Explainability in JupyterLab and Beyond: Interactive XAI Systems for Integrated and Collaborative Workflows

XAI tools are often disconnected from where data scientists actually work: The paper identifies a significant gap—while most machine learning development happens in Python environments like JupyterLab and Jupyter Notebook, explainability tools are typically built as separate standalone interfaces, creating friction in real-world workflows.

Three practical design patterns can embed explainability directly into notebooks: The authors propose specific technical approaches (one-way communication, two-way data synchronization, and bi-directional callbacks) that allow interactive explanation tools to be integrated seamlessly into the computational environments where ML practitioners already spend their time.

Integration supports more collaborative and human-centered AI development: By embedding XAI tools into existing workflows rather than requiring separate applications, the paper argues that teams can more easily incorporate diverse perspectives and user needs into model development—a key principle of responsible AI.

Open-source toolkit (bonXAI) provides a concrete starting point: The authors released an open-source toolkit demonstrating these design patterns in action with a PyTorch text classification example, making it practical for other developers to build similar integrated explainability tools rather than starting from scratch.

Better tool integration can democratize access to model explanations: By making XAI tools more accessible within familiar computational environments, these approaches could help more practitioners—not just AI specialists—understand and scrutinize how machine learning models make decisions.