Are Logistic Models Really Interpretable?

Logistic Regression is harder to understand than assumed: A user study found that skilled participants could not reliably predict how logistic regression models would classify new cases, even when given the trained model weights—challenging the common assumption that this "simple" model is inherently interpretable.

Model interpretability matters for real-world use: When AI model weights are published for public use, people need to actually understand how those models work to use them responsibly. The paper demonstrates this gap is a genuine problem, not just a theoretical concern.

A better alternative exists: Linearised Additive Models (LAMs) are proposed as a more understandable alternative that can be applied to existing trained models without retraining them, making them practical to implement.

Improved understanding doesn't sacrifice accuracy: In tests on financial modeling datasets, LAMs achieved significantly better performance on reasoning tasks compared to logistic regression, while maintaining comparable accuracy metrics (ROC-AUC and calibration scores).

Transparency requires rethinking "simple" models: The findings suggest that organizations publishing AI models for trustworthiness and transparency should reconsider whether their chosen models are actually interpretable to their intended users, rather than assuming simplicity equals understandability.