Part of the probleem is how broad the term ai is, and how narrowly it is used. People just mean autoregressor llms and maybe diffusion models, while the term ai is much broader than even machine learning (for instance formal reasoning), which is again broader than backpropagation with gradient descent (for instance boosted trees) which is again broader than generative ai (for instance classifiers and deep learning). All of these are definitely useful in science and engineering and have been for decades, although llms are now beginning to find uses as well.