with James Turner
12th European Congress of Analytic Philosophy, Madrid, 24–28 August 2026

Modern AI systems built on artificial neural networks (ANNs) display behaviours that, in animals, would typically be explained by appeal to internal representations. This has prompted debate over whether such systems could possess internal representation. We advance this discussion by applying a teleosemantic framework, according to which states’ contents depend on the proper functions of the mechanisms that produce/consume these states. The key question is: could ANN‑based systems have components with proper functions? Focusing on a long‑standing teleosemantic tradition, we examine whether current ANN architectures could underpin an engineered “AI frog” with functions relevantly similar to those of organic frogs—specifically, detecting and snapping at flies. We argue that such engineering is possible. Although AI frogs are not shaped by natural selection, their training involves a form of differential retention that can ground proper functions in their artificial neurons, akin to how differential retention grounds function in organic neurons.