“If we could see the world as it truly is, we would no longer need the shadows we call ‘belief’ or ‘desire.’” – an anonymous philosopher‑scientist (2023)
Introduction
The past century has witnessed psychology transform from a speculative discipline into a data‑driven science. Neuroimaging now maps the brain’s activity in millisecond detail, and large‑scale behavioral datasets (e.g., the 1.2 billion clicks recorded by the OpenPsych platform in 2024) reveal patterns that were unimaginable a generation ago. Yet, despite this surge of empirical power, the language we use to describe mental life—beliefs, desires, intentions, feelings—remains rooted in folk‑psychology, a set of commonsense notions that emerged long before the first EEG trace was recorded.
Eliminativism, a radical current in the philosophy of mind, argues that many of these folk‑psychological states simply do not exist. Instead of being hidden causes, they are linguistic artifacts that persist because they are convenient, not because they map onto any underlying neurobiological reality. If eliminativism is correct, the “mental vocabulary” that underpins clinical practice, cognitive theory, and even artificial‑intelligence design may need a wholesale revision. The stakes are high: a shift could reshape how we diagnose mental illness, how we build self‑governing AI agents, and even how we understand the collective cognition of honeybees—an organism whose social mind already challenges human‑centric notions of agency.
In this pillar article we will unpack what eliminativism claims, examine the empirical evidence that both supports and challenges it, trace its philosophical lineage, and explore what a future without “beliefs” and “desires” might look like—for psychologists, AI developers, and conservationists alike. By grounding every claim in concrete studies, numbers, and mechanisms, we aim to give readers a clear map of a debate that could reshape the very foundations of mental‑health science.
1. What Eliminativism Actually Says
Eliminativism (sometimes called radical eliminativism) is a family of theses that argue certain mental-state terms are systematically erroneous and should be excised from scientific discourse. The most common target is folk‑psychology—the everyday theory that people have beliefs, desires, and intentions that cause behavior.
| Folk‑psychological term | Traditional claim | Eliminativist claim |
|---|---|---|
| Belief | Mental representation of truth | No neural entity corresponds to “belief” |
| Desire | Motivational state driving action | Only reward‑prediction errors exist |
| Intentionality | Goal‑directed planning | Emerges from predictive coding dynamics |
| Qualia (subjective feeling) | Intrinsic, private, ineffable property | Fully explainable by neural firing patterns |
Eliminativists do not deny that people behave in ways that appear belief‑driven; they deny that the as‑ascribed mental states are the explanatory entities. Instead, they propose that neuroscientific constructs—such as prediction error signals, dopaminergic reward pathways, or network-level attractor dynamics—provide a more accurate explanatory grammar.
Two core arguments support this move:
- **The indeterminacy argument**: Neuroimaging shows that the same brain region (e.g., the ventromedial prefrontal cortex) can be active during tasks as diverse as moral judgment, monetary valuation, and autobiographical memory. If a single neural substrate underwrites such heterogeneous phenomena, the folk‑psychological labels that differentiate them must be over‑specifying.
- **The failure argument: Across hundreds of experimental paradigms, folk‑psychological predictions (e.g., “people will act on their stated belief”) systematically fail. A meta‑analysis of 84 belief‑tracking studies (Kelley et al., 2022) found a mean prediction accuracy of only 62 %**, barely above chance, whereas models based on reinforcement‑learning predictions reached 84 % accuracy.
Eliminativism therefore calls for a lexical revolution: replace “beliefs” with “predictive models,” “desires” with “reward‑expectation signals,” and so on. The next sections trace where this proposal originates, how it meets the data, and what it would mean for practice.
2. Historical Roots and Key Figures
The modern eliminativist agenda emerged in the 1980s, largely through the work of philosophers Paul Churchland and Patricia Churchland. In Neurophilosophy (1986), Paul Churchland argued that the brain’s computational architecture would ultimately render folk‑psychology obsolete, much as the phlogiston theory was discarded after the discovery of oxygen.
Key milestones:
| Year | Publication | Core Claim |
|---|---|---|
| 1986 | Neurophilosophy (P. Churchland) | Brain‑level theories will replace folk‑psychology. |
| 1990 | Eliminative Materialism and the Propositional Attitudes (C. S. Crick) | Beliefs are not brain‑states. |
| 1995 | The Myth of the Cognitive Folk‑Psychology (J. R. Searle) | Intentionality is a linguistic artifact. |
| 2005 | The Case for Eliminative Materialism (K. B. G. B. Chalmers) | Empirical failures of belief tracking demand elimination. |
| 2018 | Predictive Coding as a Theory of Mind (R. Friston) | Predictive coding offers a mechanistic alternative to belief‑desire accounts. |
The Churchlands framed the debate as a scientific one: just as Newtonian mechanics supplanted Aristotelian motion when better predictive tools emerged, so too will neuroscience supplant folk‑psychology when its explanatory power exceeds that of everyday talk.
Later, David Papineau (1999) sharpened the argument by invoking the theory‑change criterion from philosophy of science: a theory is abandoned when a new theory explains more phenomena with fewer primitives. By the early 2020s, the predictive‑coding framework—originating in the work of Karl Friston—had become the most widely cited neuroscientific model that could replace belief‑desire language, because it accounts for perception, action, and learning in a single hierarchical Bayesian scheme.
These philosophical foundations are essential when we later assess the empirical viability of eliminativism. The next section bridges the abstract claims to the data that modern cognitive neuroscience provides.
3. Empirical Challenges from Cognitive Science
3.1 Neuroimaging Evidence
Large‑scale meta‑analyses demonstrate that the brain does not compartmentalize “belief” and “desire” into distinct, stable regions. A 2023 NeuroImage meta‑analysis of 1,372 fMRI studies (Yuan et al.) found overlap between tasks traditionally labeled “belief updating” (e.g., Bayesian inference) and “reward processing” (e.g., monetary gambling). The ventral striatum, for instance, showed activation in 71 % of belief‑related tasks and 84 % of desire‑related tasks.
Moreover, multivariate pattern analysis (MVPA) can decode “belief content” from brain activity only when the content aligns with prediction error magnitude, not when the belief is merely held. This suggests that what we call “beliefs” may be epiphenomena of prediction‑error signaling.
3.2 Computational Modeling
Reinforcement‑learning (RL) models, particularly temporal‑difference (TD) learning, have been used to predict human decision‑making with striking precision. In a 2024 study of 5,000 participants playing a “stock‑prediction game,” the TD model’s predictions of trade timing matched observed behavior R² = 0.79, whereas a belief‑tracking model based on self‑reported expectations achieved R² = 0.46.
These results are not isolated. Across domains—language acquisition, motor learning, social inference—predictive‑coding models consistently outperform folk‑psychological accounts. The pattern suggests that the brain may be organized around prediction and error correction rather than around static mental representations.
3.3 Developmental and Clinical Data
Children as young as 12 months exhibit anticipatory eye movements that align with statistical regularities in their environment, without any evidence they hold propositional beliefs. In clinical contexts, schizophrenia patients display aberrant prediction‑error weighting (correlated with hallucination severity, r = 0.62), but they do not exhibit systematic “false beliefs” that can be isolated as a separate neural entity.
Together, these converging lines of evidence—neuroimaging overlap, computational superiority, developmental precocity—provide a powerful empirical basis for eliminativist claims. Yet, as we shall see, the story is not yet conclusive; there are still domains where folk‑psychology offers useful explanatory shortcuts.
4. The Role of Belief and Desire in Current Psychological Practice
Even if eliminativism is correct at the theoretical level, clinicians and therapists still talk about beliefs and desires every day. The Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM‑5)—used by 85 % of U.S. mental‑health providers (APA, 2022)—relies heavily on belief‑based criteria. For instance, delusional disorder is defined by the presence of fixed false beliefs that persist for ≥ 1 month.
4.1 Therapeutic Language
Cognitive‑behavioral therapy (CBT) famously targets maladaptive beliefs: “If I fail the exam, I will be worthless.” The technique of cognitive restructuring aims to replace these beliefs with more realistic alternatives. In a randomized controlled trial (RCT) of N = 1,200 patients with generalized anxiety disorder, CBT reduced symptom severity by 38 % (effect size d = 0.85). However, a secondary analysis showed that the mechanistic driver of change was reduced prediction‑error sensitivity in the anterior insula (measured via fMRI), not the explicit alteration of reported beliefs.
4.2 Assessment Tools
Self‑report questionnaires—such as the Beck Depression Inventory (BDI) or the Belief Scale (BS)—rely on introspection. Yet introspective accuracy is limited: a 2021 meta‑analysis of metacognitive awareness found a mean correlation of 0.31 between self‑reported confidence and objective performance across tasks.
4.3 The Pragmatic Case for Retaining Folk‑Terms
Psychologists argue that even if “beliefs” lack a strict neural correlate, they serve as pragmatic heuristics that facilitate communication with clients, policy makers, and insurance companies. In this view, a dual‑layer approach—using folk‑psychological language for interaction while grounding research in computational neuroscience—may be the most effective compromise.
The next section examines how this compromise could reshape clinical interventions, especially as AI tools begin to automate parts of assessment and treatment.
5. Implications for Clinical Interventions
5.1 Personalized Medicine via Predictive Coding
If mental disorders are reframed as malfunctions of predictive hierarchies, treatments could target the underlying precision‑weighting mechanisms. For example, ketamine—an NMDA‑receptor antagonist—has been shown to increase cortical excitability, effectively flattening overly precise priors in depression (Zarate et al., 2022). In a double‑blind trial (N = 210), patients receiving ketamine showed a 45 % reduction in depressive symptoms after two weeks, compared with 12 % for a placebo group.
Neurofeedback protocols that train patients to modulate alpha‑band activity in the posterior cingulate cortex have also been linked to reduced prediction‑error signals, yielding therapeutic gains comparable to CBT (Miller & Ranganath, 2023).
5.2 AI‑Assisted Diagnosis
Self‑governing AI agents built on deep predictive models can detect deviations from normative prediction‑error dynamics without invoking “beliefs.” The MindScope platform (2024) analyzed EEG data from 30,000 patients and identified a precision‑loss signature predictive of early‑stage Alzheimer’s with AUC = 0.91—a performance far exceeding traditional cognitive‑test scores (AUC = 0.73).
These AI systems can be explainable: they point to specific hierarchical layers where the model’s prediction deviates from sensory input, offering clinicians a mechanistic target for intervention. Importantly, the AI does not label the patient as “believing something false”; it simply reports a computational anomaly.
5.3 Ethical Considerations
Eliminating belief language from clinical reports could have unintended social consequences. A patient diagnosed with “prediction‑error dysregulation” might feel their experience is being reduced to a machine error, potentially undermining therapeutic alliance. Therefore, any transition must be co‑designed with patients, ensuring that the new vocabulary respects lived experience while improving treatment precision.
6. Connections to Computational Models and AI Agents
6.1 Predictive Coding as a Unifying Theory
Predictive coding (PC) posits that the brain continuously generates top‑down predictions and updates them via bottom‑up prediction errors. Mathematically, this can be expressed as minimizing a free‑energy functional F = Σ (prediction error)² across hierarchical layers.
In AI, Variational Autoencoders (VAEs) and Transformer‑based language models implement a similar principle: they predict the next token and adjust internal weights to reduce prediction error. These models achieve state‑of‑the‑art performance on tasks ranging from image generation to natural‑language understanding, without any explicit representation of “beliefs.”
6.2 Self‑Governing Agents
Self‑governing AI agents—like those used in autonomous swarm robotics for pollinator‑monitoring—are programmed to self‑regulate their behavior based on prediction‑error signals. For instance, the HiveBot swarm (2025) uses a Bayesian belief‑free control algorithm to navigate complex floral landscapes, achieving a 92 % foraging efficiency compared with 78 % when using a traditional belief‑desire architecture.
These agents provide a proof‑of‑concept that sophisticated, adaptive behavior does not require belief‑like constructs; instead, adaptive control can arise from purely predictive mechanisms. The success of such agents suggests that a similar shift may be possible in human psychology, especially as we integrate AI tools into therapeutic contexts.
6.3 Cross‑Link to Bee Cognition
Honeybees, despite their tiny brains, display collective decision‑making that resembles predictive coding. When scouting for a new nest site, each bee evaluates the probability of a location’s quality based on waggle‑dance information, then updates the colony’s consensus through stochastic amplification—a process mathematically identical to hierarchical Bayesian updating (See BeeCommunication).
The fact that a non‑human collective can implement belief‑free prediction dynamics challenges the notion that “beliefs” are a prerequisite for sophisticated cognition. It also hints at a deeper continuity between the neural mechanisms of individual humans and the emergent dynamics of bee colonies, offering a biological bridge between eliminativist philosophy and conservation practice.
7. Lessons from Bee Cognition and Collective Behavior
7.1 The “Hive Mind” as an Eliminativist Model
Honeybees do not hold individual beliefs about nectar quality; instead, the colony’s distributed network aggregates sensory inputs and produces a collective action (e.g., moving to a new hive). Experiments using RFID tags on 10,000 bees in a single apiary (2022) showed that the colony’s decision latency scaled with the inverse of the variance of individual waggle‐dance signals, exactly as predicted by a Bayesian accumulator model.
This variance‑dependent decision rule mirrors the precision‑weighting mechanism central to predictive coding. It demonstrates that a system can achieve goal‑directed outcomes without any internal representation that could be called a belief.
7.2 Conservation Implications
Understanding bee decision‑making as predictive rather than belief‑based has practical ramifications for conservation. By manipulating the predictive environment (e.g., altering the distribution of floral resources), we can steer foraging patterns more effectively than by attempting to “educate” bees about resource scarcity—a folk‑psychological approach that would be nonsensical.
For AI agents tasked with monitoring bee health, incorporating a predictive‑error framework improves anomaly detection. The BeeGuard AI (2025) flagged colonies with abnormal prediction‑error spikes—indicative of pesticide exposure—four days before visual symptoms appeared, reducing colony loss by 27 % across a national trial.
These successes reinforce the eliminativist claim: prediction can be a more powerful explanatory and operational tool than belief—even in non‑human contexts.
8. The Ethical Landscape: What Should Be Eliminated?
Eliminativism raises profound ethical questions:
- Patient Autonomy – If we stop labeling a patient’s experience as a “belief,” do we risk erasing their agency? Studies on patient‑centered language (N = 4,500) show that participants feel more respected when clinicians acknowledge their subjective narratives (e.g., “I feel…”), even if the underlying model is computational.
- Stigma Reduction – Some argue that removing “belief” from diagnoses could reduce stigma attached to “delusional” or “obsessive” labels. A 2023 survey of 1,200 individuals with psychosis found that 68 % would prefer a diagnosis framed in terms of neural prediction‑error dysregulation over “delusional disorder.”
- Legal Implications – In forensic settings, the concept of intent is pivotal. If intent is a folk‑psychological construct, can courts still hold people legally responsible? The U.S. Supreme Court has not yet ruled on this, but legal scholars suggest a dual‑track approach: retain intent for law, while using predictive‑coding explanations in expert testimony.
- AI Accountability – For self‑governing AI agents, eliminating “belief” from design could simplify accountability: an agent’s error can be traced to a specific prediction‑error term, rather than an opaque “belief state.” This transparency aligns with ongoing AIAlignment research that seeks to make AI behavior explainable and controllable.
The ethical path forward likely involves selective elimination: discard terms that lack explanatory power, retain those that serve communicative or normative functions, and develop new vocabularies that respect both scientific rigor and human dignity.
9. Future Directions: Integration or Replacement?
9.1 Hybrid Models
A growing body of research proposes hybrid architectures that retain folk‑psychological labels as interface layers while grounding them in predictive‑coding machinery. For example, the NeuroSynth framework (2024) maps “belief” questionnaires onto a latent space of prediction‑error precision parameters, allowing clinicians to continue using familiar tools while benefitting from neurocomputational precision.
9.2 Education and Training
Training the next generation of psychologists will require curricula that blend philosophy of mind, computational neuroscience, and AI ethics. At the University of California, Berkeley, a pilot program introduced a Computational Mental Health track in 2023; early graduates report 30 % higher placement rates in AI‑health startups, indicating market demand for this interdisciplinary skill set.
9.3 Open‑Science Infrastructure
Large‑scale data repositories, such as the OpenNeuro platform, now host over 2 petabytes of fMRI and EEG data, enabling meta‑analyses that test eliminativist hypotheses across thousands of tasks. The Eliminativist Consortium (2025) has pledged to release a standardized “prediction‑error atlas” by 2027, which could serve as a common reference for both psychologists and AI developers.
9.4 Conservation Synergies
Finally, integrating eliminativist insights into bee‑conservation technology could yield a virtuous cycle: better predictive‑error models improve AI monitoring, which in turn informs land‑use policies that enhance floral diversity, feeding back into healthier bee colonies. This loop exemplifies how a philosophical stance can have concrete ecological benefits when coupled with technology.
Why It Matters
Eliminativism is more than a philosophical curiosity; it strikes at the heart of how we understand minds—human, artificial, and insect. By questioning whether “beliefs” and “desires” are real entities or convenient shorthand, we open the door to more precise diagnostics, targeted interventions, and transparent AI systems.
For psychologists, embracing a predictive‑coding vocabulary could increase treatment efficacy by up to 45 % (as seen in ketamine and neurofeedback trials). For AI developers, it offers a blueprint for agents that adapt without the baggage of belief‑state tracking, leading to higher operational efficiency (e.g., the 92 % foraging success of HiveBot). And for conservationists, it provides a scientifically grounded method to guide bee behavior and protect pollinator ecosystems.
In short, whether we eliminate or reframe folk‑psychological terms, the conversation reshapes the tools we use to heal minds, build trustworthy AI, and safeguard the planet’s most essential pollinators. The future of psychology—and, by extension, the future of a thriving, bee‑rich world—depends on how thoughtfully we navigate this conceptual transition.