A philosophical investigation into the limits of empirical projection, algorithmic over-fitting, and the linguistic structures that govern our expectations of reality. 6 mins read.
Imagine an emerald miner who has spent a lifetime pulling green gems from the deep earth. Every single specimen unearthed has been deep green. Relying on classical inductive reasoning, the miner confidently asserts: "All emeralds are green." This prediction feels safe, intuitive, and scientifically sound. It is backed by an unblemished record of past observations.
Now enters the philosopher Nelson Goodman, offering a quiet, devastating challenge. He introduces a new color predicate: grue. An object is grue if it is observed before a specific future time T and is green, or if it is not observed before T and is blue.
If we set the arbitrary time T to midnight tonight, every emerald observed up to this second satisfies both descriptions: they are green, and they are also grue. Consequently, our entire history of empirical data supports the prediction that all emeralds are green, but it also perfectly supports the prediction that tomorrow's emeralds will be grue—meaning they will suddenly appear blue. We are left with two radically different predictions about tomorrow, both perfectly justified by the exact same historical evidence.
The modern algorithmic landscape mirrors this exact vulnerability. When machine learning models over-fit to historical data, they are essentially inventing complex, grue-like predicates to satisfy past inputs, only to collapse when exposed to the unmapped future.
The Collapse of Syntactic Induction
Before Goodman published his landmark work in 1954, philosophers of science assumed that inductive logic could be formalized just like deductive logic. They believed that the validity of an inductive argument depended solely on its grammatical and logical structure. If "Object 1 is an A and is B" is repeated enough times, it structurally supports "All As are Bs."
Goodman shattered this assumption. The logical form of the argument for "all emeralds are green" is identical to the argument for "all emeralds are grue." Yet, we instinctively reject "grue" as a ridiculous, artificial concept. This reveals that the validity of induction does not rest on pure logical syntax. There is a deep, unwritten rule governing which predicates we allow ourselves to project into the future.
| Predicate Type | Example | Projectibility Status | Why We Accept/Reject |
|---|---|---|---|
| Natural / Entrenched | Green, Conducts Electricity | Projectible | Deeply woven into historical, linguistic, and scientific usage. |
| Pathological / Time-Dependent | Grue, Bleen | Non-Projectible | Lacks historical entrenchment; relies on arbitrary temporal boundaries. |
The Concept of Entrenchment
How do we rescue science from this infinite sea of bizarre, competing predictions? Goodman’s own solution was pragmatic rather than purely logical. He introduced the theory of entrenchment.
Some words and concepts are highly entrenched in our language because we have used them successfully for generations. "Green" is a projectible predicate because it has a long, robust history of projection. "Grue" is a non-projectible linguistic interloper with no historical track record. We do not choose "green" because it is inherently more real; we choose it because our linguistic community has collectively agreed to run its cognitive machinery on those tracks.
Referenced Works & Texts
- Nelson Goodman, Fact, Fiction, and Forecast, Chapter III: "The New Riddle of Induction" (1954). Explores the breakdown of confirmation theories.
- Rudolf Carnap, Logical Foundations of Probability, Chapter IV (1950). Discusses the formal semantic frameworks that Goodman sought to challenge.
If you found this valuable, consider supporting our work.
Join PhiloCrux community.
Unlock high-density masterclasses and investigations into ideas surviving outside the algorithmic consensus. Support independent thought and get full access to our digital library.
Join Now