All roles / AI trainer jobs

AI trainer jobs, and what they really pay

Almost everything written about this is affiliate content recycling invented numbers. This is what the reporting, the court filings and the companies’ own documents actually say.

There are two markets, and they are moving in opposite directions

The single most useful thing to understand is that "AI trainer" describes two different jobs with different buyers, and confusing them is why the pay figures you see range from $15 to $1,000 an hour.

General annotation work, the kind with an open signup and a qualification quiz, pays roughly $15 to $30 an hour. That tier is being actively dismantled. xAI cut 500 of its roughly 1,500 annotators in September 2025, telling them it no longer needed generalist tutor positions. Scale AI cut 200 staff and 500 contractors the same year. Handshake’s chief executive said on the record that generalists are no longer needed.

Credentialed expert work is expanding at the same time, and pays $100 to $200 an hour routinely. At the top, the New York Times reported in July 2026 that Surge AI pays $300 to $500 an hour for management consultants and investment bankers and up to $1,000 for venture partners, and that Mercor pays up to $350 for psychiatry experts. Bloomberg independently reported the $1,000 figure in February 2026.

These are not contradictory reports. They are the same event. The important consequence: the lower tier is not an entry route to the upper one. They are separate markets, and time spent in the first does not qualify you for the second.

The platforms

Ten places that actually pay for this work, current as of 2026-08-30. Rates are quoted as the source states them. Where a company publishes a figure about itself it is labeled as such, because an average hourly rate supplied by the company selling the labor is marketing until somebody audits it.

Surge AI

Credentialed

Direct expert network. Recruits from named firms and institutions for narrow domain work.

Pay: $200 to $400/hr on the one rate published on its own careers page. Built In reported $250 to $450 for medical fellows, $300 to $500 for consultants and bankers, and $500 to $1,000 for VC partners and startup CEOs.
Getting in: Email talent@surgehq.ai with the role in the subject line. Requirements are extreme: the constitutional law listing asks for Supreme Court clerkship or senior counsel at a top firm.
Worth knowing: A May 2025 California class action alleges annotators were recruited on promises of $20 to $40 an hour while paid less, and were not paid for training videos, qualification exams or assessments. The company says the claims are without merit. Current status of the case could not be determined.

DataAnnotation.tech

Open entry

The mass-market channel. Court filings name the defendant as Surge Labs doing business as DataAnnotation, so this and Surge are the same company under two names.

Pay: Its own FAQ advertises $25 to $30+/hr general, $50 to $100+/hr for coding, STEM and professional tracks. TIME documented someone earning roughly $400 to $500 a week part time at $20 to $25/hr, the low end of the advertised band.
Getting in: Unpaid starter assessment of about an hour, one attempt only, no retakes. Rejections usually get no notification.
Worth knowing: Glassdoor sits at 4.0 across 643 reviews, better than its peers. The recurring complaint is not theft but drought: accepted workers reporting months with no projects offered.

Mercor

Credentialed

The fastest growing platform in the market. Matches credentialed professionals to lab projects.

Pay: Its homepage shows $112/hr as an average contracted rate, which is a company figure skewed by high-end specialists. Its own blog reports paying over $2M a day across 30,000 weekly active contractors, which works out closer to $465 a week each. Built In reported physicians at $130 to $170/hr and lawyers at $110 to $130.
Getting in: A roughly 20 minute AI video interview with a voice agent, up to three retakes. Then ID verification, background check and tax forms.
Worth knowing: Two things worth reading before signing. Deliverable-based contracts carry no minimum hourly guarantee and pay only on successful completion. And all hourly work must be tracked through a desktop surveillance application. In March 2026 a supply-chain compromise exposed contractor data; Mercor concluded a limited subset was affected, Meta paused its contracts, and contractor lawsuits followed. Fresh breach allegations appeared on cybercrime forums on 26 August 2026 and are unresolved.

Outlier

Both tiers

Scale AI’s contributor platform, handling generative AI work: prompt writing, response ranking, RLHF, coding evaluation.

Pay: Its own listings advertise about $30/hr average for US coding work and $12 to $32.50 for physics outside the US. Glassdoor shows a $34/hr median for coders across 13 submissions as of August 2026. On Indeed, 38% of 101 respondents said they were paid fairly.
Getting in: Government ID, a phone number from your country of residence, a resume and a LinkedIn profile. Minimum an associate degree, with many projects requiring more.
Worth knowing: Outlier’s own listing text states that "for non-core work, such as during initial project onboarding, lower rates may apply", which is the company confirming onboarding pays below the headline. Scale has agreed to pay $12.5M to settle California misclassification claims, with a claim deadline of 3 September 2026. A worker petition over withheld earnings passed 1,000 signatures in a day in July 2026.

Handshake AI

Credentialed

The campus recruiting company, refounded around AI training. It cut 15% of its US workforce in October 2025 to fund the pivot.

Pay: All figures are published as ceilings rather than typical rates: up to $40/hr for AI evaluation, up to $120 for philosophy, up to $125 for game design and PCB work. It has hired improv actors at up to $74/hr.
Getting in: US work authorization required. You build a profile and wait; projects find you on a rolling basis with no fixed timeline.
Worth knowing: The clearest documented payment problem in this market. Business Insider reported in March 2026 that five contractors on OpenAI projects had accounts suspended and four were never paid. The company’s position on record: "This decision is final. There is no appeal process, and any work associated with this violation is not eligible for payment." A separate small claims case ended with a court ordering Handshake to pay $6,475. Its own terms state that where a project sets a maximum handling time, payment stops there even if the task takes longer, and that setup, meetings and time lost to technical faults are unpaid.

xAI

Both tiers

Hires tutors directly rather than through a marketplace, which is unusual and matters.

Pay: $45 to $75/hr for US-based image tutors and $35 to $45/hr for Chinese language tutors, stated on the listings themselves.
Getting in: Apply through xAI’s own job board like any other role. Portfolio and credential based rather than assessment gated. Full time, part time or contract.
Worth knowing: You are applying to a named employer with a published band, so there is no marketplace deactivation risk and no unpaid qualification maze. The risk is different and demonstrated: on a Friday evening in September 2025, xAI cut 500 of roughly 1,500 annotators, telling them it no longer needed generalist tutor positions, while announcing it would grow the specialist team tenfold.

Prolific

Open entry

Research participation rather than annotation, though the line is blurring. It now runs a formal AI tasker tier for evaluation work.

Pay: The only platform here with an enforced floor: £6 or $8 an hour minimum, £9 or $12 recommended. Underpaying studies trigger enforcement, and researchers must pay adjustments within 22 days or have their account suspended.
Getting in: Signup is often queued rather than assessed. The waitlist has reportedly passed two million people.
Worth knowing: The least problematic platform in this set on payment reliability. The real constraint is not theft but ceiling: study availability is demographically gated and unpredictable, and the floor is a floor, not a typical rate.

Mindrift

Open entry

Toloka’s crowd platform, merged into Mindrift. Most guidance written before 2026 points at a worker platform that no longer exists under that name.

Pay: No rate figures could be verified from any primary source. Every number in search results traced back to affiliate sites. Treat any specific claim about Mindrift pay as unsourced.
Getting in: Free signup, open internationally. Its own help center warns that the set of available tasks will not grow overnight.
Worth knowing: Payouts moved from on-demand withdrawal to a fixed two-week schedule during the migration, which is a real change to cash flow if you are relying on this income.

Appen

Open entry

The old guard, now operating its worker platform as CrowdGen. Lost its Google contract in 2024 and has since returned to growth, though the growth is almost entirely in China while the Western business shrinks.

Pay: No reliable current rate data exists. The widely quoted Glassdoor figure is a model estimate built on two submissions and should not be cited. Historically this work sat near local minimum wage.
Getting in: Open signup with qualification tests per project.
Worth knowing: Its own 2025 annual report is the most useful source here: it reports over a million contractors and says worker satisfaction fell "due to project availability concerns". That is the company confirming that work drying up is the workforce’s main grievance.

Invisible Technologies

Both tiers

Operations-as-a-service running RLHF and training work for OpenAI, Amazon, Microsoft and Cohere. Clients pay for outcomes rather than hours, which pushes piece-rate pressure downward.

Pay: Published figures are internally contradictory. Glassdoor reports a $54/hr median while the individual submissions on the same page run from $5 to $29. The median is a model output, not observed pay.
Getting in: Application and assessment. You supply your own computer.
Worth knowing: The worst worker sentiment in this set: Glassdoor 3.2 with only 49% recommending, and recurring reports of layoffs without notice or severance. No litigation or journalism on wage withholding was found, so the negative evidence here is review aggregate only.

Telling fraud from a platform that just treats you badly

Employment scams are large and growing. The FBI logged 24,688 employment fraud complaints in 2025 with about $363 million in losses. FTC-reported losses to job scams rose from $90 million in 2020 to $501 million in 2024. The Better Business Bureau puts the median loss on task-based scams at $2,300, well above ordinary job fraud, and found that half of employment scam reports arrived by text message.

The difficulty is that legitimate platforms in this market do things that feel fraudulent. Unpaid qualification exams, sudden deactivation, work vanishing without explanation and paid work rejected without a reason are all documented practices at real, funded, well-known companies. So "this feels wrong" is not a usable test.

The test that works: money moves in one direction only.

You are being defrauded, with no exceptions, if you are asked to:

pay a fee, buy equipment, or purchase credits to unlock tasks
deposit money to release earnings you have supposedly accrued
take payment in cryptocurrency
deposit a check and wire part of it onward
move the conversation to Telegram, WhatsApp or Signal before any interview
process other people’s identity documents or payments through your own accounts

A real platform pays you and never bills you, runs onboarding on its own domain, issues a 1099 and asks for a W-9, and pays through PayPal, Stripe or direct deposit. Be especially careful with onboarding pages: legitimate signup genuinely does ask for your name, address, phone and government ID, which is exactly what makes a cloned onboarding flow worth building. Reach these platforms by typing the address yourself rather than through a link in a message.

One documented wrinkle worth knowing about: researchers at AlgorithmWatch found a black market in verified accounts, sold for around $70 so workers in lower-income countries can reach higher-paying regional tiers. Buying or selling access this way will get the account terminated with any balance on it.

What it costs you in tax

Every major platform engages US workers as 1099 independent contractors, which is being contested in court but is the current reality. Nothing is withheld. You owe self-employment tax of 15.3% on net earnings above $400, on top of ordinary income tax, and generally must make quarterly estimated payments if you expect to owe $1,000 or more for the year. Half the self-employment tax is deductible.

You also have no minimum wage protection, no overtime, no unemployment insurance and no severance. Deactivation is not termination, so it carries no notice and no appeal.

A change specific to this year: from tax year 2026 the threshold at which a platform must send you a 1099 rose from $600 to $2,000. Your obligation to report the income did not change. Somebody earning $1,500 across three platforms this year may receive no forms at all and still owe tax on all of it, so keep your own records rather than waiting for paperwork. None of this is tax advice; check it against IRS guidance or an accountant.

Practically, a $100 an hour contract rate is worth something closer to $70 an hour compared with a salaried wage once self-employment tax and absent benefits are counted, and that is before unpaid onboarding and rejected work.

What actually gets you into the tier that pays

The rate ladder tracks external billing rates almost exactly, which tells you what is being bought. Venture partners, consultants and bankers sit at the top, then medicine and law, then graduate-level STEM, then everyone else.

What moves you up is a terminal or licensed credential, a PhD, MD, JD, CPA or PE, or deep seniority in a profession that bills at a high rate. Recognizable employer history does real screening work: one documented expert cohort averaged more than seven years of experience and came from firms like Goldman Sachs, McKinsey and Latham & Watkins.

The most useful finding for anyone already technical comes from Epoch AI’s work on reinforcement learning environments: domain knowledge and expert-level prompting matter more than machine learning background. The work has moved from labeling answers toward designing tasks and graders that a model cannot shortcut. Enterprise workflow tasks, meaning real business processes in real software, are the named growth area.

What does not get you there: general writing ability, prompt engineering as a standalone claim, a bachelor’s degree without domain depth, or working more hours.

Common questions

How much do AI trainer jobs actually pay?

Two separate markets with different buyers. Generalist annotation pays roughly $15 to $30 an hour and is shrinking. Credentialed expert work pays $100 to $200 an hour, and at the very top, reporting has documented $300 to $500 for consultants and bankers and up to $1,000 an hour for venture partners. The lower tier is not a ladder to the upper one.

Is DataAnnotation.tech legitimate?

Yes, in the sense that matters most: money flows to you and never from you, and most workers are paid. Court filings establish that it is Surge Labs operating under another name. The common complaints are an unpaid entry assessment with a single attempt, and being accepted but then offered no work for long stretches.

Which platform pays the most?

Surge AI’s direct expert channel, if you can get in. It is also the hardest: its listings ask for credentials like a Supreme Court clerkship or a Pulitzer. For most people with a strong professional record, Mercor and Handshake AI are the realistic route to three-figure hourly rates.

Do I have to pay to start?

Never. No legitimate platform in this market charges a fee, asks you to buy equipment or credits, pays in cryptocurrency, or sends you a check to deposit and forward. Money moves in one direction only. That single test separates fraud from platforms that merely treat workers badly.

Will this work still exist in a few years?

Spending is rising sharply and the composition is shifting hard toward expertise. Whether the top rates persist is genuinely contested. Bloomberg reported in February 2026 that contractors themselves expect two years, maybe five, and there is a structural reason for that: once a model absorbs a specialty, projects in that specialty disappear.

How is this taxed in the US?

As self-employment. You are a 1099 contractor, nothing is withheld, and you owe self-employment tax of 15.3% on net earnings above $400 on top of income tax, usually in quarterly estimated payments. From tax year 2026 the threshold for a platform to send you a 1099 rose from $600 to $2,000, which does not change your obligation to report the income. Keep your own records rather than waiting for forms.

How long this lasts

Spending is rising steeply. Mercor’s contractor payouts went from $1.5 million a day in October 2025 to $4 million a day by July 2026. Anthropic leadership has discussed spending over $1 billion on reinforcement learning environments in a year.

Whether the top rates hold is contested, and the honest answer is that nobody knows. Bloomberg reported in February 2026 that contractors themselves expect two years, maybe five. There is a structural reason to take that seriously: once a model absorbs a specialty, work in that specialty disappears, and the people doing it are paid to make their own niche obsolete. Treat this as income with a horizon rather than a career track.

On sources. Everything above comes from journalism, court filings, government data or the companies’ own published documents, and allegations are described as allegations. Search results for these terms are dominated by affiliate sites that recycle each other’s invented figures, and none were used. Some things could not be verified and are marked that way on the platform entries. Worker forums were not reachable during research, which is a real gap, since that is where the richest reporting on deactivation and effective hourly rates lives. Written 2026-08-30. Corrections to jobs@aievalsjobs.com.

This board itself covers the other side of this market: staff roles in evals, post-training, red teaming and alignment, at 626 openings across 94 companies.

Browse the 626 open roles →