Reinforcement learning from human feedback or RLHF, is a training method where a model's outputs get ranked or rated by human reviewers, and that feedback is used to adjust the model so it produces more of what people rated highly and less of what they rated poorly. Instead of learning purely from a fixed dataset, the model gradually shifts its behavior based on ongoing human judgment.

‍

‍Why it matters

A model can be technically accurate and still produce answers that are unhelpful, off tone, or subtly wrong in a way that's hard to catch with a simple accuracy score. RLHF is what lets a model learn softer qualities like helpfulness, tone, and honesty, things that are much easier for a person to judge by reading a response than to define with a hard coded rule. This is a big part of why RLHF became closely tied to training conversational AI models to actually sound useful rather than just technically correct.

‍

‍How it's done in practice

The typical process involves showing human reviewers several different responses to the same prompt and asking them to rank which one is best, rather than writing a new response from scratch. Those rankings train a separate reward model, which then guides further training of the main model. Because this requires a large volume of careful, consistent human judgment, teams doing RLHF at scale rely heavily on structured reviewer guidelines and calibration, the same underlying idea as calibrating labeling reviewers against each other.‍

‍

Where it's used

This became widely known through the training of large conversational AI models like ChatGPT, where human reviewers ranked different responses to the same prompt so the model could learn which style of answer people actually preferred. It's now standard practice across most major AI labs building chat based assistants, not a one-off technique used by a single company.