What is RLHF? Reinforcement learning from human feedback, explained simply

RLHF (reinforcement learning from human feedback) is a way of training an AI model using people's judgments. Humans compare pairs of model answers and pick the better one; those choices train a "reward model" that predicts what people prefer; then the main model is tuned to produce answers the reward model scores highly. The human comparisons are the part companies pay AI trainers to do.

Updated October 28, 2026 4 min read Numbers from 1,815 open listingsHow we check

On this page 8 sections

Why isn't training on the internet enough?
How does RLHF work, step by step?
What does an RLHF task look like in practice?
Why do experts get paid more for RLHF work?
What has changed since the original RLHF papers?
How do I get RLHF work?
New AI training jobs
Other questions people ask

Why isn't training on the internet enough?

A language model's first stage of training teaches it to predict the next word in huge amounts of text. That produces a model that can write fluently, but it has no particular reason to be helpful, honest or safe. It will continue a question with another question, or confidently invent a source, because that is also what text on the internet does.

RLHF is one of the steps that turns that raw model into an assistant. OpenAI's InstructGPT paper reported that people preferred outputs from a 1.3 billion parameter model trained with human feedback over the original 175 billion parameter GPT-3 (Ouyang et al., 2022). Human judgment did more than a 100x larger model.

How does RLHF work, step by step?

The three stages of RLHF and where human workers come in
Stage	What happens	The human job
1. Demonstrations (SFT)	The model is fine-tuned on examples of ideal answers.	Write prompts and model-quality answers to them.
2. Comparisons	The model writes several answers to a prompt. People rank them. A reward model learns to predict those rankings.	Compare answers side by side, rate them on a rubric, explain the choice.
3. Optimization	The model is trained to produce answers the reward model scores highly, while staying close to its earlier behavior.	Mostly automated. Humans check the results and find new failures, which feeds back into stage 2.

A simple way to picture it: you are not teaching the model facts. You are teaching it taste. Which of these two answers would a careful expert rather receive?

What does an RLHF task look like in practice?

A typical comparison task on a platform like Outlier or Mercor goes like this:

You see a prompt, for example a user asking how to split rent fairly between roommates with different-sized rooms.
You see two model responses, A and B.
You rate each one on several dimensions the project defines, often accuracy, instruction-following, completeness, safety and writing quality, usually on a 1–5 or 1–7 scale.
You pick an overall winner and say how strongly you prefer it.
You write a short justification that names specific problems: "B's math is wrong in step 2; A ignores the shared-bathroom detail the user gave."

The justification matters most. Reviewers check whether your reasoning is specific and correct, and vague notes like "A is more helpful" are the fastest way to fail a quality review.

Why do experts get paid more for RLHF work?

Early RLHF needed generalists to catch obvious problems: rude tone, ignored instructions, made-up facts. Current models rarely make those mistakes. What they still get wrong is subtle: a drug dose that is off for a kidney-impaired patient, a contract clause that wouldn't hold up in a specific jurisdiction, code that passes the obvious test and fails on an edge case.

Only someone who knows the field can catch those, so labs pay accordingly. That is why platforms like Mercor and micro1 recruit doctors, lawyers and engineers, and why the top of the pay range sits in coding, medicine, law and finance.

What has changed since the original RLHF papers?

The idea of learning from human preference comparisons goes back to research on teaching agents from human feedback in 2017 (Christiano et al.). Since then:

DPO and similar methods skip the separate reward model and train directly on the preference pairs (Rafailov et al., 2023). The human comparison data is still needed.
AI feedback now handles some easy judgments, so human work has shifted toward harder, expert-level tasks and toward writing rubrics that automated graders follow.
Agent and coding tasks ask humans to judge multi-step work: did the model use the tools correctly, and did the final code actually run?

How do I get RLHF work?

Almost every platform on Rubric is doing some form of RLHF or preference data. Start with what AI training work is, then pick a platform that matches your background on Find my platform, or filter open jobs by your field. New listings land in new this week first.