RLHF: How AI Gets Its Common Sense

You've seen the chatbots that hallucinate — the ones that invent a fake lawyer, a fake law, or a fake auntie's name. RLHF is the fix. It stands for Reinforcement Learning from Human Feedback. Here's how it actually works, in plain terms.

First, the AI gets trained on a massive pile of text — books, forums, recipes, tweets. It learns patterns. But it has no real sense of what's helpful, safe, or true. It's like a kid who memorized the entire bodega menu but has never cooked a single dish.

So humans step in. They're called annotators — usually contractors from the Philippines, Kenya, or Latin America, paid a few dollars an hour. They read the AI's responses and rank them. Which one was more accurate? Which one was less harmful? Which one actually helped the user? Those rankings become the AI's