QuestionsXXII
Why trust that the future will be like the past?
The sun has risen every day of recorded history. Hume asked a question no one has fully answered: what makes that a reason to expect it tomorrow?
- Reading time
- 4 min
- Difficulty
- Intermediate
- Branch
- Knowledge
- First posed
- David Hume, A Treatise of Human Nature, 1739, and An Enquiry Concerning Human Understanding, 1748
The question
Every piece of bread you have eaten has nourished you. So you expect the next piece will. Every time you have dropped a glass, it fell. So you expect the next one will. This kind of reasoning, from observed cases to unobserved ones, is called induction. Science depends on it. So does every ordinary expectation you have.
In 1739 David Hume asked what justifies it. It is not logic: there is no contradiction in supposing that the next piece of bread will poison you, so the conclusion does not follow from the premises with logical necessity. Then it must be experience. But the only experience we have is that induction has worked in the past. To conclude that it will work in the future is to use induction to justify induction, which is arguing in a circle.
Hume concluded that our inductive expectations are not founded on reason at all. They come from custom or habit: after seeing one thing follow another many times, the mind simply comes to expect it. This works well for us. But it is not, Hume argued, a rational justification.
“Custom, then, is the great guide of human life.”
David Hume, An Enquiry Concerning Human Understanding (1748)
Why it matters
If Hume is right, the foundations of science are not purely rational. Every scientific law is a generalisation from finite observations, and no finite set of observations logically guarantees the next one.
The problem also bears on how we reason under uncertainty everywhere: in medicine, in law, in machine learning. Systems that learn from data are, in effect, automated inductive reasoners, and they face the same question of why past patterns should hold.
Key terms
- InductionReasoning from observed cases to unobserved ones, such as from past experience to the future. Its conclusions are probable at best, never guaranteed.
- DeductionReasoning in which, if the premises are true, the conclusion must be true. Contrasted with induction, which only makes its conclusion probable.
- Uniformity of natureThe assumption that the future will resemble the past and unobserved cases will resemble observed ones. Hume argued it cannot be justified without circularity.
- FalsifiabilityKarl Popper's criterion for a scientific theory: it must make predictions that could, in principle, turn out false.
- BayesianismThe view that rational belief comes in degrees, measured as probabilities, and that we should update those degrees on new evidence according to Bayes' theorem.
The positions
Induction cannot be justified
Hume was right. There is no non-circular rational justification of induction. We rely on it from habit.
Science does not need induction
Science proceeds by conjecture and refutation, not induction. Theories are never confirmed, only tested and possibly falsified.
Induction is rational by definition
Asking whether induction is reasonable is like asking whether the law is legal. Being reasonable about the unobserved simply means reasoning inductively.
Probability can help
Bayesian probability theory explains how evidence should change our confidence, and shows that inductive reasoning is coherent.
Arguments and objections
Induction cannot be justified
- Best argument
- Every proposed justification either assumes the uniformity of nature, the principle that the future resembles the past, which is itself an inductive conclusion, or appeals to the past success of induction, which is circular. No one has found a way out in nearly three centuries.
- Strongest objection
- If this is true, there is no rational difference between expecting the sun to rise and expecting it not to. That seems absurd, and suggests that the standard of 'justification' being demanded is too high.
Science does not need induction
- Best argument
- Karl Popper argued that while no number of white swans proves 'all swans are white', a single black swan disproves it. Falsification uses only deduction. Science advances by proposing bold theories and trying hard to refute them. Those that survive are not proven, merely not yet refuted.
- Strongest objection
- We still have to decide which surviving theory to rely on when building a bridge or taking a medicine, and choosing the best-tested one seems to involve exactly the inductive confidence Popper wanted to avoid.
Induction is rational by definition
- Best argument
- P. F. Strawson argued that induction is part of what we mean by 'rational belief' about the unobserved. There is no higher standard against which induction could be judged and found wanting. Demanding a deductive justification for induction is a category mistake.
- Strongest objection
- This seems to dissolve the problem by definition rather than solve it. We can still coherently ask whether a method of reasoning that we call 'rational' actually tends to lead to the truth.
Probability can help
- Best argument
- Bayesians model belief as degrees of confidence updated by evidence according to the laws of probability. Over time, agents who update this way converge on the truth in a wide range of cases. This does not prove the future will resemble the past, but it shows how to reason well given that you start with some reasonable expectations.
- Strongest objection
- Everything depends on the starting probabilities, the priors. Someone who starts out expecting nature to be irregular will draw very different conclusions from the same evidence. Bayesianism tells you how to update, but not where to begin.
Who argued what
- Induction cannot be justified David Hume
- Science does not need induction Karl Popper
- Induction is rational by definition P. F. Strawson
- Probability can help Thomas Bayes, Frank Ramsey, Bruno de Finetti
Variations
Russell's chicken
Bertrand Russell described a chicken that is fed every morning by the farmer. Each day confirms its expectation that it will be fed tomorrow. Then one morning the farmer wrings its neck. 'More refined views as to the uniformity of nature would have been useful to the chicken.'
Grue
Nelson Goodman invented the word 'grue': green if observed before some future date, and blue thereafter. Every emerald ever observed is both green and grue. So why do we expect the next emerald to be green rather than grue? Goodman's 'new riddle of induction' shows that the problem is not only whether to trust patterns, but which patterns to trust.
Go deeper
Machine learning faces the same problem
A model trained on past data assumes that the future will resemble its training set. When the world shifts, a pandemic, a new market, a new kind of input, performance can collapse. Computer scientists call this distribution shift. The 'no free lunch' theorems of machine learning formalise a version of Hume's point: averaged across all possible worlds, no learning method outperforms any other. Learning only works because we assume our world has some particular regularities.
Schools of thought: Empiricism
Related ideas
Further reading and sources
- EncyclopediaStanford Encyclopedia of Philosophy: The Problem of Induction
- WikipediaProblem of induction
- BookAn Enquiry Concerning Human Understanding, David Hume
- BookThe Problems of Philosophy, Bertrand Russell
- BookFact, Fiction, and Forecast, Nelson Goodman
Open questions
- What, if anything, makes it reasonable to expect the sun to rise tomorrow?
- If induction cannot be justified, should that change how you act?
- Why green, and not grue?