RLHF (Reinforcement Learning from Human Feedback)
RLHF trains an AI model on people's ratings of its answers, so it learns to give the kind of reply people prefer.
RLHF, short for reinforcement learning from human feedback, is a way of training an AI model using people's judgements. People compare the model's answers and say which they prefer. The model is then trained to give more answers like the preferred ones.
It's a large part of why chat assistants sound the way they do: helpful, polite, balanced, and sure of themselves. That tone affects how the platforms talk about your brand.
How it works
First, a model learns from a large amount of text, which teaches it language and facts but not manners. Then people are shown pairs of its answers to the same question and pick the better one. Those choices are used to train a second model that scores answers the way the people would. Finally, the main model is trained to get high scores.
OpenAI described the method in its 2022 paper on InstructGPT, and most chat assistants since have been trained in a similar way.
What it means for brands
RLHF teaches models what a good answer looks like to the people who rated them. A few habits follow from that.
Answers tend to sound balanced. Ask for "the best" and you'll often get several options with pros and cons, rather than one name. Being one of the few brands named matters more than being called the best.
Answers tend to sound confident, including when they're wrong. A model trained to sound helpful may state an old price as plainly as a current one.
Answers can lean towards the question. Ask "Why is Fernhill Coffee so expensive?" and the answer may simply go along with the idea, without asking whether it's true. So the way people phrase questions about you shapes what they hear.