Do opaque algorithms have functions?

The functions of technical artifacts are closely associated with design. Increasingly, however, we depend on technologies that are not designed: algorithms produced using machine learning (ML). Machine learning uses automated optimization processes to produce algorithms that are often opaque even to developers. I argue that these opaque ML models cannot be ascribed functions on the leading design-based account, the ICE theory of Houkes and Vermaas (Technical functions: On the use and design of artefacts, Springer, 2010). Specifically, I argue that the form of rational justification that the ICE theory provides for the uses of artifacts in typical cases cannot be provided for opaque algorithms, since developers lack the necessary “support beliefs” about how the algorithm works. This does not mean that opaque ML models have no proper functions at all; rather, it lends support to non-intentional theories of technical functions, modeled after function theories in the sciences. However, these theories are meant to perform different work than the ICE theory. As a result, the kind of justification they provide for artifact use is weaker than that provided by the ICE theory and fails to justify many ongoing uses of algorithms.

The paper therefore has two upshots. First, opaque ML models lack intentional technical functions. Second, while they may have functions on alternative theories, in many cases these functional ascriptions still leave the use of algorithms unjustified. Rather than extensions of our agency, opaque algorithms are a kind of found object with unknown properties that may or may not be suited to the purposes to which we put them.