Against (theory-neutral) method (in consciousness science)

Abstract

This article challenges the assumption that the science of consciousness can proceed from a theory-neutral foundation. I argue that even ostensibly theory-neutral (or theory-light) programmes inevitably rely on substantive background commitments that cannot be cleanly bracketed. The analysis demonstrates that the aspiration to eliminate or minimize theory-dependence in favour of pure observation risks collapsing into naïve empiricism. More broadly, the paper contends that there is no context-independent scientific method—certainly not one that seeks to purge theoretical commitments from the neuroscience of consciousness without significant epistemic cost.

Highlights

• Theory-neutral or theory-light Natural Kind Reasoning in consciousness science is untenable, as empirical observation and measurement are irreducibly shaped by substantive theoretical, conceptual, and institutional commitments.

• Attempts to minimize theory-dependence risk collapsing into naïve empiricism, thereby discarding indispensable explanatory resources without delivering a genuinely neutral or unified methodology.

• There is no context-independent or universally legitimate scientific method for the neuroscience of consciousness; methodological pluralism and theory-ladenness are unavoidable and epistemically productive.

Introduction

The scientific study of consciousness is methodologically complex, shaped by deep theoretical disagreements stemming from both its nature and the means by which it should be measured. Competing frameworks—such as Global Neuronal Workspace Theory (GNWT) (Dehaene et al. 2006), and Integrated Information Theory (IIT) (Tononi 2017), etc.—differ not only in explanatory commitments but also in the theoretical implications they ascribe to empirical findings. Each framework thereby influences how experimental results are interpreted and which measures are considered evidentially significant. In response to theoretical fragmentation, theory-light or theory-neutral versions of Natural Kind Reasoning (NKR) have been proposed (Bayne and Shea 2020, Birch 2022, Mckilliam 2024). These approaches aim to identify stable, theory-independent categories of conscious phenomena through the empirical clustering of observable features. This identification process is understood to be iterative and experimentally driven, rather than guided by theoretical frameworks. The methodological hope is that by allowing measurement practices to converge empirically, theoretical bias can be mitigated. The central contention of this article is that the promise of theoretical neutrality does not hold: in the neuroscience of consciousness, there is no single methodology one can rely on, and efforts to minimize theoretical bias cannot escape the inevitable plurality and contestability of scientific practices. Theory-light or ostensibly theory-free approaches, in discarding essential theoretical resources in the pursuit of neutrality, succumb to a phenomenally naïve empiricism. This article thus argues that, in this context, any attempt to establish a universal methodology—even one aimed merely at dispensing with cumbersome theoretical commitments—fails to advance the scientific investigation of consciousness.

To substantiate its claim, the article draws on non-empiricist accounts of scientific evidence, which reject the notion that evidence consists of unmediated perceptual data (Bogen et al. 1988; Bird 2022, 2023), emphasizing instead that it emerges from theory-laden, socially embedded epistemic practices. To support this view, the article first argues that perceptions—and hence scientific observations—are inherently shaped by conceptual and institutional commitments and are cognitively penetrated (Hohwy 2014; Pezzulo et al. 2024). Furthermore, adopting the Bayesian adversarial collaboration framework, it argues that in the context of experimental validation, theoretical commitments both guide and constrain experimentation in consciousness science (Del Pin et al. 2021, Corcoran et al. 2023, Melloni et al. 2023). While this article’s insistence on the indispensability of theoretical commitments aligns with recent efforts that underscore the need for explicit conceptual frameworks in consciousness research, it advances the discussion by adopting a distinctly Feyerabendian stance: there can be no universal or singular methodology capable of legitimately regulating the relation between theory and observation in the neuroscience of consciousness.

The paper is structured as follows: section ‘Natural kind reasoning’ provides an overview of NKR in consciousness science. Section ‘Theory-ladenness of experimentation’ substantiates the paper’s critique by drawing on progressive cognitive science to challenge the theory-neutrality of observation, while also demonstrating that experimental design is guided and constrained by theoretical commitments. Section ‘Demarcation and discrimination’ explains why the assumption of theory-neutrality is untenable in the context of NKR, particularly given the impossibility of demarcating and discriminating between legitimate and illegitimate theoretical influences on observation. As the paper’s coda, Section ‘Coda of the paper: against method’ sets out its Feyerabendian position.

Natural kind reasoning

A central methodological challenge in consciousness science is how to empirically test consciousness, with experimental paradigms designed to adjudicate between competing theoretical frameworks. NKR has been proposed with the aim of identifying the natural kind of phenomenal consciousness (Block 2007, Shea 2012). The problem though is that competing theories—such as Global Workspace Theory (GWT), which attributes consciousness to the global broadcasting of information across cognitive systems (Baars 2005); and IIT, which defines consciousness in terms of the system’s capacity to integrate information into unified wholes (Tononi 2015)—emphasize fundamentally different explanatory targets, resulting in divergent empirical interpretations. This fragmentation is further intensified by a form of circularity in the evaluation of evidence, where empirical results are routinely filtered through pre-existing theoretical commitments. As a consequence, discordant findings are frequently reinterpreted or marginalized, with the result that validation tests become intrinsically theory-dependent rather than neutral arbiters of competing claims (Phillips 2018, Bayne et al. 2024). This fragmentation undermines the prospects of identifying a unified natural kind of consciousness, given its reliance on background theories that diverge in their ontological commitments and explanatory aims.

Bayne and Shea’s (2020) articulation of NKR seeks to resolve the problem of deep theoretical disagreement in consciousness science by employing independently validated operationalizations of consciousness, converging on a common explanandum, without presupposing any particular theory of what consciousness is. Bayne and Shea propose that the clustering of various marks—dispositions that reliably co-occur—provides evidence for an underlying natural kind of consciousness. They suggest that this natural kind may arise as a homeostatic property cluster or from a shared causal history, emphasizing flexibility in how consciousness is conceptualized, independent of specific theoretical assumptions. While they do not deny that NKR could go beyond pre-theoretic indicators of consciousness, they suggest that it is possible, to some non-negligible extent, to ‘begin with a variety of dispositions and capacities (“marks”) that are putatively associated with consciousness. The marks will include the various pre-theoretic indicators of consciousness, such as verbal report and volitional behavioural control’ (Bayne and Shea 2020: 70). In this vein, they argue that NKR can overcome the limitations of existing approaches, which are either tied to pre-theoretical notions or rely on contested theories.

Theory-ladenness of experimentation

To appreciate the contention of this article, it is crucial to distinguish raw data—i.e. unprocessed observational outputs—from what scientists treat as the explananda of their theories. In this context, a phenomenon is an epistemic product of disciplined abstraction rather than a single unmediated observational encounter (Bogen and Woodward 1988, Woodward 2009). This usage diverges from familiar talk in philosophy of mind and consciousness science, where ‘phenomenon’ often refers to subjective experiential states or qualitative features of consciousness. However, phenomena in the methodological sense (i.e. in the Bogen–Woodward sense) are not subjective experiential states, but the relatively stable, repeatable patterns abstracted from many heterogeneous data points. The claim is not that such patterns fail to exist in the world, but that their identification as scientific phenomena requires substantial inferential and modelling work—for example, decisions about calibration, noise treatment, background conditions, and the construction of low-level models that filter and organize the data into a coherent target of explanation. Also, science is not best understood as a repository of abstract models aimed solely at truth, but as a cognitive and social enterprise oriented toward expanding the epistemic resources of a community. The weight of claims—such as those concerning unobservable entities or natural kinds—derives not from individual priors in isolation but from shared standards of inference, calibrated by rational inquirers who remain responsive to the available evidence. Taken together, these considerations show that it is neither feasible nor coherent to draw a principled wedge between theories on the one hand and observation, measurement, and experimentation on the other. In what follows in this section, I articulate these points by drawing on case studies of salient scientific breakthroughs.

The science of perception: perception as theory-laden inference

Here, I elaborate on the claim that it is phenomena, not data, that figures directly in scientific practice. To suggest that theories can be dispensed with in the science of consciousness—or in science more generally—is to imply the possibility of observations untainted by theoretical commitments. In contrast, I argue that, given the top-down character of perception (which is the basis of observation), observational reports are already shaped by the brain’s hypotheses about the structure of the world and, indeed, about the agent’s own states. According to this proposal, perception does not begin with raw sensation; it is driven by pre-existing hypotheses encoded in generative models—structured, hierarchical, top-down representations of the world instantiated in neural circuitry—and consists in a continuous, reciprocal interaction between top-down predictions and bottom-up sensory input. These hypotheses specify probabilistic expectations about causes of sensory data, and incoming signals serve to update those expectations through error-driven inference. This mechanistic dependency renders scientific reasoning at the cognitive level inherently theory-dependent. This is particularly evident in hierarchical predictive processing models based on the Free Energy Principle (FEP), where perception operates as an inferential process guided by prior beliefs and structured hypotheses rather than passive sensory reception (Friston and Frith 2015b, Friston et al. 2017, Pezzulo et al. 2024).

Within the FEP framework, the brain continuously generates predictions and selects the most plausible interpretation of sensory input. The discrepancy between predicted and actual sensory signals—registered as prediction error—drives this inferential process, ensuring that perception optimally balances prior expectations with incoming evidence. The imperative to minimize prediction error drives perception through two primary mechanisms:

Perceptual inference: explaining how the brain refines its internal model to accommodate unexpected sensory input, thereby adjusting its ‘theory’ of the world to improve future predictions. This corresponds to Bayesian belief updating and can be interpreted as a cognitive-level form of hypothesis revision.

Active inference: explaining how the organism modifies its behaviour to align sensory input with prior expectations, effectively reducing surprise by making the world more predictable. This is evident in sensorimotor control, where actions are selected to minimize prediction errors, making perception an enactive process rather than a passive one.

This makes perception itself a continuous, sub-personal form of explanatory inferences, where the brain selects the ‘best explanation’ for its sensory input based on the available hypotheses within its generative model (Hohwy 2013, 2014, Pietarinen and Beni 2021). This means that perception is a cognitive (hypothesis-laden, if you will) process, structured by implicit, probabilistic hypotheses about the external world. Sensation does not precede theory; it is already theory-laden. It follows, first, that there is no such thing as raw sense data according to viable theories of perception; perception itself is hypothesis-driven and cognitively penetrated. The model-based hypothesis-laden nature of perception, as formalized in predictive processing and active inference, ensures that all perceptual data are already structured by implicit models of the world. As a result, explanatory inferences—both at the cognitive level and within scientific methodology (The same model-based structures that guide perception also underlie scientific theorizing: cognitive generative models are extended and formalized into scientific models, which inherit their dependence on prior assumptions and structural commitments. This continuity makes scientific inference no less theory-laden than perception itself, a point developed further in section ‘Coda of the paper: against method’)—cannot escape their dependence on theory. This reinforces the earlier point (drawn from Woodward 2009) that what matters in scientific practice are not raw data but phenomena, since perception itself delivers information already shaped by generative models and inferential assumptions; what scientists work with are these structured, theory-mediated constructs rather than unprocessed sensory inputs. Theoretical assumptions pervade all levels of explanation, from sub-personal perceptual mechanisms to the inferential choices guiding scientific practice. The aspiration for a purely empirical, theory-free approach to identifying the natural kind of consciousness appears untenable.

Against the backdrop of this section, it may be contended that the FEP primarily generates hypotheses about the causal structure of an organism’s environment, rather than specifying a theory of consciousness per se. Proponents of NKR may build on this to argue that the critique advanced in this section does not directly concern theories of consciousness or their associated experimental practices. However, my account of the relation between theories and observations under the FEP extends directly to theories of consciousness, since organisms operating under the FEP do not merely construct models of their environment but also continuously model the consequences of their own actions upon it, including the sensory and cognitive effects of those actions. This recursive self-modelling not only generates structured expectations about external states but also about the agent’s own internal processes, thereby embedding representational commitments about experience within the generative model itself. This bidirectional inferential loop forms a natural foundation for a form of deep computational neurophenomenology, whereby first-person structure and third-person measurement are systematically intertwined, both instantiated and interpreted within the same inferential architecture (Sandved-Smith et al. 2025). Under the FEP, the same principles that govern adaptive action and perception also apply to consciousness, with the difference that modelling conscious experience occurs at a higher level of complexity. Self-organizing systems do not merely predict the immediate consequences of environmental states; more sophisticated systems maintain temporally extended generative models that integrate information about their own actions and simulate counterfactual outcomes (Friston et al. 2018). In this respect, consciousness emerges as the capacity of the system to perform temporally deep inference about itself as an embodied and temporally extended agent, where self-evidencing encompasses not only environmental contingencies but also the structure of subjective experience (Friston 2018, Solms and Friston 2018). Because organisms under the FEP actively model both the environment and the consequences of their own actions, measurement outcomes will systematically reflect those predictive models rather than raw, theory-free signals.

Perhaps the more pressing question, however, is how the hypotheses the brain generates about the environment and the self are related to the theoretical commitments that shape scientific practice in consciousness research. Scientific theory-making is not insulated from, but continuous with, the cognitive modelling that organisms employ in navigating their world. The inferential architectures that underpin self-modelling and environmental modelling provide the cognitive basis upon which scientific theorizing is constructed (Beni 2019, 2024). Below, I shall elaborate.

The transition from brain-level hypothesis to scientific hypothesis proceeds through a sequence of operations. Individual agents encode recurring patterns and contingencies in embodied, tacit models; these models instantiate regularities at multiple temporal and structural scales. Scientific activity selects, amplifies and externalizes particular regularities from this tacit store by translating embodied expectations into publicly manipulable forms (It is also worth mentioning that even at the level of basic perceptual and practical cognition, hypothesis formation is not an intrinsically individualistic process. From the start, the neural systems that generate perceptual expectations are calibrated by structured social signals—facial orientation, gaze direction, vocal prosody—and by participation in coordinated activities. Implicit action-understanding mechanisms map others’ movements onto one’s own motor repertoire, providing a channel through which observed regularities are encoded as internal predictions; higher-order mentalizing systems attribute goals and dispositions, enriching the predictive space with socially inflected causal structure (Frith 2008, Frith and Frith 2012, Friston and Frith 2015a). These systems operate within environments densely saturated with artefacts, routines and interactional norms that structure the informational inputs available to the developing agent. Consequently, the priors that govern even basal hypothesis formation reflect the distributed cognitive ecology in which the agent is embedded: representational states are co-constructed across individuals, material media and patterned practices, rather than generated solely within individual neural systems (Hutchins 1995). The same socially mediated and materially scaffolded mechanisms that bias perceptual modelling—coordination, representational propagation, transformation across media, and collective calibration—are redeployed at larger scales within scientific practice (Knorr-Cetina 1999, Beni 2021). Coordinated roles, hierarchies, and division of epistemic labour implement distributed computations that no single investigator could reliably execute.)—symbols, protocols, instruments, and explicit models. This provides one possible precise articulation of how organisms generate, update, and deploy internal models, but the explanatory role they occupy in the present account can be filled by any defensible theory of cognitive function (apart from the FEP) that treats cognition as the construction and refinement of structured representational capacities. Historically, as I will explain in section ‘Coda of the paper: against method’ of this article, closely related proposals were formulated in connectionist terms by Paul Churchland (1989), who likewise argued that scientific theorizing develops by sharpening and externalizing the multi-layered representational resources already operative in subpersonal cognition. The details of the underlying cognitive architecture—predictive hierarchies, distributed networks, or other mechanistic formalisms—are secondary to the general constraint that scientific practice inherits its basic inferential machinery from the cognitive systems of individual agents. Accordingly, the continuity claim I advance is robust across competing models of cognition: what matters is that scientific hypotheses arise from the systematic externalization, stabilization, and communal calibration of the same modelling capacities that support everyday perception, action, and self-maintenance.

This account does not equate a perceptual prediction with a formulated scientific theory; rather, it locates scientific theories as successive, institutionalized elaborations of the same modelling capacities that underwrite perception. The practical criteria that distinguish scientific hypotheses—explicit articulation, formal testability, instrumentally extended measurement, and communal validation—arise during the externalization and social-correction phases, not prior to them. Nevertheless, because the selection and initial framing of experimental contrasts, measurements and interpretive categories are constrained by the prior expectations embedded in perception and everyday cognition, the ideal of theory-neutral observation is undermined: the very inputs that feed experimental design and data interpretation are already structured by antecedent hypotheses. Hence the continuity between brain-level modelling and scientific theorizing licences a principled form of theory-ladenness for experimental practice without collapsing science into folk psychology. In this sense, the theoretical frameworks adopted in consciousness science can be seen as elaborations of the same modelling capacities, extended and formalized through collective inquiry. This continuity suggests that attempts to insulate scientific methodology from theoretical commitments are misguided: the commitments are already embedded at the cognitive level. I return to this theme in section ‘Coda of the paper: against method’.

Accounts of theory testing in consciousness science

One’s stance on experimentation in consciousness science is shaped not only by one’s theory of observation but also by one’s theory of experimental validation. Contra the theory-neutral version of NKR, which rests on a naïve theory of confirmation—one in which theories are treated as passive entities that do not shape the trajectory of experimentation—this section adopts an alternative model of theory testing grounded in Bayesian adversarial collaboration. In this view, theories such as IIT and GWT are not static constructs awaiting definitive verification or falsification. Instead, they operate as evolving frameworks that generate testable, often quantitative, predictions, thereby actively structuring experimental design and interpretation. As new data are acquired, degrees of belief in these competing theories are updated incrementally through Bayesian inference, reflecting a dynamic and ongoing process of theory revision rather than a search for final confirmation (Corcoran et al. 2023). This proposal is consistent with the collective epistemology outlined in Bird’s (2022, Chapter 8) account of explanatory inferences and abduction as a socially distributed Bayesian process embedded within the broader scientific community.

In Chapter 8 of his 2022 book, Bird argues that the Bayesian inferences central to scientific reasoning should not be seen as purely personal but should instead reflect the distributed and communicative nature of scientific practice. He suggests that the relationship between evidence and hypothesis is not contingent on individual perspectives but is governed by intersubjective norms that are accessible to all participants in the scientific community. To my mind, this perspective aligns well with the adversarial collaboration proposal for experimentation in consciousness science (Olcese et al. 2021, Corcoran et al. 2023, Melloni et al. 2023, Beni 2025). In this view, experimental phenomena—such as expectations about neural signatures and connectivity patterns traced through neuroimaging and electrophysiological techniques—are not tested in a theoretical vacuum. Rather, they are evaluated against the theoretical backgrounds of specific theories, such as IIT or predictive processing, ensuring that experimental results are interpreted within an appropriate theoretical context. This view underscores the central role of theories embedded in priors, emphasizing that all experimental outcomes undergo Bayesian updating. However, the priors—the initial weight assigned to each theory's predictions—are shaped by the researcher's theoretical commitments, in the sense that ‘competing theoretical hypotheses of varying degrees of specificity (precision) can be encoded as prior constraints over the parameters of a generative model prescribed by each theory’ (Corcoran et al. 2023: 3505).

To recap, the testing of theories of consciousness is not a mechanical comparison of predictions against data but an interpretive activity situated within established conceptual frameworks. And, far from being theory-blind, the interpretation of experimental results is fundamentally influenced by the theories embedded in these priors. For instance, a researcher with strong priors in IIT will update their beliefs more significantly in response to integration measures, while a proponent of GNWT may interpret the same data as less decisive, particularly if frontoparietal broadcasting and long-range synchronization fail to meet the thresholds expected by the theory. This is because ‘Bayesian inference [when cast into the adversarial collaboration framework] permits adversarial parties to ascribe different prior beliefs about the probability of their favored theoretical hypotheses, leading to different posterior beliefs—i.e., belief updating—in the face of the evidence’ (Corcoran et al. 2023: 3507). This process highlights how priors encode background assumptions and determine the epistemic weight of new evidence, reinforcing the non-neutrality of the scientific process. Overall, this proposal underscores the idea that the significance of empirical data is not inherent but contingent upon the theoretical framework guiding its interpretation, and thus it speaks directly to a non-empiricist perspective on scientific knowledge, according to which scientific inferences do not serve as a theory-independent bridge between data and explanation. Instead, it functions within, and is constrained by, the internal norms and assumptions of the theoretical framework to which it is applied. Below, I shall delve into further details.

Under the adversarial collaboration framework, the experimental design is not intended to remove theoretical bias but rather to adjudicate between competing theoretical constructs through targeted empirical tests. In this context, introspective reports are not treated as neutral observational givens; rather, they function as data embedded within a theory-laden inferential architecture. Melloni et al. (2023) exemplify this approach by explicitly structuring their experimental paradigm around five theory-derived predictions aimed at distinguishing between IIT and GNWT; IIT predicts that the neural correlates of consciousness (NCC) are primarily localized within the posterior ‘hot zone’—comprising parietal, occipital, and lateral temporal regions—whereas GNWT posits that every conscious experience is accompanied by activation of a fronto-parietal network, including the prefrontal cortex, in conjunction with high-level sensory cortices. Accordingly, consistent identification of primary neural activity correlating with conscious experience in the posterior hot zone would support IIT, whereas evidence favouring fronto-parietal activation patterns would be taken as corroborative of GNWT. Similarly, IIT predicts that the contents of consciousness should be maximally decodable from posterior cortical regions, on the grounds that conscious content is intrinsically specified by the integrated cause-effect structure localized within those areas. In contrast, GNWT predicts that information about the content of experience should be decodable from both the fronto-parietal network and high-level sensory cortices, reflecting the global broadcast of information characteristic of conscious access.

To be clear, there are, therefore, two mutually reinforcing loci of theory-ladenness in this context. First, perceptual priors and embodied expectancies—the sub-personal predictive dispositions that regulate attention, salience attribution, and the parsing of causal structure—selectively determine which experimental contrasts are even intelligible as candidates for evidential discrimination. This establishes a baseline constraint on inquiry: the space of possible hypotheses is already sculpted by the cognitive machinery responsible for modelling the environment. Secondly, the scientific theories of consciousness (IIT, GNWT) constitute institutionalized, formally articulated extensions of precisely those modelling mechanisms. As argued in section ‘The science of perception: perception as theory-laden inference’, such theories do not arise ex nihilo; they are higher-order codifications of the same inferential architecture that governs everyday predictive cognition. Their role within scientific practice is therefore doubly loaded: they inherit the epistemic contour of sub-personal model formation and, in virtue of their formalization, impose additional normative structure on experimental design.

The main point here is that adversarial collaboration framework does not dispense with theoretical commitments; rather, it operationalizes them as normative parts of experimental design, as, the very act of specifying which aspect of consciousness is targeted (state versus content, phenomenal versus access), translating this into concrete empirical criteria (e.g. decodable stimulus category from posterior cortex during no-report conditions), and preregistering contrasts and analytic pipelines amounts to embedding theory into the design (Del Pin et al. 2021, Consortium et al. 2023).

The Bayesian articulation of the adversarial collaboration approach further elucidates how theory-ladenness operates in the construction of empirical tests for consciousness. The key insight here is that prior expectations reflect structured beliefs derived from the generative models associated with specific theories, given that competing theoretical hypotheses of varying degrees are embedded as prior constraints over the parameters of a generative model prescribed by each theory (Corcoran et al. 2023). A proponent of IIT, for example, would encode higher prior probabilities for outcomes that reflect its theoretical commitments—for instance, that elevated pre-stimulus excitability or increased local synchrony within posterior cortical regions enhances the likelihood of conscious perception. These priors function as theory-constrained probability distributions over plausible explanatory models and are embedded within the generative architecture prescribed by the theory. Such theoretical priors do not merely influence the interpretation of results post hoc; they also shape the design of the experiment itself—determining what is measured, when it is measured, and how the data are modelled. As data are acquired, Bayesian updating proceeds via inference to a posterior distribution; however, this process remains constrained by the structure of the prior, ensuring that belief revision is guided by the theory’s explanatory architecture rather than proceeding in a theory-neutral fashion. Conversely, a researcher operating within the GNWT framework would assign higher prior probabilities to outcomes consistent with global broadcasting dynamics—for example, transient activation in prefrontal and parietal cortices, or decodable category-specific information in task-relevant prefrontal and high-level sensory regions. These priors similarly guide experimental design by specifying which neural signatures are relevant, how they should be temporally and spatially resolved, and what constitutes evidence for workspace ignition. In both cases, theoretical commitments are not external to the experimental process but function as structural constraints on both inference and operationalization. The experimental design thus reflects not evidential neutrality but structured contestation between incompatible theoretical priors. This harkens to the fundamental insight that theories in consciousness science do not merely explain data; they determine what counts as data, how it is interpreted, and what would constitute successful prediction or confirmation.

Demarcation and discrimination

In the previous section, I justified this claim by showing that experimental design in consciousness science is never theory-free: theories structure what is measured, how it is measured, and how results are interpreted. To ignore this is to forget that data are not self-interpreting but always filtered through generative models and background assumptions; in this light, NKR’s aspiration to neutrality effectively reinstates the myth of the ‘given’ (or raw data, or pure observation), the very stance dismantled by well-known critiques (Sellars 1956). In this section, I develop this point into a deeper critique directed at a more charitable—and more realistic—reading of the NKR project.

If NKR’s ambitious claims to theory-neutrality amount merely to retaining broadly acceptable theoretical commitments while discarding those deemed contentious or fine-grained, its contribution reduces to a restatement of the uncontroversial point that theory choice should be cautious. In that sense, its purported novelty dissolves into a methodological commonplace; the approach does not overcome theory-ladenness but merely draws an arbitrary line between what is treated as unproblematic and what is labelled problematic—a line that does no substantive work, because the interpretive flexibility that motivates NKR in the first place arises precisely within the supposedly safe background commitments.

With respect to my reference to an ‘arbitrary line’: it is the ostensible partitioning of methodological commitments into those treated as innocuous and those treated as theoretically encumbering, a partition effected by stipulation rather than by an independently justifiable criterion. This move can be read through Quine’s critique in Two Dogmas of Empiricism: Quine (1953) showed that the putative boundary between analytic truths and empirical hypotheses cannot be sustained without recourse to circular or conventional stipulations. Applied to NKR, the parallel is straightforward—declaring certain methodological assumptions ‘safe’ requires precisely the sort of tacit stipulation Quine diagnosed. Accordingly, the proposed demarcation does not track a principled difference in epistemic status but rather overlays a rule atop the very inferential apparatus whose neutrality is in question. If that sorting rule is treated as self-evident, the view collapses into a form of naïve empiricism—a position already widely judged untenable within the philosophy of science.

One can grant that certain background assumptions—for instance, that haemodynamic fluctuations provide a workable proxy for neuronal activity—are, in practice, widely shared and thus appear methodologically uncontroversial. Although these assumptions are often treated as routine, they do conceptual and inferential work—they determine how signals are interpreted, how models are fitted, and what counts as evidence for or against a given hypothesis (Henson et al. 1999, Larson-Prior et al. 2006, Astolfi et al. 2010). The main point here is not only that seemingly straightforward premises (e.g. that a measurement reliably reflects the underlying process) already presuppose theoretical commitments about adequacy, resolution, and lawful mapping, but also, more specifically, that routine practices—such as localizing functions to cortical regions—embed further assumptions about functional modularity and spatial segregation, i.e. that cognitive constructs can be cleanly instantiated in discrete neural regions, which is also open to debate (Hipolito and Kirchhoff 2019). Thus, what might be presented as a minimal or ‘safe’ assumption (e.g. that fMRI tracks neural activity) already shades into more contentious commitments (e.g. that such activity can be localized to discrete cortical modules). Of course, these are not the kinds of theoretical commitments that NKR theorists seek to eliminate; they are tolerated precisely because they are indispensable to the operation of the measurement apparatus. Yet, once such theory-ladenness is admitted at the ground floor of data acquisition, its influence is not easily contained. The inferential infrastructure built atop these premises inevitably inherits their issues. At each level, the transition from raw measurements to meaningful scientific phenomena presupposes substantive theoretical choices, with no clear demarcation between what is methodologically necessary and what reflects theoretical importation, meaning that demarcating which fluctuations count as noise and which as signal is not a neutral step, but a judgement shaped by a background theory (McAllister 1997). There is no principled criterion for separating coarse-grained assumptions (e.g. that haemodynamic responses track neural activity) from fine-grained ones (e.g. that a particular decoding signature reflects workspace ignition rather than metacognitive report). Hence the limitations of the pure insertion principle (The pure insertion principle presumes that adding or removing a single cognitive component leaves all other processes unchanged. Pace this principle, the brain is arguably a nonlinear, interactive system, and cognitive processes are not modular units that can be switched on or off in isolation. Instead, the engagement of one process typically modulates or reshapes the activity of others, producing interaction effects that undermine the logic of simple subtraction or insertion designs (Friston et al. 1996). Accordingly, experimental contrasts that rely on pure insertion risk misattributing neural activations, since the observed signal may reflect complex interdependencies between components rather than the isolated contribution of a single process. By the same token, attempts to isolate a specific cognitive process inevitably presuppose commitments regarding task decomposition, functional localization, and the influence of interacting components on the observed signal—assumptions that cannot be treated as uncontroversial or objectively self-evident (Roskies 2010, Schlicht 2018). What is framed as a neutral design choice (e.g. subtracting one task from another) already embodies substantive assumptions about modularity and independence, which are themselves contested. So, observed neural differences cannot be treated as direct evidence of the processes of interest, because neural activation patterns therefore do not straightforwardly correspond to theoretical constructs, but require interpretive mediation. Among other things, this is because the interpretation of activation patterns depends on a scaffolding of auxiliary assumptions, which makes their experimental individuation problematic, as there is typically neither an isomorphism nor even a robust homomorphism (a structure-preserving mapping) between the conceptual vocabulary of theories and the measurements produced by neuroimaging (Overgaard and Kirkeby-Hinrup 2021). To apply this into contemporary consciousness science, large-scale meta-analyses show that methodological choices alone—paradigm, report structure, and measurement modality—systematically predict which theory an experiment appears to support, even with constant results (Yaron et al. 2022). Theoretical outcomes are often foreseeable from the methods section, indicating that evidential salience is already conditioned by upstream inferential commitments. This directly challenges the NKR assumption that a theory-neutral evidential base can be isolated from method.

In the same vein, evidence that a neural region correlates with conscious processing—the candidate NCC—cannot by itself resolve theoretical disputes. This is because such correlations do not establish that the identified activity is either necessary or sufficient for consciousness, nor that it plays the same role across different experimental contexts (The conceptual vocabularies of competing theories of consciousness do not map neatly onto neuroimaging data, because there is no systematic or structure-preserving mapping (i.e. no one-to-one or reliably homomorphic correspondence) between theoretical constructs and activation patterns, and because the same neural signal can be interpreted as evidence for different mechanisms depending on the theoretical framework applied (Brett et al. 2002, Aru et al. 2012). Key constructs—such as conscious access, workspace ignition, or integration—remain underspecified, and identical conscious experiences may be realized by different neural mechanisms, making one-to-one mappings untenable. The relation between observed neural activity and conscious processes is indirect, mediated by auxiliary assumptions, and therefore open to competing interpretations. The implication is clear enough: the very same intervention can be read as minimally theory-laden or as deeply theory-dependent, depending on the interpretive framework.

All of the considerations rehearsed in this section harken to long-standing insights in the philosophy of science: what matters in scientific practice are not isolated data points but phenomena. Which aspects of the data are treated as explanatorily salient is never given in advance, but is determined by antecedent commitments that guide both the framing of experiments and the interpretation of their results(Shapere 1982, Bird 2023). Taken together, these considerations indicate that attempts to distinguish between theoretically dispensable and indispensable assumptions cannot be sustained, since the salience of any datum or pattern already depends on prior commitments about what is explanatorily relevant. On this basis, the project of a theory-light NKR framework fails, because its proposed division rests on the very theoretical mediation it seeks to avoid.

Coda of the paper: against method

Previously, it has been argued that in the context of theory appraisal, empirical results cannot serve as straightforward confirmations or disconfirmations of competing frameworks, since their evidential force is always mediated by the broader theoretical structures in which they are embedded. From a Lakatosian perspective, this amounts to recognizing that theories are not isolated hypotheses but elements of research programmes, consisting of a hard core of central commitments and a protective belt of auxiliary assumptions (Negro 2024). Accordingly, the appraisal of empirical findings is diachronic and contextual: results (as well as pieces of empirical evidence) acquire epistemic weight only when interpreted against the progressive or degenerative trajectory of the programme as a whole, rather than in isolation. The philosophical contribution of this paper is to extend this line of argument by adopting an explicitly (Feyerabend 2020) stance. While it affirms the indispensability of theories to experimentation, it channels this point into a more fundamental methodological claim: that any attempt to legislate a priori methodological criteria is untenable. Whether this takes the form of demarcating progressive from degenerative strands of a research programme, or distinguishing theoretically significant commitments from dispensable ones, such prescriptions themselves rest on substantive theoretical commitments rather than on neutral methodological ground. The very notion that one can, in advance, specify clear criteria for retaining or discarding theoretical elements is not a neutral arbiter but an additional theoretical stance, and as such cannot serve as a universally binding rule of scientific practice. This conception of scientific practice turns on the relation between cognitive hypotheses instantiated in the brain and scientific theories. Below, I shall elaborate.

As noted in section ‘Theory-ladenness of experimentation’, hypotheses embedded in generative models are structured, hierarchical representations that encode relational and modal features and are continuously revised by Bayesian updating to minimize prediction error. Crucially, scientific theories are not ontologically distinct artefacts but formal, social externalizations of this same inferential architecture: they codify priors, representational formats and structural commitments at a supra-personal scale, using instruments, notations, and procedures to stabilize and propagate those commitments. Even simple organisms regulate their behaviour by minimizing prediction error: chemotactic bacteria, for instance, adjust their motion relative to nutrient gradients by updating rudimentary generative models that track discrepancies between expected and encountered chemical concentrations. More complex animals deploy richer hierarchical models: a hunting cat continuously predicts the trajectories of its prey, adjusting its own movements to minimize mismatch between expected and unfolding sensorimotor contingencies. In yet more sophisticated organisms, predictive control extends to counterfactual domains, with agents acting to minimize expected free energy by evaluating the anticipated sensory and epistemic consequences of future actions. Scientific communities can be understood as further extensions of this trajectory: collectives of predictive agents who coordinate their modelling capacities, externalize hypotheses into shared symbolic systems, and construct instruments and procedures to reduce uncertainty about mechanisms of action and prediction (Beni 2021, Beni and Friston 2024). Viewed in this light, scientific theories of cognition and consciousness are not sui generis intellectual artefacts but systematic externalizations of the same predictive architecture that governs adaptive behaviour across this spectrum. Accordingly, the assumptions embedded in perception—the priors that organize attention, salience, and causal parsing—directly condition which contrasts are conceived as informative, how measurements are framed, and which explanatory models are pursued. Observations and experimental results cannot be read in isolation from these inferential antecedents: scientific hypotheses are successive externalizations and formal refinements of the brain’s internal hypotheses about environmental structure and causal mechanism. And therefore, if perceptual experience is theory-laden, scientific experimentation is ipso facto theory-laden: the laboratory is a scaled, instrumentally disciplined continuation of the same cognitive modelling enterprise, not a domain in which theory can be cleanly bracketed.

Observing a scientific theory or experimental result in isolation conveys little; it is inseparable from the inferential structures from which it emerges. This continuity undermines the very premise of NKR, since any attempt to purge theory from experimentation ignores the cognitive architecture that makes experimentation possible in the first place. From this continuity, it follows that if perceptual experience is theory-laden, so too is scientific experimentation, since the latter emerges from the same cognitive mechanisms that structure perception.

On a historical note, Paul Churchland (1989) was among the first to articulate this continuity explicitly, reframing the structure of scientific theories in terms of connectionist architectures whose distributed, high-dimensional representational spaces could accommodate Feyerabend’s principle of proliferation. According to this principle, scientific progress depends on the systematic generation and preservation of competing theories, since only in the presence of alternatives can the evidential scope and explanatory power of any one framework be properly assessed. Subsequent developments in computational neuroscience and predictive coding provide a more sophisticated account of generative modelling than the classical connectionism Churchland employed, yet the underlying point remains: theories are best (or at least fairly enough) understood as cognitively grounded, dynamically updated representational systems. Now, interestingly enough, in the same paper, Churchland also sets his cognitive approach to scientific theorizing alongside Feyerabend’s methodological pluralism, using the former to explain why the latter prevails in scientific practice. On Churchland’s account, representational change in science follows the same learning dynamics that govern neural networks—weight adjustments, representational re-encoding, and the gradual reorganization of high-dimensional state spaces. These processes render theory change intelligible without invoking opaque forms of incommensurability, while simultaneously vindicating Feyerabend’s historical claim that scientific progress is driven by heterogeneous and competing frameworks. Proliferation thus appears as a functional consequence of the cognitive machinery underlying model formation, rather than a merely descriptive observation about scientific history. Feyerabend’s (2020) insight bears directly on current debates about the methodology of consciousness science as well: no single procedure or putatively neutral method can claim universal authority.

The vitality of inquiry depends on the coexistence of divergent models and the continual testing of their relative adequacy, and thus requires the rejection of methodological dogmas—including those that seek to exclude or minimize theory, or that presuppose (in the spirit of naïve empiricism) a clean division between indispensable and dispensable theoretical components. To insist on such divisions is not merely naïve, but actively obstructive, since it enshrines an illusion of purity at the expense of the pluralism on which science depends. From this perspective, NKR’s claim to separate what is retainable from what must be discarded rests on a dogmatic empiricist tenet (with a nod to Quine (1953)): the untenable assumption that theories are either dispensable or can be neatly divided into innocuous and problematic components. What is presented as theory-light or theory-free in NKR in consciousness research rests on a pretence: it presupposes that theoretical commitments can be separated into indispensable cores and dispensable auxiliaries. Yet, as argued above, empirical data are never pristine; they are always already theory-penetrated, both in their generation and interpretation, and theoretical resources are inseparable from the practices by which data are gathered, structured, and interpreted, and any attempt to legislate against them a priori enacts a fiction of neutrality that obscures the true conditions of scientific inquiry. The aspiration to theory-lightness thus amounts to a disguised endorsement of one set of commitments under the guise of neutrality. Against this, the present argument holds that the entanglement of theory and evidence is irreducible, and that the progress of consciousness science depends not on constraining theories in advance, but on their explicit articulation and critical negotiation. Recognizing the indispensability of theoretical commitments leaves us with a residual problem—how to adjudicate between rival accounts when evidential interpretation remains elastic. The view developed here offers no quick resolution of that underdetermination; indeed, the argument indicates that no philosophical or methodological recipe is likely to settle it. It is not remediable by imposing new constraints or purging existing theoretical resources. This resonates with the sober core of Feyerabend’s stance: inquiry cannot be regimented into a methodological template, and attempts to enforce one merely displace, rather than resolve, the sources of disagreement. The contribution of the present paper is thus diagnostic rather than prescriptive: it clarifies why underdetermination persists and why methodological purism cannot eliminate it. Progress in consciousness science depends not on the pursuit of neutrality, but on the explicit articulation and critical negotiation of theoretical commitments.

Author contributions

Majid D Beni (Writing—original draft [lead], Writing—review & editing [lead])

Conflict of interest

The authors declare that they have no conflict of interest.

Funding

Permission to reproduce material from other sources.