A buzzword - generally I would say it refers to using unsupervised training of some neural network model and then generating new data in the domain from the model.
I believe it originally grew out of the phrase "generative model" (which models the joint data probability distribution) but most of the prominent 'generative AI' models (like GPT) are not actually generative models but discriminators.
Regardless, outside of simple things like matrix multiplication, the theory of backpropagation (although not the implementations), language modeling as a concept, etc. - almost all of these techniques are within the last decade and a half or so.
It is, unfortunately, not the same thing as a generative model.