# Please Stop Explaining Black Box Models for High Stakes Decisions

**Authors:** Cynthia Rudin  
**arXiv:** [1811.10154](https://arxiv.org/abs/1811.10154)  
**Date:** 2018-11-26  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.10154v1/

## Key points

- Black box models are increasingly used for high-stakes decision-making in society.
- Trying to explain these models after the fact can perpetuate bad practices and potentially cause harm to society.
- Recent work on explainable ML contains critical misconceptions that can negatively impact the widespread use of machine learning models in society.
- There is not necessarily a tradeoff between accuracy and interpretability, particularly when data naturally have a good representation.
- Deep learning models tend to be black boxes because they are highly recursive, but complicated black boxes are not necessary for top predictive performance.
- There tends to be little difference in performance between more complex classifiers and much simpler ones after preprocessing in data science problems where structured data with meaningful features are constructed as part of the process.
- The way forward is to design inherently interpretable models rather than relying on explanations after the fact.
- This will allow policy-makers to demand more effort in ensuring safety and trust in machine learning models for high-stakes decisions while avoiding unnecessary tradeoffs between accuracy and interpretability.

## Summary

The use of black box models for high-stakes decision-making in society has become increasingly prevalent. However, the practice of trying to explain these models after the fact rather than creating inherently interpretable models can perpetuate bad practices and potentially cause catastrophic harm to society. While there is a vast amount of literature on interpretability, much recent work on explainable ML contains critical misconceptions that can have a lasting negative impact on the widespread use of machine learning models in society. One such misconception is that there is necessarily a tradeoff between accuracy and interpretability, which is often not true, particularly when data naturally have a good representation. Deep learning models tend to be black boxes because they are highly recursive, but this does not mean that complicated black boxes are necessary for top predictive performance. In fact, there tends to be little difference in performance between more complex classifiers and much simpler ones after preprocessing in data science problems where structured data with meaningful features are constructed as part of the process. The way forward is to design models that are inherently interpretable rather than relying on explanations after the fact. This will allow policy-makers to demand more effort in ensuring safety and trust in machine learning models for high-stakes decisions while avoiding any unnecessary tradeoffs between accuracy and interpretability.