We welcome Haueis and Colaço’s argument that metabolic realities should both constrain and generate cognitive models. Their distinction between evaluative constraints that rule out implausible accounts and generative principles that guide plausible ones is especially useful. Their rejection of global glucose-depletion models of mental effort and their emphasis on local metabolic demands tied to specific neural substrates challenge any strong version of medium independence. We agree that the brain’s energy budget is not background scenery but an active constraint on sustainable information processing. Our claim is that cognitive load theory (CLT) provides a mature applied framework for testing and operationalizing these metabolic constraints in learning, instruction, and, increasingly, human–AI symbiosis. In that sense, CLT can help move metabolic realism from a valuable philosophical demand to a cumulative research program.

The key link is straightforward. The working-memory limits central to CLT are not just functional metaphors; they are the behavioral expression of the local bioenergetic constraints Haueis and Colaço describe (Sweller, van Merriënboer, & Paas, Reference Sweller, van Merriënboer and Paas2019). Working memory is the limited workspace for processing novel, biologically secondary information. When that workspace is overloaded, performance declines, and effort rises. CLT has long explained this pattern in terms of capacity limits and allocation trade-offs. A metabolic perspective helps explain why such limits exist. Local ATP availability, synaptic maintenance costs, and metabolic by-products place hard bounds on how many novel elements can be processed and integrated at once. Decades of CLT findings on mental effort, instructional design, and learning outcomes therefore offer a natural empirical test bed for metabolic models of cognition. Our recent synthesis extends this logic to AI context windows, treating them as bounded workspaces analogous to human working memory and showing that similar load-management principles may scale to hybrid systems (Wang et al., Reference Wang, Liu, Zou and Paas2026; see Figure 1). The parallel is not identity, but it is theoretically productive because both systems face finite processing budgets and must balance storage, selection, and coordination under those budgets.

Figure 1.

Working memory as the functional manifestation of local metabolic constraints. Left panel: human working memory (3–5 chunks) limited by ATP availability and synaptic maintenance costs. Right panel: AI context window (thousands of tokens) limited by energy per bit and attention dilution. Arrows show metabolic offloading via embodied actions (human) and shared workspaces/retrieval-augmented generation (human–AI symbiosis), converting high-ATP neural computation to lower-cost substrates.

Figure 1. Long description

Panel A: A diagram of human working memory. It shows a brain with a label indicating a limit of 3-5 chunks due to ATP availability and synaptic maintenance costs. Arrows indicate embodied offloading to lower-cost motor and perceptual substrates. Panel B: A diagram of an AI context window. It shows a window with thousands of tokens limited by energy per bit and attention weight. Arrows indicate metabolic offloading via shared workspaces and retrieval-augmented generation to lower-cost external substrates.

A first implication concerns secondary knowledge. Haueis and Colaço note that evolution has shaped neural efficiency. CLT formalizes this point through the distinction between biologically primary and secondary knowledge (Paas & Sweller, Reference Paas and Sweller2012). Primary knowledge, such as face recognition or basic language, is acquired with little conscious effort because evolution has already tuned the system for it. Secondary knowledge, such as reading, algebra, and scientific concepts, requires sustained high-cost processing in systems not optimized for these culturally recent tasks. Instructional designs that reduce extraneous load provide direct evidence of this metabolic burden. Replacing conventional problems with worked examples, for instance, reliably lowers reported mental effort while improving retention and transfer. These effects can be interpreted as metabolic optimization: fewer resources are wasted on redundant search, leaving more available for schema construction and automation. The primary–secondary distinction is therefore not merely psychological. It also marks differences in the bioenergetic demands of learning and in the kinds of instructional support culturally recent knowledge requires.

A second implication concerns embodied offloading. Haueis and Colaço argue that metabolic demands tie information processing to particular substrates. Recent CLT work on embodied cognition shows how learners can exploit this constraint rather than merely suffer from it (Zou et al., Reference Zou, Zhang, Mavilidi, Chen, Herold, Ouwehand and Paas2025). Gestures, object manipulation, and other environmental interactions can offload computationally demanding operations onto motor and perceptual systems. When learners trace diagrams or manipulate concrete models, some of the burden otherwise borne by frontoparietal networks can be redistributed. Experimental work shows that such strategies often reduce mental-effort ratings while improving immediate performance, retention, or transfer. On this view, embodied actions are not peripheral aids but metabolically efficient ways of reorganizing computation across the brain, body, and environment. We have argued that shared workspaces and retrieval-augmented generation play a comparable offloading role in human–AI systems by reducing avoidable load on both the person and the model while preserving performance.

A third implication is methodological. CLT’s instructional-efficiency metric can be interpreted in the ATP/bit terms proposed by Haueis and Colaço (Paas & van Merriënboer, Reference Paas and van Merriënboer1993). For decades, CLT research has combined performance with mental-effort ratings to estimate how efficiently instruction turns invested effort into learning gains. Reframed biophysically, this becomes an approximation of energy spent per unit of durable, retrievable information. That translation is indirect, but it generates clear predictions: designs that minimize unnecessary processing should yield better efficiency profiles, and hybrid human–AI pipelines should benefit from dynamic load balancing when either partner approaches overload. CLT intervention studies already provide many opportunities to evaluate such predictions across formats, sequencing principles, and coordination structures.

This also suggests concrete tests of the evaluative standards Haueis and Colaço advocate. Randomized studies that manipulate element interactivity, format, or modality while measuring effort, performance, and physiological proxies such as pupil dilation or EEG could assess whether a cognitive model respects local bioenergetic constraints. When an instructional manipulation reduces effort without impairing learning, it offers indirect evidence for a more favorable energy-information trade-off. Future work could tighten this bridge by combining CLT paradigms with direct metabolic imaging, including in hybrid systems. In short, CLT does not merely illustrate the target article’s proposal. It offers an empirical and interventional program for turning metabolic considerations into testable predictions about learning, instruction, and coordinated human–AI cognition.