RLHF stands for Reinforcement Learning from Human Feedback. It is the technique that turned raw, statistically-trained language models into the polite, helpful assistants people now talk to every day. Behind it sits a quiet army of contractors who spend their workdays comparing model responses and clicking on the better one.
If that sounds simple, it is — and it is not. The clicks are easy. Judging which response is genuinely better, and articulating why, is the actual skill. Companies pay well for people who can do it consistently.
- RLHF contractors compare model outputs and rank them by quality, safety, or accuracy.
- Pay ranges from about $15–25/hr for general writing tasks to $40–100+/hr for specialist expertise.
- The work is remote, asynchronous, and broadly available outside the US.
- Domain experts (coders, lawyers, doctors, mathematicians, linguists) command the highest rates.
- Major platforms include DataAnnotation, Outlier, Surge AI, and Mercor.
What RLHF actually is (in plain English)
A language model fresh out of pre-training has consumed an enormous amount of text and can predict what word comes next. It does not yet know which of two responses to a user's question is more helpful, more honest, or less harmful. That requires a sense of preference — and preferences live in human heads, not in raw text.
RLHF is the process of extracting those preferences. A human is shown two or more candidate responses to the same prompt. They pick the better one and, often, write a short justification. Their choices become training signal: the model learns to produce responses that resemble the ones humans prefer.
This is what makes the difference between a model that completes your sentence and a model that actually answers your question.
What you actually do all day
Most RLHF tasks fall into one of three shapes.
Preference comparison
You read a prompt, then read two AI-generated responses (sometimes labeled A and B, sometimes more). You pick the one that is better, and you usually write a short note explaining why — "Response B is more accurate; Response A invents a citation that does not exist." Each comparison takes anywhere from one to ten minutes depending on prompt complexity.
Single-response rating
You see one prompt and one response. You rate the response on a Likert-style scale (e.g. 1–7) across multiple dimensions — helpfulness, factuality, safety, instruction-following. Sometimes you also flag specific spans of text as problematic.
Prompt + response writing
Less common, more highly paid. You write a prompt in a target domain (say, a coding challenge or a legal hypothetical), then write the ideal response yourself. The pair is used as a "gold standard" for evaluating model outputs against. This is where domain experts earn premium rates.
Who pays for RLHF and why
The buyers are the major AI labs — OpenAI, Anthropic, Google DeepMind, Meta, and a long tail of smaller model developers. They do not usually hire contractors directly. Instead they outsource to specialist data companies, which run their own contractor platforms.
The four most common ways into this work, in roughly descending order of platform size:
- Outlier — Scale AI's contributor brand. Generalist and specialist tracks, large volume, login-walled feed.
- DataAnnotation — RLHF and evaluation work, primarily writing-focused. Application-based.
- Surge AI — Selective intake, strong pay relative to general annotation work. RLHF and labeling.
- Mercor — Frames the work around credentialed expert contribution. Public job board at
work.mercor.com/explore.
If you want to compare two of the biggest in detail, read our Mercor vs Outlier breakdown.
How much RLHF pays
Pay is the most variable thing about this work, and most published figures online are out of date or platform-specific. The honest ranges, as of 2026:
| Tier | Typical Pay | What It Looks Like |
|---|---|---|
| General Writing | $15–25/hr | Open-domain prompt comparison, response rating. Available to anyone with strong English. |
| Coding | $30–55/hr | Comparing model code, rating bug fixes, judging architectural answers. Often requires demonstrated software experience. |
| STEM & Math | $40–80/hr | Evaluating advanced math, physics, or research-grade scientific reasoning. Usually requires graduate-level background. |
| Domain Expert | $60–150/hr | Medicine, law, finance, engineering. Credentials required. Lower volume, higher per-hour pay. |
| Languages | $20–60/hr | Pay varies dramatically by language — common languages pay less, rare or technical ones pay much more. |
Two important caveats: most platforms pay hourly only for tracked working time, with strict expectations about productivity. And the highest tiers are real but selective — the gating is credentials, calibration, and the ability to write clear justifications, not just task volume.
Geographic eligibility
RLHF work is widely available outside the United States, more than most general gig platforms. The major contributor platforms accept applicants from the EU, UK, India, Latin America, parts of Africa, and most of Asia, though specific tasks are sometimes restricted.
The most common restriction is language eligibility, not country eligibility. A task that asks contributors to evaluate English coding responses will accept fluent English speakers anywhere; a task that requires native Spanish speakers will gate on language regardless of country.
If you are outside the US and tired of resources that pretend you do not exist, this is one of the reasons AITasks.live treats geographic eligibility as a first-class filter on every listing.
What makes a good RLHF contractor
The work rewards a specific cluster of skills more than raw subject knowledge, even at the expert tiers.
- Calibrated judgment. You can recognize that response A is slightly better than response B for three identifiable reasons, even when both are fluent.
- Concise written justification. Two sentences that name the actual flaw. Not "the second one is better written" — but "the second response correctly applies the chain rule; the first drops the negative sign in step three."
- Tolerance for repetition. The hundredth comparison of the day requires the same care as the first.
- Honesty about uncertainty. When you do not know which is correct, you say so rather than guessing confidently. Platforms reward this; they have ways of detecting fake confidence.
- Speed paired with accuracy. Hourly platforms care about throughput. Per-task platforms care about quality. The best contractors are decent at both.
How to start
- Pick one or two platforms to apply to first. Casting a wide net is less effective than getting calibrated on one platform. Outlier and DataAnnotation are the most accessible starting points for generalists; Mercor is the right starting point if you have credentials.
- Take the entry assessments seriously. Most platforms gate access behind a sample evaluation. These determine your rate tier as much as your acceptance. Read each prompt and response carefully and write justifications you would defend to a manager.
- Pick a niche if you have one. A senior software engineer applying as a "generalist contributor" leaves money on the table. The same engineer applying to coding-specific tracks earns 2–3x more for the same work.
- Track your hours and platform mix. Multi-platform contractors typically out-earn single-platform contractors, but only if they manage availability and avoid burnout. Treat it like a real job pipeline.
Common misconceptions
"It is going to be automated away."
The opposite, mostly. As frontier models get smarter, the bar for useful human feedback rises — but so does the value of each high-quality comparison. The work is shifting upward toward specialist evaluation faster than it is shrinking.
"You need an AI background."
You don't. You need clear thinking, careful reading, and a domain skill. Most contributors are subject-matter experts in something else (writers, coders, lawyers, doctors, linguists) who happen to be good at articulating judgments.
"It is the same as data labeling."
Related, but not the same. Data annotation is closer to "draw a box around the cat." RLHF is closer to "explain which of these two essays is more persuasive, and why." The work product is reasoning, not coordinates.
Where AITasks.live fits
The hardest part of RLHF work is not the work itself — it is finding which platforms have active tasks that match your skills and your country of residence, today. That information is scattered across platform feeds (some login-walled), Reddit threads, and Discord servers.
AITasks.live is the directory we wished existed when we were piecing together our own pipeline.