Revisiting Explainable AI through Model-Independent Concept Dictionaries

Explainability methods need to work across different AI models: Traditional explainability techniques are often tied to specific model architectures, making it difficult to apply consistent standards across organizations. DictXAI addresses this by using interpretable concepts defined in the input domain rather than relying on model-specific internal components, enabling more standardized explanations regardless of which AI system is being used.

Current AI explanations often hide the real problems: Existing explainability methods assume that basic input features (like individual pixels) are already understandable to humans, which they often aren't. DictXAI instead uses "dictionaries" of meaningful, predefined concepts to explain decisions, which can reveal actual data artifacts and biases—such as identifying when an AI model relies on irrelevant patterns rather than genuine medical indicators (a phenomenon called "Clever Hans effects").

Better explanations can catch AI failures before deployment: The paper demonstrates that DictXAI can identify when AI systems malfunction by tracing predictions back to specific, identifiable patterns in the data. This actionable insight means organizations can potentially catch and fix problematic AI behavior during development rather than after deployment causes real-world harm.

Explainability methods should be flexible enough for different domains: DictXAI works with various types of interpretable concepts—from learned image patterns to standardized medical waveforms—showing that a single explainability approach can be adapted across different fields and data types, making responsible AI practices more scalable.

Better explanations improve human-AI collaboration: The research shows that DictXAI explanations help humans and AI systems align on complex decisions, particularly in high-stakes domains like biomedical analysis, which is essential for responsible deployment in sensitive applications.