From recognising to producing
Earlier machine learning mostly *judged* things — is this spam, what number is this. Generative models *make* things. Having learned the statistical shape of their training data, they can write a paragraph, draw an image or draft code that is plausibly like what they saw, yet new. For text this is driven by large language models; for images, by diffusion models.
Under the hood it is still deep learning, most often built on the Transformer architecture. What you get out depends heavily on what you put in, which is why prompt engineering and grounding techniques like retrieval-augmented generation matter so much in practice.
Plausible is not the same as correct
A generative model optimises for output that looks right, not output that is verified true. Treat its confident text as a draft to check, not a fact to trust.