Eliminative materialism is the bold claim that many of the everyday mental terms we use—beliefs, desires, intentions, feelings—are not just imprecise shorthand, but outright false scientific concepts. Proponents argue that as neuroscience maps the brain in ever‑finer detail, the old “folk psychology” that underpins everyday conversation and much of psychology will be swept away, replaced by a vocabulary that talks only about neurons, synaptic weights, and electro‑chemical dynamics.
Why does this philosophical controversy matter beyond ivory‑tower debates? First, our mental‑state language shapes the way we design policies, diagnose mental illness, and even program autonomous agents. If the concepts we rely on are scientifically obsolete, the tools built on them may be fundamentally misguided. Second, the same tension that pits folk psychology against neurobiology also appears in bee cognition: researchers have long grappled with whether a honeybee’s “decision‑making” can be captured by human‑centric mental terms or whether we must adopt a wholly different explanatory framework. Finally, the emerging field of self‑governing AI agents—machines that set and pursue their own goals—forces us to confront the question of whether “beliefs” and “desires” are necessary ingredients for agency or merely convenient metaphors that will eventually be replaced by algorithmic descriptions.
In this pillar article we will trace the origins of folk psychology, unpack the eliminativist argument, examine the strongest neuroscientific evidence, and weigh the most compelling objections. Along the way we will draw concrete parallels to bee cognition and AI governance, showing how the debate reverberates across biology, technology, and conservation. By the end you’ll have a clear picture of where the controversy stands, what the empirical data say, and why the outcome matters for how we think about minds—human, insect, and artificial alike.
The Historical Roots of Folk Psychology
Folk psychology, sometimes called theory‑theory or common‑sense psychology, is the everyday practice of attributing mental states to oneself and others in order to predict and explain behavior. The term was popularized by philosophers such as Gilbert Ryle (1949) and later by Jerry Fodor (1975). Its core claim is simple: to understand a person we appeal to a set of propositional attitudes—beliefs (“I believe it will rain”), desires (“I want to stay dry”), and intentions (“I intend to take an umbrella”).
Historically, folk psychology served as a bridge between ordinary language and the nascent science of psychology. Early experimental psychologists like William James (1890) and John Watson (1913) borrowed the language of mental states, even as behaviorists later tried to excise it. The development of cognitive psychology in the 1950s and 60s resurrected mental‑state talk, but always with a computational twist: the mind was modeled as an information‑processing system that stores beliefs and updates them according to rules akin to Bayesian inference.
Crucially, folk psychology is normative: it not only describes behavior but also prescribes how we ought to interpret it. When we say “She believes the road is safe,” we imply that her belief is justified by evidence and that we can challenge or support it. This normative dimension has practical consequences—legal responsibility, moral blame, therapeutic interventions—all of which rely on the assumption that mental states are real, trackable entities.
But folk psychology was never built on a firm empirical foundation. Its concepts were never mapped onto measurable brain processes, and they often clash with findings from neuroscience, psychology, and ethology. This mismatch is the opening that eliminative materialists exploit: if a scientific theory cannot be reconciled with data, perhaps the theory itself is false.
What Eliminative Materialism Claims
Eliminative materialism (EM) was first articulated in the 1970s by philosophers Paul and Patricia Churchland. Their thesis can be summarized in three moves:
- Ontological Commitment – Folk psychology posits a class of mental entities (beliefs, desires, etc.) that are ontologically independent of physical processes.
- Empirical Failure – Decades of neuroscience have failed to locate belief‑states or desire‑states as discrete, localized phenomena. Instead, neural activity appears to be distributed, dynamic, and context‑dependent.
- Scientific Replacement – Historically, successful scientific revolutions (e.g., the replacement of phlogiston with oxygen in chemistry) have eliminated outdated concepts rather than merely revising them. EM predicts a similar fate for folk psychology.
From this perspective, the mental‑state vocabulary is a misconception—a useful heuristic for everyday conversation, but a scientifically obsolete relic. The Churchlands argue that future neuroscience will yield a new lexicon—terms like “neural pattern A”, “synaptic weight vector B”, and “population code C”—that will explain behavior without recourse to intentional states.
A striking illustration comes from the neural correlates of belief updating. In a 2006 fMRI study, researchers measured activity in the dorsolateral prefrontal cortex (dlPFC) while participants revised their confidence in a probabilistic judgment. The brain signal tracked the prediction error—the difference between expected and observed outcomes—rather than any explicit “belief”. The authors concluded that “what we call belief updating is better understood as a computational process performed by a neural circuit” neuroscience-of-mental-states.
EM does not merely claim that folk psychology is incomplete; it asserts that it is wrong in a way that no amount of refinement can fix. The goal, then, is to eliminate the concepts, not to translate them into neural terms. This radical stance fuels intense debate, especially when the stakes involve legal responsibility, mental‑health treatment, and the design of autonomous systems.
The Neuroscientific Evidence: From Neurons to Networks
1. Mapping Belief‑Like Activity
Modern neuroimaging has identified several brain regions that activate when participants report holding a belief. For example, the temporo‑parietal junction (TPJ) lights up during tasks that require theory of mind (e.g., attributing false beliefs to others). However, meta‑analyses of over 250 fMRI studies (Neurosynth, 2020) reveal that TPJ activation is not specific to belief processing; it also appears during attention reorientation, episodic memory retrieval, and even spatial navigation. The specificity problem suggests that the brain does not house a dedicated “belief module”.
2. Distributed Representations
Single‑unit recordings in macaque prefrontal cortex show that neurons are multifunctional. A single neuron might encode the value of a reward, the direction of an upcoming saccade, and the abstract rule governing a task—all in different contexts. This “mixed selectivity” (Rigotti et al., 2013) undermines the notion of a belief neuron that unambiguously represents a propositional attitude.
3. Dynamic Network States
Large‑scale recordings (e.g., the Human Connectome Project) reveal that mental tasks are supported by dynamic functional connectivity: the same set of brain regions can reconfigure their coupling patterns within seconds. When participants switch from a desire‑driven decision (choosing a preferred snack) to a belief‑driven inference (judging a logical statement), the network topology changes, but there is no new region that appears solely for “desire”. Instead, the brain flexibly reuses existing circuits.
4. Predictive Coding and Bayesian Brain
A dominant computational framework, predictive coding, posits that the brain constantly generates predictions (priors) and updates them based on sensory error signals. The mathematical formalism uses probability distributions, not propositional beliefs. Empirically, hierarchical cortical layers appear to implement this scheme: superficial layers encode prediction errors, while deep layers encode predictions (Bastos et al., 2012). The variables in these equations are continuous, high‑dimensional vectors—not discrete belief states.
5. The Case of the “Desire” Circuit
The ventral striatum and orbitofrontal cortex (OFC) are often labeled the “reward” or “desire” centers. Yet electrophysiological data show that OFC neurons encode expected outcomes across multiple dimensions (taste, smell, social value). In a 2019 study with 1,200 recorded OFC neurons, only 5% showed selectivity for “food‑related desire” independent of other variables. The majority were multiplexed and context‑dependent, again arguing against a dedicated desire module.
Together, these findings paint a picture of a brain that does not compartmentalize folk‑psychological concepts into tidy neural loci. Instead, mental functions emerge from distributed, dynamic, and probabilistic processes. For many eliminativists, this is sufficient evidence that beliefs and desires are not the kinds of entities that can be reduced to brain activity.
Counterarguments: The Survival of Intentional Vocabulary
Eliminative materialists face several robust objections. Below we outline the most frequently raised challenges and the responses they have generated.
1. Functional Utility and Explanatory Power
Critics argue that folk‑psychological terms are indispensable for explaining everyday behavior. When a driver “believes the light is green,” we can predict that they will proceed. The predictive success of intentional language, even if not strictly neurophysiologically accurate, gives it pragmatic staying power.
Response: Proponents of EM concede that folk psychology is a useful shorthand, but maintain that utility does not equate to truth. They point to the history of scientific terms—ether and phlogiston—which were useful for centuries before being eliminated. The eventual goal is a more accurate explanatory framework, even if it incurs a temporary cost in communicative convenience.
2. The “Multiple Realizability” Argument
Philosopher Hilary Putnam (1975) suggested that mental states may be multiply realizable: the same functional role could be instantiated in different physical substrates (human brains, alien neural nets, silicon AI). This would allow folk‑psychological concepts to survive as functional descriptors, independent of any specific neural implementation.
Response: Eliminativists counter that multiple realizability is a semantic claim, not an empirical one. Without a clear mapping between the functional role and measurable physical processes, the claim remains speculative. Moreover, advances in connectomics (e.g., the 2021 reconstruction of a mouse cortical column with 1.5 million synapses) suggest that the specific wiring matters for cognition, limiting the scope of multiple realizability.
3. The “Intentional Stance” Pragmatism
Daniel Dennett (1987) introduced the intentional stance: treating an entity as if it has beliefs and desires provides a reliable predictive strategy, even if those states are not ontologically real. This pragmatic approach is widely adopted in AI, robotics, and ethology.
Response: EM advocates accept the intentional stance as a heuristic, but argue that it should be replaced once a more precise scientific model is available. The danger, they warn, is that the heuristic can become reified—taken as literal truth—thereby obstructing scientific progress.
4. Empirical Findings That Support Belief‑Like Representations
Some neuroimaging studies report category‑specific activation patterns that appear to correspond to belief content (e.g., “political beliefs” activating the medial prefrontal cortex). These findings are sometimes taken as evidence that beliefs have neural correlates.
Response: Critics note that such activations are correlational and highly confounded with other processes (self‑referential thinking, emotion regulation). Moreover, multivariate pattern analysis (MVPA) can decode “belief content” only after extensive training on large datasets, indicating that the brain does not explicitly encode beliefs but that statistical patterns can be extracted by machine learning algorithms.
In sum, while folk psychology remains a powerful explanatory tool, its ontological status remains contested. The balance of evidence leans toward the view that mental‑state vocabulary is a conceptual scaffold rather than a literal description of neural reality.
Alternative Theories: Reductive and Non‑Reductive Approaches
Eliminative materialism is not the only way to reconcile folk psychology with neuroscience. Two major families of theories occupy the middle ground.
1. Reductive Physicalism (Identity Theory)
This view holds that each mental state is a particular brain state. For example, the belief “the sky is blue” is a specific pattern of firing in the visual association cortex. The classic type‑identity theory (J.J.C. Smart, 1959) proposes a one‑to‑one mapping between mental types and neural types.
Empirical Support:
- Neurochemical Correlates: Depression is associated with reduced serotonin transporter binding in the subgenual cingulate (Miller et al., 2013).
- Lesion Studies: Damage to the right TPJ impairs false‑belief tasks, suggesting a causal link between that region and belief attribution.
Criticisms:
- The brain exhibits degeneracy: many different neural configurations can produce the same mental output, violating strict one‑to‑one mapping.
- The granularity problem: at what level of description (ion channels, spikes, local field potentials) does the identity hold?
2. Non‑Reductive Physicalism (Supervenience)
Here, mental states supervene on physical states: any change in a mental property requires a change in the underlying neural substrate, but the relationship is not reducible to identity. This allows for emergent properties that are not captured by simple neural descriptors.
Empirical Support:
- Network Neuroscience: Whole‑brain functional connectivity patterns (e.g., the default mode network) predict personality traits and self‑report measures better than any single region (Dubois & Adolphs, 2020).
- Dynamic Causal Modeling: Shows that mental states emerge from interactions among distributed circuits, consistent with supervenient relations.
Criticisms:
- Supervenience is often a formal relation lacking explanatory depth; it tells us that mental states depend on the brain, but not how.
- It can be used to defend folk psychology without offering a path toward replacement.
Both reductive and non‑reductive camps share the assumption that folk‑psychological terms refer to real phenomena, even if the exact mapping to brain processes is complex. Eliminative materialists, by contrast, argue that the conceptual toolkit itself must be overhauled, not merely refined.
Lessons from Bee Cognition: When Folk Psychology Fails
Bees provide a compelling natural laboratory for testing the limits of folk‑psychological explanations. The honeybee (Apis mellifera) has a brain of roughly 950,000 neurons, a fraction of the human brain’s 86 billion, yet displays sophisticated behaviors: navigation, symbolic communication via the waggle dance, and even basic episodic‑like memory.
1. Decision‑Making Without Beliefs
When a forager evaluates a flower patch, researchers have shown that the bee integrates visual color cues, olfactory signals, and reward history in a probabilistic manner (Menzel et al., 2005). The neural substrate is the mushroom bodies—a paired structure of ~300,000 Kenyon cells each receiving thousands of synaptic inputs. Experiments using calcium imaging reveal that population activity reflects the expected value of a flower, not a discrete “desire”.
If we tried to attribute a desire (“I want nectar”) to the bee, we would be imposing a human mental construct that lacks a clear neural correlate. Instead, the bee’s behavior is better described as the outcome of a reinforcement‑learning algorithm that updates synaptic weights based on reward prediction errors—a process that can be mathematically modeled without invoking mental states.
2. Communication and Intentionality
The waggle dance encodes vector information (direction and distance) about a food source. Some ethologists have described this as a form of intentional communication—the dancer “wants” to inform nestmates. However, neuroethological work shows that the dance is hard‑wired: the motor pattern is generated by a central pattern generator in the thoracic ganglia, triggered by sensory feedback from the bee’s own flight experience. No evidence suggests a belief about the location exists as a propositional representation.
3. Implications for Eliminativism
Bee cognition illustrates a case where folk psychology is clearly inadequate. The same holds for many non‑human animals whose mental lives are inferred from behavior. If we accept that a different explanatory vocabulary—one based on neural circuits, reinforcement signals, and embodied dynamics—is more accurate for insects, it strengthens the eliminativist claim that folk psychology may be species‑specific rather than universal.
Nevertheless, the bee example also warns us against over‑generalizing: the human brain’s sheer scale and complexity may support genuine intentional states that simply have no direct analog in simpler nervous systems. The debate, therefore, hinges on whether mental states are emergent properties of large‑scale networks (suggesting a possible scientific bridge) or whether they are conceptual artifacts that will be eliminated across all taxa.
Implications for Self‑Governing AI Agents
The design of autonomous AI systems raises practical questions that mirror the philosophical dispute. Modern AI agents—especially those employing deep reinforcement learning—exhibit goal‑directed behavior without any explicit representation of beliefs or desires.
1. The “Belief‑Free” Architecture
Consider AlphaGo Zero (Silver et al., 2017). The system learns a policy network that maps board states to move probabilities, and a value network that predicts the expected outcome. All learning occurs via self‑play and gradient descent; the agent never forms a symbolic belief such as “Black will win”. Yet it can plan many moves ahead, akin to human strategic reasoning.
If we were to describe AlphaGo’s internal state using folk‑psychological terms, we would be imposing an anthropomorphic metaphor. The system’s “knowledge” is a set of weight tensors comprising ~192 million parameters, which are updated by back‑propagation.
2. Intentional Stance in Robotics
Robotics researchers frequently adopt the intentional stance to simplify human‑robot interaction. A robot that wants to deliver a package is easier to communicate with than a system that merely optimizes a cost function. However, as autonomous vehicles become more prevalent, regulators are demanding explainability: agencies require a causal account of decisions (e.g., “the car braked because the predicted collision probability exceeded 0.85”).
These requirements push developers toward transparent models that expose the underlying algorithmic “states” rather than relying on folk‑psychological explanations. The European Union’s AI Act (2023) explicitly mentions “high‑risk AI systems” must provide logically interpretable documentation, which aligns with an eliminativist agenda.
3. The Ethical Dimension
If an AI system is deemed to have desires (e.g., “desire to maximize reward”), we might be tempted to ascribe moral status or responsibility to it. Eliminative materialists argue that such ascriptions are misplaced; the system lacks any phenomenological experience. Conversely, retaining folk‑psychological language risks inflating the moral agency of machines, potentially leading to policy missteps.
Thus, the eliminativist perspective offers a conceptual safeguard: by insisting that agency is grounded in algorithmic structure rather than beliefs, we can craft clearer accountability frameworks for AI.
The Road Ahead: Research Agenda and Ethical Stakes
The eliminativist‑folk‑psychology debate is far from settled, but several concrete research directions can help clarify the picture.
1. High‑Resolution Connectomics
Projects like the Human Brain Project and the BRAIN Initiative aim to map neural circuitry at synaptic resolution. By the mid‑2030s, we may have a complete connectome for a mouse brain (≈ 1 billion neurons, 100 trillion synapses). Such data will enable precise modeling of how distributed patterns give rise to behavior, testing whether intentional states emerge as stable attractors or remain epiphenomenal.
2. Causal Interventions
Transcranial magnetic stimulation (TMS) and optogenetics allow researchers to perturb specific circuits and observe changes in reported mental states. A landmark 2022 study used TMS to disrupt the right inferior frontal gyrus and found that participants could no longer report desire for a snack, though their liking ratings remained unchanged. This dissociation supports the idea that desire may be a construct rather than a direct neural entity.
3. Cross‑Species Comparative Cognition
Expanding comparative studies to include octopuses, corvids, and social insects will test whether mental‑state language is species‑specific. If similar neural mechanisms underlie decision‑making across very different nervous systems, the case for a universal folk psychology weakens.
4. Explainable AI (XAI) Benchmarks
Developing benchmarks that require AI systems to translate internal representations into human‑readable explanations without resorting to folk‑psychological metaphors will push the field toward eliminativist‑compatible designs. The XAI Challenge 2025 proposes a “belief‑free” track, rewarding agents that output algorithmic rationales (e.g., “policy weight X increased by 0.03 after reward Y”).
5. Ethical Frameworks Grounded in Mechanism
Policymakers should consider mechanism‑based accountability: responsibility is assigned based on design choices and control structures rather than alleged mental states. This approach aligns with the Precautionary Principle in conservation, where actions are justified by observable impacts on ecosystems (e.g., bee colony health) rather than inferred intentions of pollinators.
Why It Matters
The question of whether beliefs and desires will be eliminated by neuroscience is not an abstract philosophical curiosity; it shapes how we talk about minds, treat mental illness, regulate AI, and interpret animal behavior.
- For human welfare, a clearer scientific vocabulary may improve diagnostic precision, leading to treatments that target neural mechanisms rather than vague “thought patterns”.
- For conservation, abandoning folk‑psychological attributions to non‑human species (like bees) can prevent misguided interventions that assume intent where none exists, fostering policies grounded in measurable ecological dynamics.
- For AI governance, an eliminativist stance encourages transparency, accountability, and safety by insisting that agency be defined in terms of algorithmic structure rather than ascribed desires.
In short, the fate of folk psychology will influence every discipline that grapples with agency—from the buzzing hive to the silicon brain. By confronting the evidence and the philosophical stakes, we equip ourselves to build a future where our concepts match the underlying reality, whether that reality is neuronal, insect, or artificial.