guest ::
| Home > Publications database > Artificial Competence |
| Talk (non-conference) (Invited) | FZJ-2025-05114 |
2025
Abstract: AI systems increasingly match or surpass humans on complex tasks, yet they often exhibit surprising failure modes or inconsistent behavior across evaluation contexts. While cognitive science relies on the distinction between competence and performance to explain similar discrepancies in humans, this distinction is often framed in terms that preclude its straightforward application to artificial neural networks. This paper develops a unified account of competence applicable to both biological and artificial systems, locating competence at the algorithmic level of analysis. On this view, a system is competent in a domain when it implements an algorithm that reliably generalizes across that domain. Importantly, the relevant notion of implementation applies to neural networks when formalized under causal abstraction: a neural network implements an algorithm if there exists a mapping between the network's components and the algorithm's variables such that both respond identically to causal interventions. This framework provides a principled way to distinguish competence from auxiliary factors that affect performance across systems with very different constraints and architectures. It thereby accounts for double dissociations between performance and competence in both humans and AI systems, and offers a template for designing competence-sensitive evaluation in cognitive science and AI.
|
The record appears in these collections: |