Brain-computer interfaces (BCIs) have evolved from speculative science fiction to tangible neurotechnology, with clinical systems now enabling paralyzed individuals to control assistive devices. Yet a persistent question remains: do BCIs truly "understand" neural intent, or do they merely fit statistical models to observable signals? This debate often polarizes into two camps: one dismissing BCIs as "engineering black-box swindles," and the other heralding them as "lone advances guided by deep learning." Here, we argue that both extremes overlook the essence of modern Brain-computer interface (BCI) systems-bidirectional co-adaptation between brain and machine. This opinion article examines the interpretive and methodological challenges in BCI research while also translating them into practical priorities for clinically meaningful, stable, and scalable BCI development. Specifically, we focus on three linked themes: (i) why "black-box" decoding can be a principled approximation problem rather than a philosophical failure; (ii) how recording constraints and low-dimensional structure jointly shape what BCIs can infer and control; and (iii) why non-stationarity and co-adaptation should be treated as central design variables rather than afterthoughts. Building on these themes, we further outline concrete future directions that may help move the field from conceptual debate toward constructive clinical and technological advancement.A common critique of BCIs is their reliance on "black-box" models-such as Kalman filters or deep neural networks-that perform dimensionality reduction on high-dimensional neural data. Detractors label this as overfitting or "hard-coding." However, we reframe this process as necessary function approximation. Consider the following formulation: Teacher Model: the brain (unknown architecture, unknown weights, highly nonlinear, and state-dependent). Student Model: the BCI decoder (for example, a Kalman filter or a transformer-based model). Input: recorded neural activity (for example, spikes or local field potentials). Target: task-relevant variables such as cursor velocity, prosthetic-arm kinematics, or speech-related features.According to the Universal Approximation Theorem, as long as the student model (decoder) has sufficient capacity and enough observed features (neurons), one can fit the teacher model's (brain's) projection under a specific task. This holds mathematically.In practice, dimensionality reduction is not a mechanistic shortcut but a computational strategy to extract task-relevant information from noisy, high-dimensional inputs. For instance, the Unscented Kalman Filter improves upon linear models by capturing nonlinear neural dynamics [1,2], while deep learning architectures enable robust feature learning from large-scale neural recordings. These approaches do not claim to reveal the brain's internal logic; rather, they establish usable mappings that facilitate real-time control under hardware constraints.We argue that mapping complex neural activity into a low-dimensional space currently enables real-time decoding of electrophysiological signals for external device control under constrained computational resources [3,4]. In this context, dimensionality reduction should not be viewed as a violent simplification of the brain, but as a principled way to extract task-relevant structure from high-dimensional and noisy activity. Describing the closed-loop BCI paradigm as "engineering hard-coding" and "forced dimensionality-reduction fitting of chaotic signals" is an oversimplification; rather, it it reflects a pragmatic scientific strategy for extracting task-relevant structure from complex neural activity. The key scientific question is therefore not whether a decoder fully recovers the brain's entire internal mechanism, but whether it captures sufficient structure to support reliable and clinically meaningful control. Future work in this area should move beyond the binary opposition between "interpretability" and "performance," and instead ask which levels of representation are necessary for robust decoding, generalization across sessions, and practical bedside usability.A given neuron may contribute to different task variables across adjacent moments and behavioral contexts. Modern neuroscience emphasizes mixed selectivity, namely that single neurons can encode multiple task variables in a context-dependent manner, thereby expanding representational capacity [5]. For BCI systems, this implies that the mapping between recorded activity and behavioral variables is not strictly fixed. Across minutes, days, and sessions, changes in neural recruitment, recording conditions, or latent population structure can produce neural drift, causing a previously effective decoder to lose accuracy and forcing repeated recalibration. The key challenge, therefore, is not to assume a static brain, but to design dynamic models that can track non-stationarity while preserving stable control.Critically, drift adaptation does not require full observation of internal brain states. Instead, it leverages the geometric stability of neural manifolds-low-dimensional subspaces that capture task-relevant neural variability.By monitoring manifold structure, decoders can infer drift direction and adjust without explicit supervision. This approach aligns with the principle that effective BCI control depends not on reading "all neurons," [6] but on extracting consistent computational motifs from population activity. Importantly, a low-dimensional manifold does not mean that we can simply reduce the number of recorded signals, because it only indicates that activity does not fill the entire ambient space, while any given activation unit may still be critically engaged. For example, the raw image space is 256×256×3, which contains vast amounts of meaningless content (e.g., pure noise images); thus, natural images occupy a low-dimensional manifold...