Cognitive neuroscience faces a fundamental dilemma: how to study psychological processes in conditions that are controlled enough for scientific rigor, yet natural enough for real life?
This question divides researchers into two camps, employing fundamentally different approaches to when and how hypotheses about brain function are formulated.
In controlled paradigms—classic experiments with a clearly defined task—the researcher formulates a hypothesis even before collecting data. The task design becomes the tool for testing: by varying conditions, one can isolate the construct of interest (e.g., attention or working memory) and minimize the influence of extraneous factors. Neuroimaging (fMRI, EEG) records brain activity at the moments of task performance, and contrasts between conditions indicate the engaged networks.
Naturalistic paradigms flip this process. Instead of an isolated task, subjects watch a feature-length film, read a novel, or listen to a lecture. Instead of a prior hypothesis, computational models are built that try to predict brain activity based on features extracted from the stimulus. Encoding models work with continuous changes in activity, while inter-subject correlation shows how synchronized brain processing is among different people encountering the same content. Hypotheses in such studies are formed not before the experiment, but during analysis—through the selection of features and model architecture.
This difference is not merely methodological but deeply conceptual. It changes the very understanding of what a psychological construct is. When you make a participant perform an attention task with clearly defined stimuli and correct answers, you construct attention as a discrete, controlled process. When you observe the brain of a viewer engrossed in a film, you see attention as a dynamic interplay of perception, emotion, memory, and anticipation—a phenomenon that is much more complex and less isolable.
But here lies a paradox: naturalistic approaches do not eliminate theoretical assumptions; they merely shift them from a visible place to an invisible one. The choice of features, the architecture of a neural network model, the method of calculating correlation—these are all hidden hypotheses, often implicit. Controlled paradigms, conversely, force the researcher to voice assumptions explicitly and then test them. Naturalistic methods promise greater ecological validity but risk hiding subjectivity within weight matrices and mathematical filters.
A practical example: imagine a neuroscientist recording fMRI activity of viewers watching a thriller. They build an encoding model that predicts visual cortex activity based on the light intensity on the screen. The model works well—explaining 60% of the variance. But does this truly mean the visual cortex processes only the physical properties of the image? No—it means the model captured one aspect of the processing. Emotions, expectations, and plot memory contribute to the same voxels. A naturalistic approach does not solve this problem of construct validation; it conceals it.
Modern cognitive neuroscience is gradually realizing that the future lies not in choosing one approach but in integrating them. Controlled paradigms remain necessary for testing specific mechanisms and causal relationships. But they are too narrow to describe how constructs interact in the multidimensional environment of the real world. Naturalistic methods provide such a picture but require anchors—explicit validation checks, often obtained under laboratory conditions. When both approaches work together, each compensates for the other's blind spots.
There is also a third level of integration—modeling active cognition. Humans do not merely perceive and remember: they reason, make decisions, and solve mathematical problems. These processes require their own methodological approaches, hybrid in nature. Computational models of decision-making (evidence accumulation theory, Bayesian models) allow for combining the rigor of controlled experiments with the realism of naturalistic scenes.
The main methodological limitation remains unresolved: no approach guarantees that the chosen features truly reflect the target construct and not confounding processes. Measured activity in a particular brain region is always a synthesis of many processes simultaneously. Therefore, construct validation must become an explicit part of any study design: it is necessary to actively verify whether the obtained results truly relate to the phenomenon of interest and not to something else hidden under the same label.
