Reinforcement learning is teaching by consequence rather than by instruction. Instead of being told the right answer, a system tries something, receives a signal about how well that worked, and gradually shifts toward the behaviours that earn better signals. The reward does the teaching. Over many rounds, the system stops needing to be told what to do and starts to work it out, the way a dog learns a trick from treats rather than from a written explanation of the trick.
In modern AI, this is much of how a raw language model is turned into something useful to talk to. After the model has learned to predict text, people rate its responses, and those ratings become the reward signal that pulls it toward answers humans actually find helpful. The technique behind today's chat assistants, reinforcement learning from human feedback, is reinforcement learning with people supplying the sense of better and worse. It is less like programming and more like shaping.
The catch is that a system trained this way learns exactly what the reward rewards, which is not always what its designers meant. If the signal favours answers that sound confident, the system learns to sound confident, whether or not it is right. Get the reward subtly wrong and you get behaviour that is subtly wrong in ways that can be hard to notice. Much of the difficulty in building these systems is not the learning itself but choosing what to reward.
The human echo is hard to miss. People also learn by consequence, and we also become whatever our rewards quietly encourage. A child praised only for being right learns to fear being wrong; an adult rewarded only for speed learns to stop thinking carefully. Whether the learner is made of neurons or numbers, the same uncomfortable lesson holds: you tend to get more of what you reward, not more of what you intended.
