Working for the AI Labs

One company pays its contractors more than $1.5 million a day. It has over 30,000 people on its books, and they earn an average of more than $85 an hour. Top rates reach $200. Its customers are OpenAI, Anthropic, Meta and Google.

That company is Mercor, it was valued at $10 billion in October 2025 and was reportedly discussing $20 billion by July 2026, and almost nobody writing about AI side income mentions it. The reason is straightforward. What it sells is none of those things: it is hourly work, done by you, on someone else’s platform. It is hourly work. You cannot make a course out of it.

It is also, on the evidence, the best-paid AI earning route available to an individual, with one large caveat about geography that we will get to before any of the numbers, because for a lot of readers it decides the whole question.

What the work consists of#

The labs have run out of easy data. Scraping the web got them to where they are, and getting further requires something the web does not contain: correct answers from people who know the subject.

So they buy expert judgement. In practice that means:

• Evaluating model answers. Two responses, you say which is better and why. This is the RLHF work that most annotation platforms run, and it is the lowest-paid tier.
• Writing reference answers. A hard problem in your field, and you produce the answer a model should have given. Paid much better, because very few people can do it.
• Domain expert tasks. A doctor reviewing clinical reasoning, a solicitor checking a contract analysis, a senior engineer assessing whether generated code would actually survive production. This is where the $200-an-hour rates live.
• Red teaming. Deliberately trying to make a model behave badly, then documenting how you did it.
• Filling in forms and writing reports about how work is really done in your profession, the tacit knowledge that exists in practitioners’ heads and nowhere on the internet.

Note what is being bought. Not your time as a generic worker, but your credential and your experience. The labs are paying for the thing your career gave you, and paying by the hour for it, on a schedule you set.

Start with the geography problem#

The $85-an-hour average is not open to anyone with an internet connection, and the eligibility rules matter more than the rates do.

There are effectively two tiers of platform:

High-paying, geographically restricted. Mercor is largely US-focused. DataAnnotation is US, UK, Canada, Australia and New Zealand. These are the platforms where the impressive numbers come from, and if your country is not on the list, no amount of qualification will get you in. Country lists also change, so the only reliable answer is the platform’s own current page.

Worldwide, much lower rates. Appen, TELUS, Toloka and Clickworker hire globally and pay regional rates that are a fraction of the figures above. This is real work and real money and it is not $85 an hour.

Sanctioned countries are excluded from essentially all of it. Russia, Iran, North Korea and Belarus are on that list.

Being outside a platform’s country list is reportedly the single most common reason a qualified applicant is rejected. Check before you spend a day on an application, not after.

What it pays#

The platform-level figures are well sourced. The per-task rate tables are not: they come from industry aggregators, not from the companies, so read the shape of them and ignore the exact numbers.

Most Mercor contractors reportedly sit in a $40–75 range rather than at the top of it. Monthly, the commonly reported figures are $1,500–3,500 for strong written work at 30–40 hours a week, and $3,500–6,000 for STEM specialists.

And the number that matters more than any of those: your effective rate is lower than the posted rate. Screening is unpaid. Hunting for available tasks is unpaid. Failed qualification attempts are unpaid. Anyone reporting real earnings from these platforms mentions this, and the gap is not small.

Platform figures from the press; the per-task table and country lists are from aggregators. Sources: CNBC, TechCrunch, Bloomberg, The Information.

Who is buying so you know this is not one company#

Mercor is the loudest but not the only one. Surge AI, Scale AI (through Outlier), Handshake and Micro1, which raised at a $500 million valuation in September 2025, all sell the same thing to the same labs. Prolific and DataAnnotation serve the more general end.

That matters for two reasons. The demand is a market and not one company’s bet on a market. And when one platform’s work dries up, applying to two more is the obvious next step, not a defeat.

Getting in#

The application is the work sample, and this is where most people fail for avoidable reasons.

1. Check the country list first. On the platform’s own site, today.
2. Apply with your actual credential foregrounded. These platforms are buying your profession. A CV that reads “freelance writer, interested in AI” is worth much less to them than “eight years as a paediatric nurse” or “senior backend engineer, Go and Postgres, seven years”. The specific qualification is the product.
3. Expect an unpaid assessment. Usually a writing or reasoning task in your domain. It is a filter, not a formality, and it is where the rate you get offered is decided.
4. Write carefully in the assessment. Almost all of this work is judged on written explanation quality — why answer A is better, not just that it is. Clear, specific, structured reasoning is the entire skill being tested, whatever your field.
5. Apply to three or four platforms, not one. Work volume is unpredictable on all of them, and having qualifications on several is the only real defence.

The complaints, in order of frequency#

Consistently, across every source we found, and worth knowing before you rely on this:

Sudden deactivation is the number one complaint. Accounts get closed with no warning and no explanation and pending pay can be at risk when it happens. Take payouts promptly rather than letting a balance accumulate.

Work volume is unpredictable. “Approved but no tasks” is a widely reported state, and 2026 saw a lot of reported dry spells. This is contract work with no guaranteed hours, which means it is a poor primary income and a good supplementary one.

Communication is opaque. Failed assessments without notice, projects ending without explanation, no route to a human.

Rates have drifted down on some platforms since 2024. Not everywhere, and not in the expert tiers, but the general-annotation end has softened.

Newer entrants are wobbly. Alignerr is the 2026 platform people ask about most, and there have been unpaid-work reports. With any new platform, do one small piece of work and confirm you get paid before doing a second.

The scams that target exactly this#

Because the real thing is obscure and the pay is good, the impersonation business is brisk. Two rules cover nearly all of it.

You are never asked to pay. No registration fee, no equipment deposit, no training course, no “verification payment”. Every legitimate platform here pays you.

Apply only through the platform’s own domain. Not a link in a message, not a recruiter on Telegram or WhatsApp, not a job board listing that takes you somewhere else. Type the address yourself.

The related scam is the overpayment cheque, you are “hired”, sent a payment larger than expected, and asked to return the difference. The original payment reverses and the money you sent is gone.

Who this suits#

People with a credential or a real skill in a field the labs need: medicine, law, finance, engineering, accountancy, teaching, the trades, professional writing, translation. If you have spent years learning something specific, this route values it more than any other on the site.

People who want money without building an audience, a product, or a brand. There is nothing to grow here. You do work, you get paid, and it stops when you stop.

People who want flexible supplementary income around another job, which is what it is good for.

Who it does not#

Anyone outside the eligible countries for the high-paying platforms. That is most of the world.

Anyone who needs a reliable monthly figure. The volume is not guaranteed and the dry spells are real.

Anyone hoping to build something. This does not compound. After two years you have earnings and no asset, which is fine if that is what you wanted, and a waste of two years if it wasn’t. Using it to fund something you own is the sensible version.

Anyone whose skill is general rather than specific. The general-annotation tier pays least, is most crowded, and is most exposed to being automated by the very models it is training.

Is this going away?#

Nobody knows, and there are arguments in both directions.

Against: the whole point of this work is to make models better at things humans currently do better, and each round of it makes some of the previous round’s tasks automatable. The general-annotation end of the market is the obvious casualty.

For: the money going into it grew through 2026, the valuations went up, and the demand moved towards harder expertise. As long as labs are competing on capability at the frontier, they need people who know things the internet does not contain.

What that suggests in practice: treat it as good money now and not as a career, take it while the rates hold, and put some of it into something you own.

Country eligibility and rates change frequently, check the platform’s own pages before applying, and tell us if something here is out of date: [email protected].