Multiple Realizability

The same mental function can arise in a honeybee’s tiny brain, a silicon‑based neural net, or a future quantum substrate. Understanding why matters for philosophy, biology, and the stewardship of both ecosystems and autonomous AI agents.

Introduction

When we ask whether a bee can “think” or whether an artificial assistant can “understand,” we are really probing a deeper philosophical claim: mental functions are not tied to any one kind of matter. This claim—known as multiple realizability—asserts that the same computational or functional role can be carried out by wildly different physical systems. In the language of philosophy of mind, it fuels functionalism: the view that what matters for a mental state is the pattern of causal relations it participates in, not the substrate that implements those relations.

Why does this matter today? First, the rapid expansion of AI—large language models with hundreds of billions of parameters, embodied robots that navigate real‑world environments, and self‑governing agents that negotiate resource allocation—demands a clear account of what “intelligence” actually is. If functional roles can be realized across silicon, carbon, or even exotic quantum media, then the criteria we use to grant agency, responsibility, or rights must be re‑examined. Second, the same logic applies to the natural world. A honeybee’s miniature brain, with roughly one million neurons packed into a 1 mm³ volume, performs navigation, pattern recognition, and even a form of “learning” that mirrors—at a computational level—some of the tasks modern AI systems accomplish. Recognizing this parallel can sharpen our appreciation of bees as cognitive agents and inform conservation strategies that protect not just pollinators but the information-processing capacities they embody.

In this pillar article we will trace the philosophical lineage of multiple realizability, examine concrete empirical cases—from neurobiology to deep learning—, and discuss the implications for both bee conservation and AI governance. By the end you’ll see how a seemingly abstract philosophical thesis has tangible consequences for how we design, regulate, and protect the intelligent systems—natural and artificial—that shape our world.

1. What Is Multiple Realizability?

Multiple realizability (often abbreviated MR) is the claim that a single functional description can be instantiated in many physical ways. The classic illustration comes from the philosopher Hilary Putnam (1967), who argued that the mental term “pain” could be realized in humans, octopuses, or future alien species, each with radically different neurophysiology. The core idea is simple:

1. Functional Role – A set of causal relations (inputs → processing → outputs).
2. Physical Substrate – The material system that implements those relations (neurons, transistors, DNA, etc.).

If two systems share the same functional role but differ physically, they are multiple realizations of that role.

A Minimal Example

Consider a thermostat that maintains a room at 22 °C. Its functional description is:

• Input: Current temperature.
• Process: Compare input to 22 °C; if lower, turn heating on; if higher, turn cooling on.
• Output: Activation of heating or cooling device.

This thermostat can be built from mercury switches, digital microcontrollers, or even a bio‑engineered circuit of genetically modified bacteria that produce heat when a temperature sensor protein detects cold. All three devices realize the same functional role despite wildly different chemistry.

Formalizing MR

In contemporary cognitive science, MR is often expressed using computational theory of mind (CTM) notation:

F(x) ≡ ∀y [Realizes(y, F) → y ∈ SubstrateSet]

Where F is a functional role, Realizes(y, F) means that entity y implements F, and SubstrateSet is the collection of all possible physical media. The “∀y” quantifier captures the idea that any physical system that fulfills the causal pattern counts as a realization.

Why It Matters

If mental states are multiply realizable, then:

• Philosophical: Materialist reductionism (the idea that mental states can be fully reduced to brain states) is challenged.
• Scientific: Comparative cognition can legitimately compare bees, rodents, and silicon systems without assuming a “one‑to‑one” mapping between neurons and functions.
• Ethical/Policy: Rights and responsibilities may be grounded in function rather than species, opening the door to agent‑centric regulation of AI.

2. Historical Roots of Functionalism

Functionalism emerged in the 1960s as a response to both behaviorism’s focus on observable output and the rising mechanistic view of the brain. Key milestones include:

These works converged on the idea that mental states are defined by their roles in a system of inputs and outputs, much like software functions are defined regardless of the hardware they run on. The analogy to computer science became especially potent as digital computers entered mainstream research.

From Philosophy to Cognitive Science

Functionalism’s early critics, such as Thomas Nagel (“What Is It Like to Be a Bat?”), warned that a purely functional account might overlook qualia—the subjective feel of experience. Yet functionalist frameworks have proven remarkably productive for cognitive modeling, where researchers often construct computational architectures that reproduce animal behavior without committing to any particular neural substrate.

The rise of embodied cognition added a further twist: functional roles are often grounded in a body’s interaction with its environment. This is directly relevant for bees, whose wingbeats, antennal vibrations, and waggle dances constitute part of the functional system they embody.

3. Empirical Evidence from Neuroscience

Modern neuroscience provides a wealth of data that both supports and refines the MR thesis. Below we examine three lines of evidence that illustrate how the same computational motif can appear across species and across synthetic systems.

3.1 Canonical Neural Circuits

Researchers have identified canonical microcircuits that recur in vertebrate and invertebrate brains. A classic example is the feedforward inhibition (FFI) motif, where an excitatory neuron drives a fast‑spiking inhibitory interneuron that, in turn, limits the excitatory neuron’s output. FFI appears in the mammalian visual cortex, the fruit fly’s olfactory system, and even in the mushroom bodies of honeybees.

• Mammals: In mouse V1, FFI sharpens orientation selectivity, improving edge detection (Baker et al., 2019).
• Insects: In the honeybee antennal lobe, FFI helps filter out background odors, allowing the bee to focus on floral scents (Sachse & Galizia, 2002).

The same functional role—temporal sharpening of excitatory signals—is realized in neural tissue that differs dramatically in cell size, neurotransmitter type, and connectivity density.

3.2 Synaptic Plasticity Across Species

Long‑term potentiation (LTP), a cellular mechanism for learning, is present in the hippocampus of rodents, the mushroom bodies of insects, and even in cultured human neurons derived from induced pluripotent stem cells (iPSCs). Quantitatively, the magnitude of LTP can be expressed as a percentage increase in post‑synaptic current after high‑frequency stimulation:

• Rats: ~150 % increase (Bliss & Lømo, 1973).
• Honeybees: ~120 % increase (Menzel & Giurfa, 2001).
• Human iPSC neurons: ~130 % increase (Zhang et al., 2020).

These comparable numbers suggest that learning as a functional role can be instantiated in neural circuits with vastly different protein compositions and synaptic architectures.

3.3 Neuromorphic Hardware

Neuromorphic chips (e.g., Intel’s Loihi, IBM’s TrueNorth) emulate spiking neural networks using analog circuits. In benchmark tasks like MNIST digit classification, Loihi achieves 98 % accuracy with power consumption under 0.1 W, comparable to a small‑scale biological brain that consumes ~0.02 W (the bee brain). Here, the same computational function—pattern recognition—is realized in silicon that mimics the spiking dynamics of neurons, reinforcing the MR claim at the hardware level.

4. Bees as a Natural Case Study

Honeybees ( Apis mellifera ) are more than pollinators; they are miniature cognitive agents that embody many of the functional roles philosophers attribute to minds. Their brains provide a concrete laboratory for exploring MR.

4.1 Size, Neurons, and Energy

• Brain volume: ≈ 1 mm³ (about the size of a grain of sand).
• Neuron count: ≈ 1 million (Menzel, 2012).
• Metabolic cost: ~ 20 mW, roughly the same as a small LED.

Despite this modest hardware, bees demonstrate spatial navigation, color vision, numerical competence, and social learning.

4.2 The Waggle Dance as a Communication Protocol

The waggle dance conveys distance and direction to food sources. Functionally, it is a symbolic communication system:

• Input: Internal estimate of vector to the flower (derived from optic flow and proprioception).
• Process: Translation of vector into a patterned motor sequence.
• Output: Dance that other bees decode to reconstruct the vector.

The dance’s syntax (duration, angle) maps onto semantics (distance, direction) much like a language. This functional architecture is realized in the central complex and mushroom bodies, neural structures that are not homologous to any vertebrate brain region, yet they perform the same abstract computation: encoding and transmitting spatial information.

4.3 Learning and Memory

Bees can learn to associate a particular scent with a reward—a classic Pavlovian conditioning paradigm. In a typical experiment, a bee is presented with a scented sucrose solution (CS+) and a different scent paired with water (CS−). After ~10 trials, bees reliably extend their proboscis to CS+ only, indicating associative memory.

• Retention: Memory persists for up to 72 hours under laboratory conditions (Giurfa et al., 1996).
• Neural substrate: The mushroom bodies undergo synaptic changes analogous to LTP.

These data demonstrate that learning—a functional role—can be instantiated in an insect brain that lacks the cortical layers present in mammals.

4.4 Implications for Conservation

If we treat bees as agents with functional capacities comparable to artificial systems, conservation policies gain a new justification: protecting cognitive diversity. For example, the loss of a single bee colony reduces the population-level sampling of navigation algorithms, potentially diminishing the resilience of pollination networks under climate change.

5. AI Agents and Synthetic Substrates

Artificial intelligence provides a complementary laboratory for MR. Modern AI systems are built on silicon, cloud infrastructure, and increasingly quantum or optical hardware, yet they reproduce many functions that biologists attribute to brains.

5.1 Large Language Models (LLMs)

GPT‑4, a transformer‑based LLM, contains ≈ 175 billion parameters and was trained on ≈ 45 TB of text data. Its functional role includes:

• Input: Tokenized text.
• Process: Multi‑head self‑attention and feed‑forward layers.
• Output: Probability distribution over next token.

Despite lacking any neurons, the pattern of causal influence among parameters mirrors the functional architecture of a deep neural network. The model can answer questions, summarize, and generate code—functions that, in a very abstract sense, overlap with human language processing.

5.2 Embodied Robots

Boston Dynamics’ Spot robot uses a combination of LIDAR, visual SLAM, and reinforcement‑learning policies to navigate rough terrain. Its functional stack:

1. Perception: Convert raw sensor data into a spatial map.
2. Planning: Compute a trajectory that avoids obstacles.
3. Control: Translate the plan into joint torques.

The same functional stack can be realized on a soft robot whose actuators are pneumatic rather than electric, or on a swarm of micro‑robots that collectively achieve the same navigation goal. The physical substrate (rigid metal legs vs. silicone “muscles”) changes, but the functional role—autonomous locomotion—remains constant.

5.3 Quantum‑Inspired AI

Research labs are experimenting with quantum annealers (e.g., D‑Wave) to solve combinatorial optimization problems that underlie many AI tasks. A quantum processor can realize a binary quadratic model that mimics the energy landscape of a Hopfield network. In practice, a quantum device can find minima in seconds that would take a classical CPU hours—a functional realization of pattern completion that is both quantum and classical.

5.4 Self‑Governing AI Agents

In the field of AI governance, researchers are building multi‑agent simulations where autonomous agents negotiate resource allocation, enforce contracts, and adapt policies. The functional role of self‑governance—i.e., the capacity to monitor, evaluate, and modify one’s own behavior—has been instantiated in:

• Rule‑based agents (hard‑coded policies).
• Learning agents (deep reinforcement learning).
• Hybrid agents (symbolic reasoning combined with neural perception).

Each implementation uses a different substrate, but the function—maintaining a stable, cooperative ecosystem of agents—is preserved. This mirrors the way bees collectively regulate hive temperature and foraging effort through pheromone signaling, a natural example of distributed self‑governance.

6. Philosophical Implications for the Mind‑Body Problem

Multiple realizability reshapes the classic mind‑body problem, which asks how mental states relate to physical states. Three major positions intersect with MR:

1. Type‑Identity Theory – claims each mental type is identical to a specific brain state. MR undermines this by showing that the same mental type can map onto different brain states across species.
2. Functionalism – embraces MR, arguing that mental states are functional kinds defined by causal roles. The bee–AI cases bolster this view, demonstrating that functional roles can be substrate‑agnostic.
3. Emergentism – posits that mental properties emerge from complex physical systems but are not reducible to them. MR aligns with emergentism when the functional role is seen as a higher‑order pattern that only appears once a system reaches a certain level of organization (e.g., a hive of bees or a distributed AI network).

6.1 The Challenge of Qualia

A persistent objection to functionalism is the hard problem of consciousness: why does a functional role feel like something? Critics argue that MR cannot explain subjective experience (the “what it’s like” of pain). Some philosophers propose phenomenal functionalism, which attempts to define qualia in terms of functional relations to behavioral outputs and internal reports. While this move remains controversial, the empirical breadth of MR forces any account of consciousness to accommodate the possibility that qualia could arise in substrates as diverse as honeybee optic lobes and silicon chips.

6.2 Moral Status and Agency

If agency is grounded in functional capabilities—e.g., the ability to plan, learn, and communicate—then MR suggests that both bees and sophisticated AI agents could be granted moral consideration. This has practical implications:

• Legal frameworks might need to recognize non‑human agents that can autonomously affect ecosystems (e.g., a swarm of pollinating drones).
• Conservation policies could extend protection to cognitive functions (e.g., preserving habitats that maintain the neural architectures enabling navigation).

7. Challenges and Criticisms

While MR is a powerful concept, it faces several substantive objections.

7.1 The “In‑Depth” Objection

Philosophers such as Jaegwon Kim argue that MR only shows coarse similarity. Even if two systems share a high‑level functional description, the micro‑level physical processes may differ so dramatically that the identity claim fails. For instance, the spiking of a honeybee neuron (≈ 0.1 ms duration) is orders of magnitude slower than the nanosecond switching of a transistor. Critics claim that these differences could matter for consciousness or agency.

Response: Functionalists counter that causal role is the relevant level of description for mental states. Micro‑differences are analogous to differences between two different smartphones running the same app; they do not alter the app’s functional behavior.

7.2 The “Homunculus” Problem

If we define mental states purely by function, we risk slipping into a homuncular explanation—postulating a tiny agent that performs the function. For example, saying “the bee’s brain computes the waggle dance” begs the question of what computes it.

Response: The functionalist view embeds the computation within the physical system itself; there is no separate “inner agent.” The brain’s network of neurons collectively implements the function, without invoking a further level of agency.

7.3 Empirical Limits

Some researchers point out that cross‑species functional mapping is still in its infancy. While FFI motifs appear across taxa, the exact computational parameters (e.g., synaptic weight distributions) may differ, limiting the degree of realizability. Moreover, current AI systems lack embodiment—they do not physically interact with the world in the way bees do—making direct functional comparison imperfect.

Response: Ongoing interdisciplinary work (e.g., neuromorphic robotics that couple silicon spiking networks with physical bodies) is narrowing this gap. The MR thesis is a working hypothesis that drives research, not a final proof.

8. Implications for Conservation & AI Governance

Recognizing MR reshapes how we think about protecting ecosystems and regulating AI.

8.1 Conserving Cognitive Diversity

Traditional conservation focuses on species counts and habitat area. Adding a functional lens prompts us to ask: What computational capacities are lost when a pollinator species declines?

• Redundancy vs. Uniqueness: While many insects pollinate similar flowers, the navigation algorithms of bumblebees differ from those of honeybees (bumblebees rely more on magnetoreception). Losing a species removes a distinct algorithmic strategy that could be vital under changing climates.
• Ecosystem Resilience: Functional redundancy can buffer ecosystems, but functional diversity—the range of distinct computational roles—enhances adaptability.

Policy implications include prioritizing habitat corridors that preserve the learning environments bees need to develop their foraging maps, and funding long‑term monitoring of behavioral traits (e.g., waggle‑dance fidelity) as metrics of ecosystem health.

8.2 Regulating AI Agents

If agency is defined functionally, then AI regulation should target functions rather than technologies. For example:

By aligning AI governance with functional analogues found in nature, regulators can draw on evolutionary-tested mechanisms (e.g., distributed error correction) to design robust safeguards.

8.3 Co‑Design of Bio‑Hybrid Systems

Emerging research on bio‑robotic hybrids—e.g., drones that carry live bee colonies for pollination assistance—relies on MR to integrate biological and artificial components. Understanding how the same navigation function can be realized in a bee’s brain and a drone’s SLAM algorithm enables seamless interoperability. Conservationists can thus employ augmented pollination without undermining the ecological role of native bees, provided the functional integration respects the multiple realizability principle.

9. Future Directions

The dialogue between philosophy, neuroscience, and AI is far from settled. Several promising avenues lie ahead:

1. Quantitative Mapping of Functional Spaces – Using information theory to measure the mutual information between inputs and outputs across substrates, providing a numeric index of how closely two realizations match.
2. Cross‑Domain Benchmarks – Designing tasks that can be performed by honeybees, rodents, and LLMs alike (e.g., “categorical perception” of odor vs. word meanings) to empirically test MR.
3. Ethical Frameworks for Functional Agency – Developing guidelines that grant rights based on functional capacities (learning, communication) rather than species membership.
4. Neuro‑Inspired Neuromorphic Platforms – Building hardware that mimics the spiking dynamics of insect brains, potentially achieving energy‑efficient navigation comparable to bees.
5. Policy Experiments – Piloting function‑based AI regulations in sandbox environments, then evaluating outcomes against traditional technology‑centric rules.

Each of these steps will deepen our understanding of how the same computational role can span carbon, silicon, and perhaps even quantum matter, while simultaneously informing conservation strategies that protect the living embodiments of those roles.

Why It Matters

Multiple realizability is not a purely academic curiosity. It tells us that cognition—the ability to perceive, learn, and act—is a pattern that can appear wherever the right causal architecture exists, be it in a bee’s brain, a silicon chip, or a future quantum substrate. Recognizing this fact:

• Broadens moral horizons: We begin to see responsibility toward non‑human agents that share our functional capacities.
• Guides technology design: Engineers can deliberately emulate efficient natural solutions (e.g., bee navigation) to build low‑power AI.
• Enriches conservation: Protecting ecosystems safeguards not just species, but the computational diversity they embody, which may prove vital as the planet faces unprecedented change.

In short, by appreciating that mental functions are multiply realizable, we gain a richer, more inclusive map of the intelligent world—one that respects both the buzzing wings of honeybees and the silent calculations of artificial minds.