Inference
The moment an already-trained AI actually does its job: you give it a prompt, and it produces an answer. If training is the long process of the model going to school, inference is it turning up to work and applying what it learned. Every response you get from an AI tool is an act of inference.
Inference is the step where a finished AI model takes your input and generates a result. It is worth separating from training, which is the earlier, one-off process of building the model in the first place. Training is slow, expensive, and done by the company that makes the model. Inference happens every single time anyone uses it.
The school analogy holds up well here. Training is the years of study that give a model its abilities. Inference is the exam question you hand it afterwards. The model is not learning anything new from your prompt during inference. It is applying what it already knows to whatever you have asked.
This distinction quietly explains a few things you may notice. It is why an AI does not remember your previous conversations unless the tool is built to feed them back in, and it is why a model cannot know about events after its knowledge cutoff. Its learning finished at training time; inference only uses what is already there.
Inference also costs money and energy each time it runs, which is why some tools have usage limits or paid tiers. Knowing the word helps you make sense of pricing pages and technical articles, where “inference cost” means the cost of running the model to answer people.