The theory-driven approach cannot stand alone

Contemporary consciousness research features a plethora of theories built on logic or limited data, where proponents of each theory seek to provide empirical support for their theory, often resulting in heated debates. The latest example is the letter in which a group of authors claim that one theory, Integrated Information Theory, is pseudoscience1. Nevertheless, the same claims have been made against other theories of consciousness2. Here, we argue that it might be time for the field to take a larger perspective and consider supplementing the strong theory-driven approach with other approaches that have become increasingly viable in recent years.

While we believe that theory-driven approaches are useful and have led to many illuminating studies on consciousness, we will first outline some limitations of this approach and some of the paradoxes that it has also led to. Recently, researchers in the field of interdisciplinary consciousness studies were asked about ten theories in a survey3, and six of the theories were considered as probably or definitely promising by more than one third of surveyed researchers. Yet, none of the ten theories were considered promising by two-thirds, and apart from quantum theories of consciousness, no single theory was considered not promising by a majority of researchers. The field thus does not appear close to a widely accepted theory.

Many theories are conceptual, psychological, or philosophical in origin, and consequently, the core claims are cached in non-empirical terms. Specific constructs like the global workspace, metarepresentation, recurrent processing, or attentional amplification rarely relate straightforwardly to testable hypotheses. Instead, theories need to rely on auxiliary premises that connect their core claims to operationalisation and measurement-specific details. When empirical findings challenge a theory, these auxiliary hypotheses are revised and the core claims remain intact, a pattern in scientific practice identified by Imre Lakatos more than forty years ago4.

In addition, most research focuses on a single theory and examines only the compatibility of data with that particular theory, while comparison of two or more theoretical predictions is much rarer. This has led to paradoxical situations where proponents of several theories claim the same empirical evidence to support their position. The crux of this issue concerns how we can test theories themselves rather than auxiliary claims that can be revised in the face of conflicting evidence.

Within the theory-driven framework, one recent promising proposal has been to pit theories against each other in adversarial collaboration where proponents of each theory have to commit to certain predictions before the data is seen5. Nevertheless, the endeavour has yet to change the minds of scientists already endorsing a particular theory.

In sum, the theory-driven approach has spawned many interesting and illuminating studies of consciousness, but at the same time, it has not been successful in settling on a consensus theory, and there is a substantial risk that this will not happen in the near future. At the same time, the approach may have limited the scope of examination, and alternative approaches have received low priority despite being used in other fields of neuroscience and having the potential to expand our knowledge of consciousness considerably. Often, when one approach has not resolved a debate despite decades of activity, it can be useful to explore the opportunities afforded by other approaches.

Theories can be deconstructed to make progress

One alternative approach is to have lower-level constructs that can be identified in empirical data as the starting point rather than overarching theories. Identifying, mapping, and investigating these constructs will deliver a set of theory-neutral ‘building blocks’ from which a novel model can be constructed.

To understand what we mean by constructs in this context, consider that scientific theories rely on theoretical constructs, i.e., specific concepts that theories use to formulate their central claims. In consciousness science, theories employ high-level constructs like "global workspace" or "recurrent processing" to express their core assumptions about what makes mental states conscious. However, these theoretical constructs need to be connected to lower-level, neurally grounded constructs to enable empirical testing.

The construct-first approach analyzes these theoretical constructs systematically, clarifying how they relate to each other and how they can be operationalized empirically. For instance, the high-level construct of "global workspace" can be further explicated in terms of lower-level constructs such as signal strength, temporal stability, and spread of recurrent activity. Similarly, "local recurrence" and "metarepresentation" can be analyzed using these same dimensions, revealing that apparently distinct theories may actually share common underlying constructs at the neural level.

This deconstruction reveals that different theories of consciousness often occupy overlapping regions in a multi-dimensional construct space. By focusing empirical efforts on testing specific constructs rather than entire theory complexes, researchers can conduct more theory-neutral investigations. This approach reduces reliance on auxiliary hypotheses—the additional assumptions theories require to connect their core claims to testable predictions—and enables the discovery of which combinations of construct values correlate with the presence or absence of consciousness.

The phenomenon can be approximated iteratively

Another alternative approach is the iterative natural kind strategy6. Instead of developing theories of consciousness directly, the iterative natural kind strategy starts by determining under which conditions different kinds of systems are conscious. This involves refining indicators of consciousness and iteratively applying inferences to the best explanation to determine which systems are conscious and which are not. Ideally, this will reveal clusters of properties underlying consciousness in different types of systems. This, in the end, may yield a mature theory of consciousness. But insights about controversial cases (e.g., in disorders of consciousness) that are derived by iterative natural-kind reasoning can also help to improve theories of consciousness without requiring a complete understanding of consciousness. For instance, the iterative natural kind strategy is compatible with the construct first approach.

Data-driven models may be superior to theories

Yet another alternative is a primarily data-driven approach. In recent years, this type of approach has become increasingly feasible in multiple areas of science as big data has become more and more prevalent. The most optimistic advocates have argued that it could mean the end of theory-driven research, and that we no longer need to “know” the model as long as it works as intended. On this view, theories could be considered obsolete and most hypothesis testing unnecessary7. From this perspective, data is first, and the best model is built from patterns observed in the data.

The approach is, of course, not without flaws. For example, models built from data depend on the data from which they are built, along with scientific choices and non-scientific factors/circumstances8. Moreover, consciousness researchers could rightly argue that our field is a special case, where we cannot simply examine if “something works”, given that understanding is critical for many of the issues that make us care about consciousness in the first place. A data-driven model that we have no insight into may be less compelling than a theory-informed model when diagnosing consciousness. However, if data-driven models outperform theory-informed models dramatically across the board, then we will have to consider what insights might be hidden in them and how to extract them.

With rich data in place, a data-driven approach will allow models to be built that have not been theoretically proposed (yet), but it will also allow the identified models to be compared to existing theoretical models in terms of similarities and differences. We do not propose that this approach should stand alone in the study of consciousness, but we do remark that it appears particularly underdeveloped in our field and appears a promising avenue for future studies.

Implications, outlook, and likely counterarguments

Comparing and testing constructs may accelerate empirical consciousness science more than direct comparisons of theories. To illustrate, contrasting hypotheses about the NCC propose visual areas and prefrontal areas, respectively. However, activations in either area may also be taken as arguments for there being multiple generators of consciousness, or indicate that the NCC is task dependent, or even that both types of activations are spurious while the “real” correlate is elsewhere. Furthermore, often meaningful empirical predictions do not touch on the metaphysical aspect of these theories. For instance, even if NCCs can be identified in the prefrontal cortex, this does not suffice to demonstrate that consciousness arises as a consequence of higher-order thought. Such a finding can only support higher-order theory by infusing theoretical inferences that are not themselves open for empirical testing.

We propose to invest efforts in alternative approaches to the study of consciousness in addition to theory testing. This does not disregard theories of consciousness or pretend that such theories do not exist. Nor does it exclude that theory testing can be worthwhile. Rather, we express the optimistic standpoint that more consensus and progress may be achieved by supplementing current efforts with other approaches that may be construct-focused and more data-driven. At the same time, we propose that the current lack of consensus in consciousness research is a result of the idea that a direct comparison of theories with empirically opaque metaphysical claims itself can create progress. Concomitantly, building models from data allows examination of the full model space, and also refinements of theories in the light of new data obtained from empirical tests of constructs.