On a Friday night in September 2025, about 500 people who trained Grok got an email telling them their system access was gone. That was roughly a third of xAI’s 1,500-person data annotation team, and the same announcement said the company would “immediately surge our Specialist AI tutor team by 10x,” according to TechCrunch’s report on the cuts. Generalists out, specialists in, overnight.
That one email explains the market for anyone who wants to get paid to train AI in 2026 better than any platform’s recruiting page. The work is real and the money is real. But the pay has split into two very different jobs, and the cheaper of the two is being automated by the same models it trains.
This guide covers what the work involves, what it pays at each level right now, which platforms fit which kind of person, and how to avoid spending 20 unpaid hours on assessments for work that dries up a month later.
What you would actually be doing
Almost all of this work feeds a process called reinforcement learning from human feedback. A model produces an answer, a person judges it, and the judgment becomes training signal. The tasks come in four main shapes:
- Comparison and rating. Read two responses to the same prompt, pick the better one, and write a short justification against a rubric. This is the bulk of entry-level work.
- Writing. Draft prompts, ideal answers, or multi-turn conversations the model learns from. Usually pays more than rating because it is slower and harder to fake.
- Domain review. A lawyer checks legal answers, a nurse checks medication guidance, a developer checks whether generated code actually runs. This is where the high rates live.
- Red teaming. Try to push the model into harmful, biased, or wrong output, then document how you did it.
The work is remote and mostly asynchronous. You claim tasks from a queue when they exist. “When they exist” is the part that matters most, and we will come back to it.
What it pays now, by level
The rates that circulate online are mostly ceilings presented as averages. Here is what the primary sources actually say.
DataAnnotation, one of the larger platforms, now advertises its generalist role at “up to $50/hr,” software engineering roles at “up to $150+/hr,” and finance and other expert roles at “up to $125+/hr”. The words “up to” carry a lot of weight there. Those are the top of the band for the best-scoring workers on the hardest projects, not what a new contributor earns in week one.
Mercor, which sells expert contractors to AI labs, publishes tier ranges that are more useful for planning. Per a breakdown of Mercor’s own published bands, entry-level labeling and basic feedback work runs about $12 to $25 an hour, specialized evaluation and coding review about $25 to $53, credentialed professionals (physicians, attorneys, PhDs) about $75 to $200+, and senior domain experts with at least four years of professional experience above $200. The same source reports primary-care physicians at $130 to $170 an hour and lawyers at $110 to $130.
At the large contractor shops, generalist pay is lower still. Workers rating Google’s Gemini and AI Overviews through Hitachi’s GlobalLogic told reporters that wages start at $16 an hour for generalist raters and $21 an hour for “super raters”, as reported in The Guardian’s investigation (excerpted at that link).
Put those together and a realistic planning table looks like this:
| Your background | Typical work | Realistic hourly range |
|---|---|---|
| No specialty | Rating, labeling, simple comparisons | $12 to $25 |
| Strong writer, coder, or analyst | Writing tasks, code review, structured evaluation | $25 to $53 |
| Licensed or credentialed professional | Medical, legal, financial, scientific review | $75 to $200+ |
| Senior expert, 4+ years | Rubric design, expert-only evaluation | $200+ |
Ranges from Mercor’s published tiers, via HeroHunt.
The honest math for most readers: 10 hours a week in the bottom tier at $18 an hour is about $780 a month, if the queue stays full. That is useful side income. It is not a salary, and the queue rarely stays full.
If you are in the top two rows, the economics are different enough that we wrote a separate analysis of why AI labs pay credentialed experts $200 an hour and what that rate is really buying.
The platforms, sorted by who they want
Platform names change faster than pay rates. Scale AI’s old Remotasks brand, for example, has largely been folded into its Outlier platform, so an old Remotasks login is no longer the place to start. Rather than chase a ranking that will be wrong in six months, match the platform to your profile:
| Platform | Built for | What to know before applying |
|---|---|---|
| DataAnnotation | Writers, coders, subject experts | Long unpaid starter assessment; silence usually means you were not accepted |
| Outlier (Scale AI) | Generalists and STEM contributors | Absorbed Remotasks; volume swings with Scale’s client roster |
| Mercor | Credentialed professionals | AI video interview, project-based work, weeks of waiting between matches are common |
| Mindrift (Toloka) | Multilingual and domain experts | Wider country eligibility than most |
| Alignerr (Labelbox) | Writers and academics | Project-based, uneven availability |
| CrowdGen (Appen) | Generalists | Lower rates, broader global access |
Two practical notes. First, several of the better-paying platforms restrict applicants by country, so read the eligibility page before you invest hours in a test. Second, aggregator sites that promise “access to all platforms” for a fee are not worth paying for. Every legitimate platform in this table is free to join.
What happened to generalists over the past year
The xAI email was not an outlier. It was one of four signals in a short window, all pointing the same way.
Scale AI, July 2025. About a month after Meta’s $14.3 billion deal with the company, Scale laid off 200 employees (around 14% of full-time staff) and ended contracts with 500 contractors. Interim CEO Jason Droege’s memo said, “We scaled too quickly.” Some major clients reportedly pulled back after the Meta tie-up, and contributor volume followed.
Google’s raters, August 2025. More than 200 GlobalLogic contractors working on Gemini and AI Overviews were let go in at least two rounds with little or no warning, many of them holding master’s degrees or PhDs. One was told by email that it was a “ramp-down.”
xAI, September 2025. The 500 generalist cuts described above, paired with a public push to hire specialists in STEM, finance, medicine, and safety.
Mercor, November 2025. Mercor cancelled one project and offered workers a replacement at $16 an hour, down from $21. One worker called it “a slap in the face.”
The pattern is not that AI training is disappearing. Grand View Research projects the data annotation tools market growing from $1.0 billion in 2023 to $5.3 billion by 2030, a 26.3% annual rate. The money is growing. It is just moving up the expertise ladder, because models now handle a large share of the simple labeling themselves, and labs only need humans where the model cannot yet judge its own answer.
If you are entering at the bottom tier, plan for that. Treat the work as something that can end with one email, because for thousands of people in 2025, it did.
Running it like an operator, not a gig worker
I have spent more than 20 years in IT operations at a large telecom, and now do fractional COO work, and the single habit that separates people who make this work from people who burn out on it is the same one I apply to any vendor: measure the real rate, not the quoted one.
A task that advertises $30 an hour and takes twice the estimated time pays $15. An assessment that takes six unpaid hours and leads to three 10-hour weeks at $20 an hour pays $600 for 36 hours, which is about $16.67 an hour, not $20. Platforms do not show you those numbers, so you have to keep them yourself. A plain spreadsheet works: date, platform, task type, minutes spent, dollars earned. After two weeks, sort by effective hourly rate and stop taking the task types at the bottom.
The other operations lesson is about dependency. In network operations, a single upstream provider is a known risk you design around. A single annotation platform is the same thing. The four cuts above all came from one client decision upstream of the worker. Being active on two or three platforms is not greed; it is redundancy.
Getting in and getting routed to better work
Onboarding is where most people quietly lose, and it is mostly unpaid. A few things improve the odds.
Lead with your strongest credential
Platforms route higher-paying queues to people who signal expertise during signup. If you are a registered nurse, a CPA, a paralegal, or a working developer, say so plainly and pick the domain track, even if a generalist track looks easier to pass. The pay gap between tiers is far larger than the difficulty gap.
Treat the assessment like a job interview
Read the rubric twice. Study every worked example they show you. Most rejections come from ignoring a formatting rule or a style guideline, not from wrong answers. A strong first score also tends to place you in better task pools from the start.
Protect your quality score early
Your first 100 or so tasks carry outsized weight. Accuracy scores below a platform’s threshold can end your access without much warning, and appeals are slow. Go slower than feels necessary in week one.
Budget for the dry weeks
Assume some weeks will have no work at all. If you are counting on this income for a bill, it is the wrong income for that bill.
Where the work can lead
The most valuable thing you get from rating thousands of AI responses is not the hourly pay. It is a detailed, specific sense of where models fail in your field: which questions they get confidently wrong, which formats they mangle, which instructions they ignore.
That knowledge sells at much higher rates outside the annotation queue. Small businesses pay for people who can set up AI tools that do not embarrass them, which is the core of an AI automation agency side hustle. If you notice the same workflow problem over and over in your domain, that is a product idea, and some people now package AI workflows as products for exactly that reason.
The people who come out ahead are using annotation the way an apprentice uses a first job: paid training with a clear exit, not a permanent seat in a queue someone else controls.