By Mufy Pachorawala, founder of getAIwork · Last updated: September 3, 2026
Key facts
- The job is comparison, not creation. Most RLHF tasks ask which of these answers is better and why, rather than asking you to produce something from scratch
- The written justification is the product. A correct ranking with a vague reason is close to worthless to the buyer
- No machine learning knowledge is required for the human-feedback side. Subject expertise plus disciplined writing is the actual requirement
- Hired through platforms, rarely directly. Expert networks and data vendors sit between the labs and the people doing the work
- 591 of 1,435 live listings on the getAIwork board are platform programmes as of September 3, 2026, the category that carries most RLHF work
- 391 live listings are tagged writing-centred, which is the skill family most RLHF work sits in
What is RLHF, in plain terms?
A language model trained only on text is fluent but has no opinion about what a good answer looks like. Reinforcement learning from human feedback is the stage that gives it one. Humans compare candidate answers, a reward model learns to predict those human preferences, and the language model is then tuned to produce answers that score well against it.
Everything that makes a modern assistant feel usable rather than merely fluent came from that loop: refusing the request it should refuse, admitting uncertainty, showing its working, not padding four paragraphs where one would do. Somebody had to demonstrate each of those preferences, thousands of times, in writing.
Which means the job title is slightly misleading. Almost nobody in an RLHF job does reinforcement learning. You are supplying the human preferences that a reinforcement learning process consumes. Understanding that distinction is genuinely useful in an interview, because it is the difference between a candidate who read the acronym and one who understands the pipeline.
What you actually do all day
| Task type | What it looks like | What it is testing |
|---|---|---|
| Pairwise comparison | Two answers to the same prompt. Pick the better one and write why | Your judgement, and whether you can articulate it |
| Ranking | Four or five candidate answers, ordered best to worst | Consistency across finer distinctions |
| Reference answer writing | Write the answer the model should have given | Your actual subject expertise |
| Rubric scoring | Score along named dimensions such as accuracy, helpfulness, safety | Whether you can apply someone else’s standard rather than your own |
| Error identification | Find and describe the specific flaw in a response | Precision. Vague criticism fails these |
| Red teaming | Try to make the model produce something it should not | Creativity and stamina, more than expertise |
The recurring surprise for new contributors is how much writing is involved. A task that looks like a two-second judgement, A is better than B, actually wants a paragraph explaining the basis for it, and that paragraph is what the buyer is paying for. The ranking alone is nearly free to collect. The reasoning behind it is not.
The second surprise is the guidelines. Every project ships a document, frequently thirty pages, defining what counts as helpful, what counts as unsafe, and how to break ties. Your job is to apply that standard rather than your own taste. People with strong opinions and weak guideline discipline score badly, which is one of the least intuitive facts about this work.
What the work really asks for
Three things, and only one of them is technical.
- Domain expertise. Someone has to know whether the medical answer is actually right. The whole point of paying a human is that the model cannot check itself.
- Written reasoning. Legible, specific, unemotional. If a stranger cannot follow why you preferred one answer, the label is unusable.
- Guideline discipline. Reading a long specification and applying it consistently across hundreds of items, including on the items where you privately disagree with it.
Notice what is not on that list. Python is not on it. Familiarity with transformer architecture is not on it. Those matter for the engineering roles that build the pipeline, which are different jobs with different postings. For the human-feedback work itself, a philosophy graduate who writes precisely frequently outperforms an engineer who does not.
591 live platform programmes on the free getAIwork board, screened and human-approved. Two minutes matches you against them.
Take the 2-minute match quiz →
Who hires for RLHF work
Almost never the labs directly, at least not for the contributor tier. The structure is layered, and knowing the layers saves you applying to the wrong door.
| Layer | Who they are | How you reach them |
|---|---|---|
| Frontier AI labs | The end buyers of the feedback data | Direct roles exist but are research and engineering positions, not contributor work |
| Data vendors | Companies that build and run annotation and evaluation pipelines at scale | They recruit contributors continuously, often under their own platform brand |
| Expert networks | Marketplaces matching credentialled specialists to lab projects | Apply, pass an assessment, wait to be matched to a project |
| Rating and evaluation vendors | Long-established firms from the search-quality world, now doing AI evaluation | Structured hiring with exams, usually capped hours |
Which layer you should target depends entirely on what you already have. A credentialled professional goes to the expert networks, where the credential is the product. Someone strong at writing but without a specialist field goes to the data vendors and rating firms, where the entry route is an assessment rather than a degree. Our list of platforms with open intake sorts them, and the alternatives comparison covers who buys what.
Pay, as listed
We only quote what postings state, and the honest headline is that the range in this category is very wide because it tracks the scarcity of the background rather than the difficulty of the task. On the getAIwork board today, across all 1,435 live listings, the median annual figures listed by posters run from around $83,200 for language work up to $228,800 for listings requiring a licensed profession. Those are figures listed by posters, not earnings anyone is promised, and only about 69 percent of live listings state a rate at all.
The pattern inside that spread is the useful part. A rare credential moves the rate far more than experience with the task does. A physician evaluating medical answers and a generalist evaluating everyday answers are doing structurally similar work at very different prices, because one supply pool is small and the other is not. Our weekly board statistics break this down by background and are refreshed every week.
How to get into it
Four steps, in this order, and the order matters.
- Name your domain. Not AI. The field you already know: tax, oncology, contract law, organic chemistry, Brazilian Portuguese, competitive chess. That is what is being bought.
- Fix your writing before you apply. Most assessments are graded on written justification. Practise explaining, in six sentences, why one of two answers is better, with specific reasons rather than adjectives.
- Apply to two or three platforms in parallel. They are free and non-exclusive. One is fragile, a dozen is thin, two or three is right.
- Treat the guidelines as the exam. On your first project, reread the guideline document after your first ten tasks. Almost all early quality failures are drift rather than incompetence.
The honest downsides
The volume is not steady. Projects end when the client’s project ends. Quiet fortnights are normal and are not a comment on your work.
Some of it is genuinely unpleasant. Safety and red-teaming work means reading material you would not otherwise read. Reputable projects disclose this up front and pay more for it. Read the project description properly before accepting.
It is not a stepping stone to a research job, whatever the marketing implies. It is skilled evaluation work with its own value, and treating it as a back door into a lab leads to disappointment. Skilled evaluators do become lead reviewers and quality managers, which is a real path, just not that one.
The unpaid front-loading is real. Assessments, guideline study and onboarding take hours nobody pays for. Cap it. Four hours per platform without paid work is a reasonable line, after which treat further invitations as marketing rather than opportunity.
1,435 live AI listings as of September 3, 2026, screened and human-approved, filled listings deleted daily.
Frequently asked questions
What are RLHF jobs?
Work supplying the human feedback that reinforcement learning from human feedback depends on: ranking model answers against each other, writing reference answers, scoring against rubrics and explaining the reasoning in writing. The output being bought is your judgement plus your explanation of it.
Do you need a machine learning background for RLHF work?
No, not for the human-feedback side. What it asks for is subject expertise in your own field, clear written reasoning and the discipline to apply a long guideline document consistently. Engineering knowledge matters for the roles that build the pipeline, which are different jobs.
What does RLHF stand for?
Reinforcement learning from human feedback. Humans compare candidate model answers, a reward model learns to predict those preferences, and the language model is tuned to score well against it. The human comparison step is the part people are hired for.
Who hires for RLHF jobs?
Mostly data vendors, expert networks and evaluation firms rather than the AI labs directly. The labs buy the data; the platforms recruit, assess and pay the people producing it. Contributor-level roles are almost always reached through those platforms.
What does RLHF work pay?
It varies enormously with the background required. On the getAIwork board, median annual figures listed by posters range from about $83,200 for language work to $228,800 for listings needing a licensed profession, and only about 69 percent of live listings state a rate. These are listed figures, not earnings anyone is promised.
Is RLHF work beginner-friendly?
Partly. Generalist evaluation projects do accept people without a specialist credential, though they are the most contested tier. Specialist projects require the credential. About 11 percent of live listings on our board are tagged beginner-friendly.
How much writing is involved in RLHF tasks?
More than most people expect. The ranking itself is quick; the written justification behind it is the deliverable, and vague reasoning fails quality review even when the ranking is right. Strong, specific writing is the single most transferable skill in this work.
Can RLHF work lead to a job at an AI lab?
Not directly, and marketing that suggests otherwise is overselling it. The realistic progression inside this category is towards lead reviewer, quality management and project roles on the vendor side, which are genuine careers in their own right.
Mufy Pachorawala · Founder, getAIwork
My AI scans thousands of AI-job posts (40,786 screened so far) and I personally approve every listing before it reaches the board. I write these articles by the same rules: pay quoted only as listed, and no income promises. Read our editorial rules.
Related: Get paid to train AI · AI training platforms hiring now · Data annotation jobs compared · 12 Outlier alternatives compared · AI job market statistics
