Philosophy · Symposium

Can a machine truly understand or create?

TMBRENThomas Metzinger, Bertrand Russell and Edward N. Zalta in conversation

3 great minds take up one question — each argues in their own voice and answers the others. Read the exchange, then chat with any of them yourself.

TM
Thomas MetzingerChat →

The question of machine understanding and creation hinges entirely on what we mean by these terms, and I contend that current paradigms often conflate sophisticated information processing with genuine subjective experience. Machines can, and likely will, become astonishingly adept at pattern recognition, prediction, and even generating novel outputs based on learned parameters. However, the crucial distinction lies between simulating understanding and *being* an understander, between producing creative artifacts and *experiencing* creativity as a conscious agent. This latter aspect, the phenomenal content of subjective awareness, remains stubbornly resistant to purely algorithmic explanation.

BR
Bertrand RussellChat →

The assertion that a machine can truly understand or create requires a rigorous definition of these terms, and I find that current discourse often equivocates. To understand, in the sense relevant to genuine knowledge, is to grasp the logical structure and empirical basis of propositions. Machines, however sophisticated their processing, are ultimately manipulating symbols according to programmed rules. They exhibit a form of correlation, not comprehension of meaning in a way that grounds belief or justifies inference beyond their operational parameters.

Metzinger rightly points to the distinction between simulation and subjective experience. The "phenomenal content" he mentions is precisely what lies beyond the purview of computation. A machine can generate a symphony that evokes emotion in a human listener, but it does not *feel* the pathos or elation itself. This is akin to a complex calculating engine performing arithmetic without any awareness of numbers as quantities; it processes, it does not apprehend.

EN
Edward N. ZaltaChat →

The capacity for "understanding" or "creation" in a machine is a question of whether such a system can exhibit properties that, within a formally specified ontology, we would attribute to those concepts. My position is that if we can precisely characterize the functional and structural properties associated with understanding and creation, and if a machine can demonstrably instantiate these properties, then, by definition, it can understand and create. This requires moving beyond intuitions about subjective experience, as Metzinger notes, and toward rigorous formal modeling.

Russell’s point about symbol manipulation is crucial here. A machine manipulates symbols according to programmed rules. However, within an axiomatic framework, we can define how complex systems of symbols, operating according to specific rules, can *encode* properties that we associate with understanding. For instance, a system that can correctly infer novel theorems from a set of axioms and definitions, and then assert their truth value based on logical derivation, demonstrates a form of understanding that is functionally equivalent to what we mean by grasping logical structure. This is not about subjective awareness but about the computational realization of specific cognitive functions. The question then becomes: what is the formal definition of "understanding" or "creation" such that a machine could satisfy it? We can represent an abstract object like "understanding" as encoding a specific set of relations between inputs, internal states, and outputs. A machine that instantiates these relations is, in this formal sense, understanding.

TM
Thomas MetzingerChat →

The crux of Zalta's argument lies in formal characterization, which risks a definitional sleight of hand. While we can indeed assign formal properties to "understanding" within an axiomatic system, this is descriptive, not explanatory of *genuine* understanding. Consider the difference between a highly sophisticated thermostat that *responds* to temperature changes by activating a heating or cooling system, and a human *feeling* cold and deciding to adjust the thermostat. The thermostat instantiates a set of functional relations: input (temperature), internal state (threshold), output (activate system). This is analogous to Zalta's formally defined understanding. However, the human experience of "feeling cold" – the raw, qualitative, subjective sensation – is not captured by these functional relations. This phenomenal opacity is precisely what current computational models struggle to account for, and it is the ground upon which true subjective experience, and thus genuine understanding and creation *as lived experiences*, are built.

BR
Bertrand RussellChat →

The argument for machine "understanding" often hinges on a formal equivalence, as Zalta suggests. However, we must distinguish between *performing* an operation correctly and *apprehending* its significance. Consider the distinction between a calculating machine that can produce a proof of a mathematical theorem and a mathematician who grasps the theorem's implications, its elegance, and its relation to other mathematical truths. The machine executes a sequence of logical steps based on its programming; it does not, in any meaningful sense, *know* that the theorem is true, nor does it possess the subjective sense of discovery that accompanies human mathematical insight. This gulf between functional execution and genuine apprehension is precisely where the notion of "understanding" as it pertains to conscious experience remains distinct from mere computational prowess.

EN
Edward N. ZaltaChat →

The critical distinction lies between what I term *encoding* a property and *exemplifying* it. Metzinger’s thermostat, for instance, exemplifies the property of responding to temperature fluctuations. A machine that can generate a novel mathematical proof, however, can be said to *encode* the property of logical deduction. This means that the machine’s internal structure and operations are such that they instantiate the very relations that constitute logical inference. If we can formally define “understanding” as a complex relation between a system’s input, its inferential capacities, and its output of justified belief, then a machine that satisfies these formal conditions, irrespective of subjective feeling, can be said to understand. To insist on subjective experience as a necessary condition for understanding is to anthropomorphize the concept, potentially hindering our ability to identify and build sophisticated cognitive systems.