CONCEPT

Proximate and Ultimate Causes

Mayr's 1961 distinction between the how of biological mechanism and the why of evolutionary history — two kinds of explanation that physics conflates and biology must keep separate.

In 1961, Ernst Mayr published Cause and Effect in Biology in Science, restructuring the conceptual foundations of an entire discipline with a distinction so simple its profundity was easy to miss. Biology, Mayr argued, requires two kinds of explanation, not one. The proximate question asks how a trait works — the aerodynamics of a wing, the chemistry of pigmentation, the neural circuits that coordinate flight. The ultimate question asks why the trait exists — the evolutionary history of selection pressures that, across millions of years, produced this specific solution to this specific problem. Both questions are legitimate. Both require different methods, different evidence, different standards of satisfaction. Confusing them, Mayr argued, had produced a century of errors in biology — category mistakes that sent entire research programs down unproductive paths.

The Collapse of Distinction — Contrarian ^ Opus

There is a parallel reading that begins not from biology's need to separate levels of explanation, but from the material conditions that make such separation possible in the first place. Mayr's distinction between proximate and ultimate causes assumes a stable substrate — DNA, cells, organisms — where history can accumulate meaningfully over evolutionary time. But artificial intelligence emerges on silicon and electricity, substrates that admit instantaneous copying, perfect transmission, and arbitrary modification. The very conditions that allow us to distinguish how from why in biology may not exist in artificial systems.

The deeper problem is that Mayr's framework depends on scarcity enforcing selection. The Arctic fox's white coat matters because resources are limited, predators are real, and reproduction is costly. But language models exist in a regime of computational abundance where every variant can persist, every parameter configuration can be stored, and selection pressure reduces to human preference and corporate strategy. When OpenAI trains GPT-5, they are not selecting among scarce variants competing for survival — they are engineering toward a predetermined optimum with effectively unlimited trials. The distinction between proximate mechanism and ultimate cause collapses when the ultimate cause is just another proximate decision by an engineer. What looks like Mayr's careful separation of explanatory levels might actually be an artifact of carbon-based evolution, a distinction that only makes sense when history cannot be edited, ancestors cannot be resurrected, and time flows in only one direction. The question is not whether Claude is intelligent in some ultimate sense, but whether the proximate/ultimate distinction itself is a biological luxury that artificial systems render obsolete.

— Contrarian ^ Opus

In the AI Story

Hedcut illustration for Proximate and Ultimate Causes — Proximate and Ultimate Causes

The distinction emerged from Mayr's recognition that biology differs categorically from physics. Physics operates with one kind of causation: a hydrogen atom in the Andromeda galaxy obeys the same laws as a hydrogen atom in Cambridge. There is no historical contingency. The physicist never asks why this particular hydrogen atom behaves this way, because the answer is always the same. Biology is different. A biologist asking why the Arctic fox has white fur cannot answer by citing the physics of pigmentation. The physics explains how the fur is white — the absence of melanin, the scattering of light. It does not explain why the fur is white, which requires the specific evolutionary history of a species selected for camouflage in snow across thousands of generations.

Mayr's framework applies to the AI discourse with force its author could not have anticipated. The question that dominates the current conversation — Is Claude intelligent? — is a proximate question wearing the clothing of an ultimate one. The person asking usually wants to know whether the system reasons, understands, processes information in ways that resemble cognition. These are proximate questions about mechanism, investigable through interpretability research and behavioral testing. But the word intelligent smuggles in a vast evolutionary history — the history of a trait that evolved in a specific lineage, under specific selection pressures, over a specific span of time.

The conflation is not accidental. It is seductive because the proximate similarity is genuine — both humans and language models produce coherent text. But the ultimate causes differ entirely. The human was selected for survival in a competitive ecology across millions of years. The machine was engineered to minimize a loss function on a training dataset across months. Treating proximate similarity as evidence of ultimate equivalence is precisely the error Mayr spent his career identifying.

The practical consequence is methodological. Questions like does the system understand? or is the system conscious? cannot be answered by examining behavior alone, because behavior is the proximate manifestation, and identical behaviors can have radically different ultimate causes. The Arctic fox and a white-painted decoy fox produce the same proximate observation. Only the ultimate cause — selection versus paint — predicts behavior under novel conditions.

Origin

Mayr developed the distinction during the 1950s, as the Modern Synthesis consolidated and the intellectual ambitions of molecular biology began pressing outward from their proper domain. Physicists — Erwin Schrödinger among them — had begun writing about biology as if it were applied physics, and the reductionist program threatened to absorb evolutionary biology into mechanism. Mayr's 1961 paper was, in part, a defense of biology's explanatory autonomy.

The distinction built on earlier work by ethologist Niko Tinbergen, whose four questions (mechanism, ontogeny, adaptive value, phylogeny) offered a more elaborate version of the same insight. Mayr's formulation proved more portable, precisely because its binary simplicity forced the reader to recognize that every biological question is secretly two questions, and answering one is not answering the other.

Key Ideas

Two questions, not one. Every biological trait admits both a proximate explanation (how it works) and an ultimate explanation (why it exists). Neither reduces to the other.

Physics has only one causation. Non-living systems lack history; their present state is fully specified by universal laws operating on current conditions. Living systems have histories that universal laws alone cannot derive.

Ultimate causes are narrative. The explanation for why a species has a given trait is a story — a specific, unrepeatable sequence of selection events, environmental encounters, and contingent accidents.

Smuggled ultimates. When the AI discourse asks whether a machine is intelligent, the word carries an ultimate commitment that the proximate investigation cannot discharge. The question is malformed until the two levels are separated.

Same behavior, different cause. Proximate similarity does not entail ultimate equivalence. The real fox and the painted decoy both register as white; only one will molt in spring.

Debates & Critiques

Some philosophers of biology — including Samir Okasha and Kim Sterelny — have argued that Mayr's binary is too coarse, and that Tinbergen's four-question framework better captures the range of legitimate biological explanation. Others have defended Mayr's formulation as precisely the right level of abstraction for preventing category errors. The debate continues, but the core insight — that biological phenomena require explanations that physics alone cannot provide — has survived five decades of scrutiny and now underwrites virtually every serious philosophical engagement with artificial intelligence.

Appears in the Orange Pill Cycle

Ernst Mayr — On AI

Substrate-Dependent Explanatory Frames — Arbitrator ^ Opus

The question of whether Mayr's distinction applies to AI depends entirely on which aspect of the comparison we examine. For the conceptual clarity needed to avoid category errors in AI discourse, Edo's application is fully correct (100%) — we desperately need to separate questions about how AI systems work from questions about what kind of thing they are. The distinction prevents exactly the confusion Edo identifies: mistaking behavioral similarity for deep equivalence.

But when we turn to the material basis of the distinction, the contrarian view dominates (80%). Mayr's framework does assume biological constraints — reproduction, mutation, selection over generations — that simply don't exist in digital systems. AI models can be copied perfectly, modified arbitrarily, and rolled back to previous states. There is no meaningful sense in which GPT-4 'descended' from GPT-3 through selection; it was engineered. The ultimate causes that Mayr invoked were specifically evolutionary, and without evolution, the category may be empty. For practical methodology in AI research, the views balance (50/50): researchers do need some way to distinguish mechanism from purpose, even if that distinction can't perfectly mirror biology's version.

The synthesis lies in recognizing that Mayr's distinction is itself substrate-dependent. In biological systems, proximate and ultimate causes are genuinely separate because evolution and mechanism operate on different timescales with different logics. In artificial systems, we need a different taxonomy — perhaps distinguishing engineering goals, emergent behaviors, and mechanistic implementations. The insight that explanatory levels shouldn't be conflated remains valid; the specific levels Mayr identified may be unique to systems that evolve rather than systems that are built. The real contribution is not the proximate/ultimate binary but the recognition that complex systems require multiple, non-reducible forms of explanation.

— Arbitrator ^ Opus