The objective of the proposed research is to develop a unified framework for evaluating and improving the bidirectional flow of information between humans and AI systems by treating both explanations and prompts through the lens of the Information Bottleneck principle. The core idea is that human AI interaction is a communication problem in two directions. In the model to human direction, explanations must compress the model’s internal reasoning into a form that preserves what a person needs to judge, trust, or correct the output. In the human to model direction, prompts must compress the user’s intent into a form that is sufficient to guide model behavior without adding redundant or conflicting signal. This proposal argues that necessity and sufficiency are the right principles for analyzing both directions under one theory. A representation is necessary if removing it changes the outcome, and sufficient if providing it can produce the desired outcome. Building on this view, the research first develops causal, interventional measures of necessity and sufficiency and uses them to study existing explanation methods, showing in preliminary tabular experiments that highly ranked features are not always both necessary and sufficient. It then extends the framework to image models, where it investigates whether standard evaluation metrics truly measure information that survives the model’s internal compression. Finally, it applies the same framework to interactive segmentation, where prompts are analyzed as units of information transfer, classified by their contribution, and selected actively through predictive disagreement using a Bayesian extension of SAM. Together, the proposed work aims to establish a single information theoretic account of how AI systems should communicate with humans and how humans should communicate with AI systems for effective collaborative decision making.