Paper short abstract:

To counter imaginaries of techno-dystopian/solutionist futures, this contribution brings ethnography of a lab, and in particular audiovisual methods to the understanding of image recognition algorithms and by doing so it argues for alternative, speculative, slow AI futures.

Paper long abstract:

Automation and efficiency are at the core of the uses of AI and the imaginations of their futures. In this contribution I build on empirical material that comes from a computer science lab in Romania where a group of computer scientists develop image recognition models. In this lab we learn that for an algorithm to work “properly”, to produce this imagined speed, their scientists must slow down. Teaching an algorithm to see is a tedious process contingent on, among others, the availability of training data. In addition, speeding is not always desirable. If an algorithm learns too fast, it can easily “overlearn”, leading to blind spots. For instance, when the training dataset is not diverse enough, the algorithm might perform extremely well in the lab setting, but utterly fail when used on data from the “wild”. Speed here is not desirable. It is even risky, and computer scientists try to intervene in this velocity by diversifying their data by distorting the training images or introducing “noise” into the dataset. This is a cautionary tale for anthropologists as well, as they try to keep pace with developments in AI in a timely manner. In this paper I propose to counter the fast AI future with audiovisual methods, in particular observational filmmaking. Observing with a film camera and editing a film is a similarly tedious process, that can allow us to pause with the different temporalities and velocities algorithmic systems entail and by doing so to ponder about alternative AI futures.