Researchers are developing new frameworks to break open AI’s "black box," building interpretability and causal reasoning into models from the start.