If a machine performs an intelligent task, how much does its construction matter? Computational functionalism identifies mental capacities by the roles they play rather than by whether they are realized in a brain or a computer. A role can be described through the inputs a system takes and the outputs or effects it produces. A clock, for example, can keep time whether its mechanism is digital or mechanical. section 9.3.1
This view makes it possible to study machine intelligence without first deciding whether machines have consciousness or need embodiment. Legg’s goal-oriented definition follows that path: it treats intelligence as an agent’s capacity to achieve goals across environments, independently of its substrate. section 9.3.1
When similar results hide different mechanisms
A debate over recurrent neural networks (RNNs) and simpler statistical language models illustrates a hazard. Judging an RNN solely by whether another model can produce similar results can obscure what its architecture contributes. The reverse leap is hazardous too: similar machine and human performance does not establish that their internal mechanisms—or their minds—are equivalent. section 4.1
The distinction matters when turning an engineering result into a claim about mind. McCarthy’s 1956 Dartmouth proposal wagered that features of intelligence could be described precisely enough for a machine to simulate them; it did not claim that simulation makes machine and mind identical. Critics including Hubert Dreyfus challenged approaches that neglected embodiment and conscious experience. A pragmatic response is to build and measure useful functions while leaving those questions open. Functional success can establish what a system does, but cannot by itself explain how it does it or settle what intelligence is. section 9.3.7 section 9.3.1
