"Generative AI” is not called that because it generates new [novel] content - quite the opposite, actually.
The word "generative" simply describes the process of how a statistical model arrives at predictions. That is it selects ("generates") the best matching values from its training data, given the user's input (aka prompt).
In general there are discriminative and generative models.
Discriminative models find patterns in the training data, and compute the output by essentially looking up the rule that relates the output. E.g. input is an image of a cat, the model finds the "cat pattern" and outputs "cat".
Discriminative: Think "rule book" - look up the rule that applies, act accordingly.
Generative models also find patterns in the data, albeit in a different way. Instead of looking up a rule directly relating the output, it looks up the part of the training data (or rather, a statistical summary thereof) that matches the input, and then outputs the most likely next value from that. E.g. the input is "where is New York?", the model pattern-matches data containing "location" and "city" data and outputs the respective value for New York, say "USA".
Generative: Think "index in a book" - lookup the page, read what's there.
"Generative AI” is not called that because it generates new [novel] content - quite the opposite, actually.
The word "generative" simply describes the process of how a statistical model arrives at predictions. That is it selects ("generates") the best matching values from its training data, given the user's input (aka prompt).
In general there are discriminative and generative models.
Discriminative models find patterns in the training data, and compute the output by essentially looking up the rule that relates the output. E.g. input is an image of a cat, the model finds the "cat pattern" and outputs "cat".
Discriminative: Think "rule book" - look up the rule that applies, act accordingly.
Generative models also find patterns in the data, albeit in a different way. Instead of looking up a rule directly relating the output, it looks up the part of the training data (or rather, a statistical summary thereof) that matches the input, and then outputs the most likely next value from that. E.g. the input is "where is New York?", the model pattern-matches data containing "location" and "city" data and outputs the respective value for New York, say "USA".
Generative: Think "index in a book" - lookup the page, read what's there.