Picture a painter who spends years studying thousands of works by the impressionist masters. He does not merely learn to copy them; he grasps their style – how they use colour, how they handle light, how they lay down brushstrokes. Ask him then to paint a portrait of your dog in the impressionist style and he can do it. He creates an entirely new, original work that nonetheless carries all the marks of the style he studied. Generative AI works on a similar principle. Instead of paintings it studies vast quantities of text, images or music from the internet. It can then produce, on request, wholly new content – write a poem, draw a picture or compose music – that is indistinguishable from what a human might have made.
Generative AI (also GenAI) denotes a class of artificial intelligence models that learn from the patterns and structures in training data and are then able to create new, synthetic content autonomously. Unlike discriminative models, which learn only to classify or predict from given inputs (recognising spam, say), generative models produce data that is similar in nature to the training data without being a direct copy of it. Typical architectures include Generative Adversarial Networks (GANs), variational autoencoders (VAEs) and transformers. The output may be text, images, music, code or other forms of complex data that imitate human creativity and knowledge.