorts every open role into one of twelve kinds of work, the same taxonomy used on the pay explorer and the role pages. The kind decides what your day looks like, what you need to know and what you will be paid, so it is the first thing to understand about the job. The table below shows each kind with the number of roles open in it right now and the median rate its listings state. Each count opens the list of those roles, and each kind links to a full page with its pay range, its platforms and a guide to the work.

A median is the middle listing that states an hourly rate, so half of the listings in a kind pay more and half pay less. Where a kind has too few stated rates to measure, its cell says so rather than guessing. The counts in the table show which kinds dominate the board on the day you read it, and law, medicine and finance is the one that pulls the pay of the whole market up.

Evaluation and rating: the core of the job

RLHF stands for reinforcement learning from human feedback, and it is what most people mean by AI training. You are shown a prompt and one or more answers the model gave, and you judge them: which answer is better, whether each one is accurate, complete, safe and well written, and what exactly is wrong. Many tasks then ask you to write the ideal answer yourself, or to rewrite the model's answer until it is right. The skill is not typing speed; it is careful reading and the ability to explain a judgment in a sentence another reviewer will agree with.

This kind of work is open to the widest range of people, because a generalist rater judges everyday questions while an expert rater judges questions in their own field. The AI evaluation guide explains the task types and the rubrics in detail. The common thread is consistency: platforms measure how often your ratings agree with a reviewer's, and people who are accurate and steady get more projects.

Professional and expert work

Law, medicine and finance roles ask a professional to review the model inside their profession: a physician grading diagnostic reasoning, an attorney rewriting an answer about a contract clause, an accountant checking a tax calculation. These roles usually ask for the credential itself (a medical licence, bar admission, a CPA) and recent practice, and they pay the most. This site keeps dedicated lists for the professions people search for by name: AI training jobs for physicians, AI training jobs for nurses, remote healthcare jobs for every clinician, remote attorney jobs and remote finance jobs.

Math and STEM and coding work belong to the same expert tier. A mathematician writes problems the model cannot yet solve and checks every step of its solutions; a software engineer writes code, reviews the model's code and grades how well it solved a programming task. The coding guide and the math and STEM guide cover what these projects test. Agentic work, the newest kind, asks you to build realistic tasks and environments for AI agents that use tools (browsers, terminals, spreadsheets) and to judge whether the agent did the job.

Language, audio and data work

Multilingual roles need a native or near-native speaker of a language to evaluate and write in it, and they exist for dozens of languages, from Spanish and Japanese to Tagalog, Georgian and Sorani Kurdish. Transcription, audio and speech roles ask you to transcribe recordings precisely, record your own voice to a script, or judge the quality of generated speech. Data annotation covers labelling images, text and video, and robotics data asks people to record or label real-world physical tasks for robots. The robot side of this work, from steering robots by teleoperation to filming chores with a head-mounted camera, is explained kind by kind, with how people get in, in the guides on RobotTraining.jobs, which we also publish.

These kinds of work make up most of the roles that ask for no expert field in their title, which matters if you are starting without a specialty (section 6 shows where to read how many there are). Their pay is lower and more dependent on where you live, but a native language is itself a credential: many multilingual roles want exactly one thing, which is that you are fluent in a language the model handles badly. The multilingual guide and the audio and speech guide explain the tests these roles use.

3. Who hires AI trainers

The companies that hire AI trainers fall into a few kinds of platform, and the kind decides how you apply, how you are vetted and how you are paid. This site tracks the main ones in its platform directory, with a full page for each covering pay, payout, who can join and what workers report. The distinction that matters most is between expert networks, which screen for a profession or a skill and match you to projects, microtask crowds, which let almost anyone qualify for short tasks, and frontier labs that hire their own trainers.

The expert networks hold most of the open roles and most of the money. Today Mercor has 0 open roles at a median of Unavailable an hour, micro1 has 0 at Unavailable, AfterQuery 0 at Unavailable, Turing 0 and SuperAnnotate 0. OpenTrain AI aggregates roles from many hiring companies alongside its own ( 0 today). Some of the best-known names, such as Outlier and DataAnnotation, publish no job board at all: you apply once and are offered projects inside the platform, so the number of open roles there is not visible from outside.

The counts change daily as projects open and close, which is the most important practical fact about this market. A platform with hundreds of open roles this week may have a fraction next month, and the reverse. That is why experienced AI trainers join several platforms: being accepted somewhere is an option on future projects, not a job. The head-to-head pages help you choose between the ones you are considering, for example micro1 vs Mercor, Handshake AI vs Mercor, xAI vs Outlier and Outlier vs DataAnnotation, and the alternatives pages list the closest options to each platform when one goes quiet.

The frontier labs are a smaller, separate door. xAI hires AI Tutors directly for Grok, with the hourly band printed in each posting for US candidates ( 0 open today); its page explains how that differs from a marketplace. Handshake AI recruits degree-holding US talent through its university network ( 0 open today). Direct roles remove the platform's cut but are not steadier: in September 2025 xAI cut about 500 tutor roles in one day while moving to specialists, as its platform page records.

4. What AI trainers are paid, and why the range is so wide

The honest answer to "how much do AI trainers make" is a range, and the range is wide for a reason. Across every open role that states an hourly rate, typical pay runs from Unavailable to Unavailable an hour with a median of Unavailable. "Typical" here is one fixed definition used everywhere on this site: from the 25th percentile of what roles pay (the middle of each advertised range) to the 90th-percentile advertised ceiling. It describes the band most roles fall in, not the floor or the record, and the pay explorer shows it for every platform and kind of work.

Three forces set where a role lands. The first is scarcity of the judgment: a lab can find thousands of fluent English raters but only dozens of practising oncologists willing to grade its answers, so the oncologist's hour is worth many times more. The second is the stakes of the data: medical, legal and financial answers that a model gets wrong cause real harm, and labs pay for reviewers whose corrections they can trust. The third is geography: many language and transcription roles are priced by country, so the same task can pay very differently in Manila and in Munich. Section 6 shows what this means for people without a specialty: roles with no expert field in their title advertise a median well below the rest of the board, and the work without experience report states both medians as they stand today.

What the averages hide is just as important as what they show. Posted pay is not earned pay. A role that states an hourly rate pays that rate for the hours you log on a live project; it does not promise any hours at all. Many projects cap weekly hours, some pause without notice, and the time you spend on unpaid qualification tests is yours. Some listings advertise only a maximum ("up to" a rate), which is a ceiling rather than a rate; this site's pay figures leave those out of the medians for that reason. The earnings calculator turns a rate into a realistic monthly figure for the hours you can actually expect.

Reading a pay figure correctly

Job listings state pay in several shapes, and reading them the same way is the only way to compare them. A single rate ("$X/hr") is the simplest: that is what an hour of logged work pays. A range ("$X to $Y/hr") usually means the rate depends on your assessment result, your credentials or your country, and most people start nearer the bottom; this site measures a range by its middle. A ceiling ("up to $X/hr") is the most a project pays anyone, which is why it is left out of every median here. And a per-task rate ("$X per task") is only comparable once you know how long a task takes, which the listing rarely says; a test run of a few tasks tells you your real hourly figure.

Two more details change what a rate is worth. The first is who pays the fees: platforms that pay through Stripe, Deel, PayPal or Wise each charge differently for withdrawals and currency conversion, and a platform's page here lists how it pays. The second is unpaid time: onboarding courses, qualification tests and reading a new project's guidelines are often unpaid, so the first weeks on a platform earn less per hour than the posted rate. Neither is a reason to avoid a platform; both are reasons to judge a rate after a month, not after a day.

Pay also differs between platforms for the same kind of work, sometimes sharply. The highest-paying platforms ranking orders them by the typical pay their open roles state, and the AI training pay report tracks how the market moves over time. If you are choosing between two specific platforms, the comparison pages put their pay side by side with the barrier to entry, which is usually the trade-off that matters: the best-paid platforms are also the hardest to get into.

5. What the roles require: degrees, licences, experience, hours and tools

The role requirements report reads the full description of every open role whose job board publishes one, and counts what each asks for. Under each figure it states how many roles it read that day and from which platforms, which is most of the board. A mention can be a requirement or a preference ("a PhD is a plus"); the report counts both, so the figures are an upper bound on what is strictly required.

Degrees

Most roles name no degree at all. The report sorts the rest by the highest degree each mentions, from a bachelor's to a PhD. The headline hides a split by kind of work that is the most useful thing in the report: Math and STEM is the kind of work where a PhD is mentioned most often, while transcription and audio work almost never names a degree.

Law, medicine and finance stands out: more than half of its roles name no degree, yet it is the most credentialed kind of work on the board. That is not a contradiction. Those roles ask for the professional licence rather than the degree, because the licence implies the degree and proves current standing. A physician role asks for an MD or DO and a licence to practise, not for "a bachelor's in biology". The same logic explains multilingual work, where the degree named most often is only a bachelor's: what those roles really screen for is the language.

Professional credentials

A minority of roles mention a professional credential. The report's count by credential puts any professional licence first, then a medical degree (MD or DO), a law degree or bar admission and a CPA or CFA, with board certification and a nursing licence (RN or NP) after them. Almost all of them sit in law, medicine and finance. If you hold one of these, it is the most valuable line on your application, and it is worth putting it in the first sentence of your profile, in the exact form the listing uses.

Years of experience

Only a minority of roles state a number of years, and where they do the median ask is a few years, with three to four years the most common band. Experience here means experience in the field (clinical practice, software engineering, teaching), not experience training AI: section 7 shows that application forms rarely ask about prior AI training work.

Hours

Only a minority of roles state weekly hours. Those that do mostly ask for part-time hours, and a smaller group asks for forty hours or more. In practice most AI trainer work is part-time and flexible: you log hours when a project has tasks and when you are available, which is why so many people do it alongside a job, a degree or a practice. The roles that state forty hours are usually full-time contracts on a single long project.

Tests and interviews

A minority of roles mention an assessment, test or trial task, and fewer still an AI-led or recorded video interview. These are the real gates of the job, and they matter more than the requirements above: a role can name no degree and still reject most applicants at the test. The AI interview guide explains the AI-run interviews used by Mercor, micro1 and Braintrust in detail, including how they are scored and how to prepare.

Tools

The tools named most often are Python and Excel or spreadsheets, followed by engineering tools such as CAD and Git or GitHub. Spreadsheets appear across finance, data and evaluation work, and the report shows for each tool what share of the roles naming it are coding roles. If you are a technical applicant, a GitHub profile with real work in it is the equivalent of a licence for a clinician.

6. Can you become an AI trainer without experience?

Yes, but less of the market is open to beginners than the advertising suggests, and the work without experience report measures exactly how much. It reads every open role's title and kind of work and counts generalist roles: roles whose title names no expert field, level or credential, in kinds of work that are not professional by definition (law, medicine and finance, and coding, have no generalist roles). Only a minority of open roles are generalist roles, and they advertise a median hourly rate well below the board's median of Unavailable.

Another large group of roles is kept out only by a word such as "Expert" or "Specialist" in its title, and most of those name a language: a "Tagalog Language Expert" is often a fluent speaker, not a credentialed one. Counted in, generalist roles would be a far larger share of the board. So the realistic answer depends on what you bring: with a native language, a large part of the board is plausibly within reach; with nothing but English and time, the door is narrower.

The chart shows where the beginner roles were on 13 Oct 2026, and the report's live table shows where they are today: multilingual work and audio and speech have the most, and robotics data has the highest share of generalist roles, but few roles in total. The pay of these roles is what you would expect from section 4: within the same kind of work, generalist roles advertise markedly less than other roles. The generalist median is still a real wage in most of the world, but it is the entry rung of a ladder, not the market's typical rate.

The descriptions tell a sharper story than the titles. Fewer than half of the roles with a full description mention no degree, no years of experience and no credential, though most roles with a generalist title do. And while many descriptions say no prior AI experience is needed, nearly all of those still ask for expertise in the title or mention a degree, years or a credential. "No AI experience needed" usually means "we will teach you the AI part; we still need your field". Only a small share of descriptions say outright that a beginner may apply, and most of those still have a title asking for a field.

What do the roles with no stated requirement still ask for? A minority mention a test, more than half a native or fluent language, and those that state hours ask for part-time hours. That is the practical recipe for a beginner: a language, a test and part-time hours. If you speak a language other than English natively, start with the multilingual roles; if you do not, the evaluation roles and audio roles are the widest doors, and the microtask platforms described on the platforms page accept people with no specialty after a short qualification.

There is also a slower, better route that most guides skip: become an expert in something the labs need. The expert premium is the defining feature of this market, and it rewards depth in fields that have nothing to do with AI: nursing, accounting, chemistry, law, a second language at professional level. If you are a student, the field you are studying is your future AI trainer credential. The learn hub has a guide for each kind of work that explains what it takes to qualify.

7. How applications work: forms, screening questions, tests and AI interviews

The applications report reads the application forms of open roles straight from the platforms' own applicant-tracking systems (Greenhouse, Ashby, Lever, Workable and others) and counts what they ask. It reads more than a thousand forms, over a third of all open roles, and states the exact count under each figure. It is the closest thing to seeing every application in the industry at once.

What the forms ask for

Most forms require a resume and most ask for a LinkedIn profile; only a small share ask for a cover letter, and almost none requires one. Nearly every form asks for a phone number and your location or country. A GitHub profile is asked by a minority of forms and a portfolio or website by fewer. The message is plain: your resume and LinkedIn profile do the work, and a cover letter is rarely read because rarely asked.

The screening questions

Beyond the documents, most forms add questions of their own. The theme asked most often is languages spoken or fluency, followed by degree or field of study and experience or years in a field; expected pay or rate, work authorization or visa sponsorship, prior AI training or annotation work and availability come further down the list. Notice how much less often prior AI training work comes up than your languages: a small share of forms ask about it. Platforms recruit for your field and your language, then teach you their task format.

The expected-pay question deserves a note because it trips people up. When a form asks your rate, it is usually matching you to a project's budget; a number far above the role's stated pay can drop you from consideration, and a number far below it can pin you there. The safest answer is the role's own stated rate or the typical pay for that kind of work. Language questions are the opposite: be exact and honest about your level, because many projects test it.

How long it takes

A typical form has about ten fields, and a sizeable minority run to sixteen or more. The longest forms belong to a few platforms that ask many questions of their own, and the shortest, such as Terac's, ask little beyond your contact details; the table by platform ranks them. A long form is not a bad sign: it usually means the screening questions replace a first interview. Greenhouse runs the largest share of the forms read, so if you apply widely you will see the same form layout again and again.

After the form: tests and AI interviews

For most platforms the form is the easy part. What follows is one of three gates. A skills assessment (a written test in your field or a sample task) is the most common: DataAnnotation and Outlier screen this way, as do many expert networks. An AI interview, a spoken conversation with an AI that scores a transcript of your answers, is the gate at Mercor, micro1 and Braintrust; the AI interview guide covers it end to end. A language test gates most multilingual work. Each platform's page on this site describes its own steps under "How to get started".

The results arrive on very different clocks. An AI interview can return a result the same day; a skills assessment at a platform with no job board can take weeks, and some never reply. This is the second reason experienced trainers apply to several platforms at once: waiting is the longest part of becoming an AI trainer, and running applications in parallel turns weeks of waiting into days.

8. How to become an AI trainer, step by step

The route below is the one the data supports. It is ordered so that each step makes the next one easier, and it works whether you are a specialist or a beginner. None of it costs money: a legitimate platform never charges you to apply, to take a test or to get work.

Step 1: Choose your lane

Start from what you already have, not from what pays most. Your lane is the kind of work where your background is a credential: your profession, your degree, your native language, your programming skill. The kinds of work in section 2 and the table in section 9 map backgrounds to lanes. A clear lane makes every later step sharper, because platforms screen for a field, and an applicant who is "open to anything" reads as an applicant with no field.

Step 2: Make your resume and LinkedIn say it in the first line

With most forms requiring a resume and most asking for LinkedIn, these two documents are your application. Put your lane in the first line of each: "Registered nurse, 6 years in pediatric ICU", "Native Japanese speaker, professional translator", "Backend engineer, Python and Go, 5 years". List the credential exactly as listings write it (MD, RN, CPA, JD, PhD in Chemistry), the years, and the tools from section 5 that apply. An AI interview reads your resume to choose its questions, so the resume also decides what you will be asked.

Step 3: Apply to three to five platforms in your lane

Use the job board filtered to your kind of work and country to see who is hiring for it today, and pick three to five platforms: one or two with open roles in your exact field, plus one or two that keep a general pool. Check each platform's page for who can join (many take workers only from certain countries) and how it pays. Apply to specific roles where a platform lists them; a role-specific application tells the platform which project you fit.

Step 4: Prepare for the gate, then take it seriously

Find out which gate each platform uses (test, AI interview or language check) and prepare for it the way you would for a job interview, because it is one. For a skills test, practise in the format the platform describes. For an AI interview, read the AI interview guide and do a practice run. Most gates allow very few attempts, and a failed attempt can close a platform to you for months.

Step 5: Do the first project well, then let quality compound

Once you are accepted, the first project is where your quality score is set. Read the guidelines twice, follow them even when you disagree, and write your rationales in plain, specific sentences. Platforms route more and better-paid work to trainers whose ratings agree with reviewers and whose writing needs no fixing. Track your hours and payments yourself from day one.

Step 6: Climb

The market rewards depth. Once you are working, move toward the better-paid kinds of work your background allows: from general rating to expert rating in your field, from transcription to language evaluation, from single tasks to reviewer and lead roles (quality assurance leads are a common step up). Keep two or three platforms active so a paused project never means zero income, and set a job alert for your kind of work so new roles reach you the day they open.

What the first month looks like

Most new AI trainers are surprised by how uneven the first month is. The first week is usually onboarding: an account, identity and tax forms, a payment method, and often a short paid or unpaid course on the platform's tools. Then comes a project's guidelines document, which can run to dozens of pages, and a set of practice or calibration tasks graded against a reviewer's answers. Only after that do paid tasks appear, and on some platforms the first project arrives weeks after acceptance, when a lab needs your profile.

When work does arrive, it tends to arrive in a burst. A project may offer as many hours as you can give for two weeks and then pause while the lab reviews the batch. Reviewers' feedback on your first tasks is the most valuable thing you will receive in that month: it tells you exactly how the project reads its own guidelines, which is rarely identical to how a newcomer reads them. People who answer that feedback by changing their work the same day are the ones who get the next project.

Payment follows its own schedule. Weekly payouts are common at expert networks ( AfterQuery, for one, pays every Friday through Stripe), monthly at older crowd platforms such as Appen, and almost every platform holds the first payment until your tax and payout details are verified. Budget for the first month as a month of setup with some income, and judge the platform by the second and third, when you know the work and the work knows you.

The six steps fit on one line: lane, profile, three to five applications, the gate, quality, then up. People who fail usually skip the first or the fourth: they apply everywhere with a generic profile, or they treat a test as a formality. People who succeed tend to look boring on paper: a clear field, a resume that says it in the first line, and patience through the waiting.

9. Where to start, by background

The same job looks different from each starting point. The table below maps common backgrounds to the kind of work where they count as a credential, with a live count of roles in that kind and the page that lists them. Each count opens the board filtered to that work.

For clinicians, lawyers and finance professionals, the advice is short: you are the scarcest and best-paid applicant in this market, the work fits around a practice, and the profession pages above list the roles that ask for your exact licence. Check your employer's policy on outside work, and treat a platform's screening as you would a credentialing process, because that is what it is.

For engineers and scientists, the competition is stiffer but the volume is high. Coding and STEM roles ask for real depth (Math and STEM is the kind of work that mentions a PhD most often), and agentic work rewards people who can build realistic tasks for AI agents, which is closer to software engineering than to rating. A public GitHub with substantial work, or published research, does more for you here than any certificate.

For multilingual speakers, the language is the credential. Multilingual work has more generalist roles than almost any other kind, and most of them ask first and foremost for native fluency. Pay varies with the language and the country, so compare the multilingual pay range with the role in front of you, and look for evaluation roles in your language, which pay more than transcription in it.

For people with no specialty, the honest route is two-track: take the generalist roles that are open (evaluation, audio, labeling, microtask platforms) to start earning and learn the work, and meanwhile build a specialty the market pays for. The country pages show how many roles accept applicants where you live, which matters more for generalist work than for any other kind.

10. Risks, red flags and what the job is not

AI trainer work is real work at real companies, but the industry has risks that recruiters do not advertise. The first is instability. Work comes in projects, projects end, and platforms change direction quickly: xAI cut about 500 tutor roles in September 2025 when it moved to specialists, and Outlier cut generalist projects after Meta took a 49% stake in its parent company in 2025, both recorded on their xAI and Outlier pages. Plan for income that rises and falls, and never let one platform be your only source.

The second is disputes over pay and accounts. Contractors can be removed from a platform, sometimes without a clear reason, and payment for work already done can be contested. The best-documented recent case is Handshake AI's: in 2026 Business Insider reported contractors whose accounts were suspended and who were not paid for completed work, and two lawsuits over withheld pay followed, as its platform page sets out. Keep your own record of hours and tasks, read the payment terms before you start, and prefer platforms whose payout reputation is documented.

The third is scams that imitate the job. Because "AI trainer" is now a popular search, fake recruiters use it. The rule that catches nearly all of them: a legitimate platform never asks you to pay (for training, equipment, a background check or a "starter kit"), never asks for your bank password, and never moves the conversation to a messaging app before you have applied through its own site. Each platform's page on this site links its official application address; apply there, not through a link someone sent you.

Contracts, taxes and confidentiality

Because most AI trainers are independent contractors, three pieces of paperwork deserve more attention than people give them. The contractor agreement sets how you are paid, when a platform may end the relationship and what happens to unpaid work if it does; read the payment and termination clauses before the first task, not after a dispute. Taxes are yours to handle: in the US, platforms that pay contractors report it on a 1099 form (Handshake AI's own page here describes its trainers as 1099 contractors paid hourly), and in most other countries contract income must be declared the same way as freelance income. Set aside a share of every payout from the first one.

The third is confidentiality. Most projects are covered by a non-disclosure agreement, because the prompts, the guidelines and the client are commercially sensitive. That has a practical consequence for your career: you often cannot name the lab you worked for or show the tasks you did. Keep a private record of the kinds of work, the hours and any quality ratings you received, which you can describe in general terms on a resume without breaking the agreement.

It also helps to know what the job is not. It is not a salaried career with benefits in most cases: you are a contractor and handle your own taxes. It is not a way to practise your profession: a physician grading a model's answers is not treating patients, and should check whether outside work is allowed. And it is not passive income: every dollar is an hour of focused, checked work. Seen clearly, it is flexible, well-paid contract work for people with judgment the labs need, and a modest but real income for many people without a specialty.

11. Where the job is heading

The direction is visible in the data already. Expertise is pulling away from generalist work: generalist roles advertise a median far below the rest of the board, and the gap holds within the same kind of work. Labs that once needed crowds to rank fluent sentences now need professionals to correct reasoning in medicine, law, finance, science and code, and the pay report shows the best-paid work concentrated there. The live numbers in section 2 let you watch that split day by day.

The second trend is agentic work: building tasks, environments and grading rubrics for AI agents that use software tools on their own. It is the newest kind of work on the board ( 0 open roles today) and it pays like expert work, because building a realistic task requires knowing how the real task is done. The third is frontier labs hiring their own trainers, as xAI does, alongside the platforms. And the fourth is screening by AI: AI-run interviews already gate a large share of expert roles, which makes the guide on how AI interviews work worth reading before you apply anywhere.

None of this makes the job disappear as models improve. Each generation of models needs harder data than the last, and harder data needs better-qualified people. What changes is who is in demand: the market keeps moving up the ladder of expertise, so the AI trainers with the longest future in this work are the ones who keep their field current, not the ones who learned one task format.

12. Questions people ask

Is AI trainer a real job? Yes. It is contract work, usually remote and hourly, paid by established companies that sell training data to AI labs, plus a few labs that hire directly. Today the board lists 0 open roles. What it usually is not is a salaried position with guaranteed hours.

How much do AI trainers make? Typical pay across open roles runs from Unavailable to Unavailable an hour, with a median of Unavailable. Professionals in law, medicine and finance earn the most (a median of Unavailable an hour today), and roles that ask for no specialty the least (the work without experience report states their median today).

Do you need a degree to be an AI trainer? Not for most roles: most roles with a full description name no degree. But a degree, a professional licence or a native language is what most roles really screen for, and math and STEM is the kind of work where a PhD is mentioned most often.

Do you need experience training AI? Rarely. Only a small share of application forms ask about prior AI training or annotation work. Platforms recruit for your field and your language and teach you their task format.

How long does it take to get accepted? It depends on the gate. An AI interview can return a result the same day; a skills assessment at a platform without a job board can take weeks. Apply to three to five platforms at once so the waiting overlaps.

Can you do it part-time? Yes. Roles that state hours mostly ask for part-time hours, and most projects let you choose when you work within a weekly cap.

Can you become an AI trainer from outside the US? Often, but not everywhere: many roles accept only certain countries, and some platforms can pay only in some countries. The country pages show how many open roles accept applicants in each country, and each platform's page lists who can join.