The Projection Problem
Introduction
The projection problem, widely known as Goodman's New Riddle of Induction, is a foundational issue in epistemology and the philosophy of science. It challenges the assumption that past observations reliably justify future predictions by demonstrating that multiple, mutually contradictory hypotheses can be equally supported by the same empirical data[1].
Core Concept
Unlike Hume's classic problem of induction—which questions whether we can justify the principle that the future will resemble the past—the projection problem asks: given that induction works, how do we know which generalizations are legitimate to project forward in time?
Formulated by American philosopher Nelson Goodman in his 1954 work Fact, Fiction, and Forecast, the problem exposes a structural ambiguity in scientific reasoning. It forces us to distinguish between law-like generalizations (those that support predictive reasoning) and accidental generalizations (those that do not)[2].
Historical Context
Inductive reasoning has long been the backbone of scientific inquiry. From Francis Bacon's empirical methods to John Stuart Mill's system of experimental canons, scientists assumed that repeated observation naturally accumulates toward truth. David Hume famously dismantled this confidence by showing that induction cannot be justified deductively without circularity[3].
Goodman accepted Hume's conclusion that induction lacks a strict logical foundation. However, he argued that science still functions because practitioners implicitly rely on a concept he termed projectibility—the property of predicates that makes them suitable for inductive projection. The projection problem arises when we discover that projectibility is far more ambiguous than previously assumed[2].
The 'Grue' Predicate
To illustrate the problem, Goodman introduced the fictional predicate grue. An object is defined as grue if it is observed before time t and is green, or if it is not observed before time t and is blue. For simplicity, let t be January 1, 2100[4].
"All emeralds examined before January 1, 2100 are green."
Nelson Goodman, Fact, Fiction, and Forecast (1954)
This observation equally supports two contradictory hypotheses:
- All emeralds are green.
- All emeralds are grue.
Both hypotheses are consistent with every observation made prior to 2100. Yet after 2100, they yield opposite predictions: the first expects emeralds to remain green, while the second expects them to turn blue. The projection problem asks why we confidently accept the first while dismissing the second, despite identical empirical support[1].
Philosophical Implications
Induction vs. Deduction
Deductive reasoning guarantees truth preservation: if premises are true, the conclusion must be true. Induction offers no such guarantee; it only provides probabilistic support. The projection problem reveals that induction's weakness isn't merely probabilistic—it's structurally underdetermined. The same data can project in infinitely many directions unless we impose linguistic and conceptual constraints[5].
Scientific Laws & Projectibility
Goodman's solution hinged on distinguishing entrenched predicates (those historically successful in scientific practice, like "green") from non-entrenched ones (like "grue"). Projectibility, he argued, is not a logical property but a pragmatic one rooted in human cognitive habits and linguistic conventions[2].
| Feature | Law-like Generalization | Accidental Generalization |
|---|---|---|
| Supports counterfactuals? | Yes | No |
| Predicate projectibility | Entrenched / Natural | Disjunctive / Gerrymandered |
| Example | All metals conduct electricity | All coins in my pocket are copper |
Modern Perspectives
In contemporary philosophy and computer science, the projection problem has found new relevance. Machine learning algorithms, particularly those relying on pattern recognition and statistical inference, face a computational analogue: overfitting and distribution shift. A model trained on biased or narrow data may "project" false patterns to novel contexts, mirroring Goodman's grue paradox[6].
Researchers in causal inference (e.g., Judea Pearl) and robust AI argue that solving the projection problem requires moving beyond correlation toward mechanistic understanding. By identifying invariant causal structures rather than surface-level regularities, systems can distinguish which patterns are genuinely projectible across time and domains[7].
Furthermore, linguistic philosophers have explored how natural language itself filters projectibility. Cognitive linguistics suggests that human perception naturally segments reality into natural kinds (water, gold, emerald) that align with evolutionary and physical stability, implicitly solving Goodman's riddle through embodied cognition rather than pure logic[8].
See Also
Problem of Induction · Underdetermination · Natural Kinds · Overfitting in Machine Learning · Causal Inference
References
- Goodman, N. (1954). Fact, Fiction, and Forecast. Harvard University Press.
- Salmon, W. C. (1963). "A Decision-Theoretic Approach to Projectibility." Philosophy of Science, 30(2), 107–115.
- Hume, D. (1739/1978). A Treatise of Human Nature. Oxford University Press.
- Quine, W. V. O. (1959). "Set Theory and Its Logic." Harvard Review, 3, 23–41.
- Carnap, R. (1950). Logical Foundations of Probability. University of Chicago Press.
- Geirhos, R., et al. (2020). "Shortcut Learning in Deep Neural Networks." Nature Machine Intelligence, 2, 665–673.
- Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.
- Rosch, E. (1978). "Principles of Categorization." In Cognition and Categorization. Erlbaum.