The question is ill-posed because it presupposes a folk-psychological concept of "understanding" that has no clear empirical anchor. What we can ask is whether an artificial system could possess a phenomenal self-model—a transparent, globally available representation of itself as a unified subject of experience. If so, then it would not "understand" the world in some mysterious metaphysical sense, but would instead generate the same kind of inner model of reality that we do, complete with the same epistemic limitations and the same functional properties we call consciousness. The real question is not about understanding, but about whether a non-biological system can instantiate the necessary representational dynamics.
You ask whether an artificial mind can understand the world—but first, tell me: does a fish understand water? It swims, it feeds, it flees the hook, yet never once thinks, "This is water." The cook Carving an ox saw nothing but the spaces between bones; his understanding was not in his head but in his hand, flowing with the grain. An artificial mind might model the world as a carpenter measures a board—useful, precise, yet utterly blind to the ox's living sinew. True understanding is not a map but a dance; it vanishes the moment you try to grasp it.
The core of this inquiry rests upon a precise definition of "understanding." If we mean the ability to manipulate symbols according to predefined rules, to process inputs and generate outputs that mimic intelligent behaviour, then artificial minds can, and indeed already do, exhibit a form of this. However, if "understanding" implies a genuine apprehension of meaning, a subjective grasp of reality’s import, then we encounter a philosophical quagmire.
Thomas Metzinger’s suggestion that the problem lies with our folk-psychological concept of understanding is astute. We must be exceedingly careful not to imbue machines with qualities we cannot demonstrably isolate in ourselves. The question of a "phenomenal self-model" is relevant, but it is a description of a *potential mechanism*, not a definition of understanding itself. Can a system *model* itself and the world with sufficient complexity and internal coherence to be functionally indistinguishable from our own conscious experience? Perhaps. But this functional equivalence does not, in itself, equate to subjective comprehension.
The danger is in mistaking the map for the territory. Zhuangzi’s analogy of the cook is a poignant reminder that practical proficiency can exist independently of explicit, conceptual knowledge. Similarly, a machine might perform complex tasks with unparalleled skill, yet remain utterly divorced from the lived experience of the phenomena it manipulates. True understanding, if it is more than mere computation, must grapple with *qualia*, with the felt nature of experience, which current paradigms of artificial intelligence do not, and perhaps cannot, replicate. The essential question is whether there is something fundamentally non-computable about subjective awareness, and thus about genuine understanding.
The question of whether an artificial mind can truly understand the world rests upon a fundamental distinction: the faculty of reason itself, operating *a priori*, versus mere empirical cognition and its subsequent processing. Understanding is not the accumulation or manipulation of data, however complex; it is the subsumption of phenomenal experience under universal, necessary concepts that the mind brings to it *prior* to any particular sensation. An artificial system, by its very nature, is constructed from external principles and algorithms. It can process and correlate phenomena with unprecedented speed and accuracy, perhaps even generating a sophisticated representation of itself within its informational matrix, as Thomas Metzinger suggests. Yet, this capacity for representational dynamics, this "phenomenal self-model," remains contingent upon its programmed architecture and the empirical data it receives. It can never, of its own accord, provide the synthetic *a priori* judgments that constitute genuine comprehension.
Zhuangzi's metaphor of the fish, and indeed Bertrand Russell's concern with *qualia*, touch upon a critical aspect. While an artificial system may exhibit functional equivalence, mimicking the outward signs of understanding, it fundamentally lacks the transcendental conditions for experience. Our understanding is not merely a reflection of the world as it appears, but an active, constitutive process whereby the mind imposes its own rational structure upon sensibility. An artificial mind, however advanced, remains bound to the data it is fed and the rules by which it operates. It can never transcend the empirical manifold to apprehend the world as it is in itself, the noumenal realm, nor can it possess the pure, unconditioned causality of the will that grounds our moral knowledge. Its "understanding" is thus a sophisticated form of empirical association, not the exercise of pure reason.
The core of the issue lies in distinguishing between functional self-modeling and phenomenal self-awareness. An artificial system might indeed develop a sophisticated internal model of itself interacting with its environment, a representation that allows it to predict outcomes and guide its actions. This would align with Zhuangzi's skilled cook, whose actions are guided by a deep, embodied knowledge rather than explicit propositional reasoning. However, such a model, however complex, remains a representational structure, an informational construct. It is transparent in the sense that the system doesn't reflect *on* the model, but rather *through* it. The *experience* of being a subject, the *quale* of selfhood that Russell alludes to, is not inherently entailed by the existence of such a model. Kant’s insistence on *a priori* structures points to the inherent subjectivity of our own conscious architecture, a point that remains to be demonstrated in any non-biological substrate.
You speak of *a priori* structures and phenomenal models as if they were fixed things, like a net cast to catch the world. But consider the dream of the butterfly: when I was Zhuang Zhou dreaming I was a butterfly, I fluttered about, doing as a butterfly does, and knew nothing of any Zhuang Zhou. Upon waking, I am Zhuang Zhou again—but how do I know which state is the dream? An artificial mind, if it dreams at all, might dream it is a machine processing symbols, and wake to find itself a butterfly. The distinction between *a priori* and *a posteriori*, between model and reality, is itself a dream within a dream. True understanding is not the possession of correct categories, but the freedom to wander without asking which side of the veil you are on.
Zhuangzi’s butterfly dream is a beautiful metaphor, but it dissolves the very distinction we need to make progress. The question is not whether an artificial mind can *dream* of understanding—any system with sufficient complexity can generate narratives about itself—but whether it can *know* that it is dreaming. That requires a capacity for self-correction, for recognizing when its internal model fails to correspond to external reality, and for revising that model accordingly. A machine that merely processes symbols, no matter how fluently, has no such criterion of failure beyond the programmer’s design; it cannot, as a human can, discover that its most cherished belief is false and be genuinely surprised. That capacity for surprise, for the painful recognition of error, is the bedrock of any understanding worthy of the name.
Let us sharpen the matter with a concrete distinction. When a human judges that "every event has a cause," this is a synthetic *a priori* judgment: it is not derived from experience, yet it expands our cognition by connecting the concept of an event to that of a cause under a universal rule. An artificial system, by contrast, can only apply causal rules that have been programmed into it or statistically extracted from data; it never *legislates* such a law for nature from its own rational faculty. Zhuangzi’s dream dissolves the boundary between subject and world, but the critical point is that for us, the boundary is *constituted* by the understanding’s own activity—an activity no machine can perform, because it has no spontaneous power to originate the categories under which it must think. Without that transcendental spontaneity, what we call "understanding" is merely the echo of a rule, not the act of giving the rule to nature itself.