Model extraction aims to create a substitute model that functionally replicates the target model's behavior. This is done by systematically querying the API and training a new model on the observed input-output pairs.