RLHF (Reinforcement Learning from Human Feedback)
A training technique where human evaluators rank AI outputs, and the model learns to produce responses humans prefer. Used to align models like ChatGPT with human values, making them more helpful and less harmful.
Why it matters
RLHF is the step that turned raw language models into helpful assistants that follow instructions and stay polite. Before it, models could complete text but often ignored what you actually wanted or produced unsafe replies. By learning from human preferences, they became far more useful and better behaved. Understanding RLHF explains why assistants like ChatGPT feel cooperative, and also why they sometimes hedge or refuse in ways that reflect their trainers' choices.
A concrete example
During training, people are shown two AI answers to the same question and pick the better one, over and over. The model gradually learns to favor responses humans prefer, like clear, honest, and non-toxic ones. A practical result you feel: when you ask for help rewriting an email and get a genuinely useful draft rather than random text, that cooperativeness largely comes from RLHF.
How to use it
This explains a lot of everyday model behaviour: why models are agreeable, why they hedge, why they apologise when challenged even when they were right. All of that was rewarded during training. Knowing it lets you compensate — asking a model to argue against a position, or to list what would make an answer wrong, gets past the trained tendency to agree with you. The technique is described in Training language models to follow instructions with human feedback.
The common mistake
Reading agreement as confirmation. A model that changes its answer because you pushed back has not re-examined the evidence; it has done the thing it was trained to do when a user expresses disagreement.
Related terms
AI Alignment
The research challenge of ensuring AI systems pursue goals that are beneficial to humans. Misaligned AI could technically achieve its objective while causing unintended harm. Alignment research aims to make AI reliably helpful, harmless, and honest.
Constitutional AI
A training approach where AI models are given a set of principles (a 'constitution') and learn to self-critique and revise their outputs to comply with those principles. Reduces reliance on human feedback for safety alignment.
Hallucination
When an AI model generates information that sounds plausible but is factually incorrect or entirely fabricated. Common with statistics, citations, and historical claims.
AI Dubbing
Automated translation and re-voicing of audio/video content into other languages while preserving the original speaker's voice characteristics, timing, and emotional delivery.
AI Orchestration
Coordinating multiple AI models, tools, and data sources in a unified pipeline. An orchestration layer manages prompt routing, context passing, error handling, and output aggregation across different AI services.
API (Application Programming Interface)
A way for software applications to communicate with each other. AI APIs let developers integrate AI capabilities into their own applications programmatically.