All Explanations are Wrong, But Many Are Useful: Exploring the Rashomon Explanation Set with Large Language Models

Multiple valid explanations can coexist without sacrificing accuracy: The paper challenges the common assumption that explaining AI models requires trading off accuracy for interpretability. Instead, the researchers found that building a set of multiple faithful explanations (called the "Rashomon Explanation Set") can actually improve model performance rather than degrade it, suggesting the accuracy-explainability trade-off may be a false choice.

Explanation quality directly impacts prediction performance: The research demonstrates that explanation fidelity—how well explanations actually guide the model's predictions—acts as a measurable bound on model performance. This means investing in better explanations isn't just good for transparency; it's fundamentally tied to making better predictions across real-world applications like customer churn, clinical outcomes, and advertising.

Large language models can generate trustworthy natural language explanations at scale: RashomonLLM uses AI to produce explanations in plain language that align with actual model decisions, tested on large-scale real-world datasets. This suggests that consumer-facing explanations don't need to be overly technical or limited to simple rules—they can be more natural and nuanced while remaining faithful to how the model actually works.

Improved performance holds up under real-world conditions: The gains in both accuracy and explanation quality remained robust across distribution shifts, temporal changes, and different random seeds, indicating this approach isn't just theoretically sound but practically reliable for deployment in dynamic business environments.