In-context learning is the ability of a large language model to pick up a new task simply from examples or instructions placed in its prompt, with no retraining and no change to its internal weights. Show the model a few input-output pairs, and it infers the pattern and applies it to the next input.
This behavior was named and demonstrated in OpenAI’s 2020 GPT-3 paper, “Language Models are Few-Shot Learners” (Brown et al.). The paper showed that a sufficiently large model could perform tasks “few-shot” — from a handful of in-prompt examples — often rivaling models specifically trained for those tasks. It was a surprising, emergent property of scale that reframed how people use these models: by writing prompts rather than retraining.
In-context learning is why prompt engineering exists. The model’s general knowledge is fixed, but how you frame the request inside the context window can dramatically change results.
Why business readers should care: In-context learning means a single deployed model can be steered to new tasks instantly through prompts, often with no engineering project at all — a fast, cheap path to value, but one bounded by what fits in the prompt.