The Problem of Induction in Machine Learning

David Hume articulated the problem of induction in 1739: no amount of observed instances logically guarantees that the next instance will follow the same pattern. Machine learning is built entirely on the assumption that it will. Every ML model assumes that the statistical patterns in training data will persist in production data. This assumption has no logical justification, only pragmatic success. When an ML model encounters distribution shift, it encounters Hume’s problem in production.

Why is the problem of induction the deepest vulnerability in machine learning?

Because every ML model makes an inductive leap from past data to future prediction. The leap works until it does not. And no amount of validation on historical data can guarantee it will work on future data, because that guarantee would itself require an inductive argument.

Hume’s problem of induction states that we cannot logically justify the inference from “all observed X have been Y” to “the next X will be Y.” Every ML model makes exactly this inference. A model trained on 10 million customer transactions assumes that the 10,000,001st transaction will follow the same distribution. When the world changes (a pandemic, a policy shift, a market crash), the assumption breaks. The model does not know the world has changed. It keeps predicting based on a past that no longer describes the present.

I deployed a demand forecasting model in early 2020. It was trained on 3 years of historical data. It was accurate to within 4.2% on validation sets. In March 2020, its predictions became meaningless overnight. The model had not failed. The world had changed, and the model’s inductive assumption (that past patterns predict future patterns) was exposed as precisely what Hume said it was: a habit of expectation, not a logical necessity.

This is not merely a technical concern about distribution shift. It is a philosophical limitation on what any data-driven system can know. The model’s confidence in its predictions is mathematical. Its justification for those predictions is, in the deepest sense, philosophical, and philosophically fragile.

Hume’s problem has no solution. But it has a practice: monitor relentlessly, deploy skeptically, and never mistake a model’s statistical confidence for certainty about the future. The future does not owe your training data anything.