Linguistic Comparisons of Black Box Models

Modern artificial intelligence (AI) models often contain millions, or even trillions, of parameters. These trained models frequently operate as black boxes (BBs). They receive inputs and produce outputs without users understanding the inner workings of the models. Human brains are sufficiently complex that, instead of predicting behavior from detailed knowledge of a person's neural structure, we observe past actions and form behavioral descriptions. For example, without understanding the inner workings of Bob's brain, we might still learn that “Bob likes to eat much more ice cream than Alice when the weather is hot and sunny.” Herein, we present a new process for generating comparative linguistic summaries to evaluate BB models. This capability holds significance in tasks like explainable AI, objectively identifying discrepancies and similarities between AI models, and self-improving closed-loop AI.