Machine Learning Uncertainty as a Design Material
Part of Post-Phenomenological AI Studies
In this project, I conducted post-phenomenological analyses of four projects from different methodological strands of design research in order to discern how ML technologies may be seen as a design material. Due to the probabilistic techniques employed, I specifically centered on model and data uncertainty. From the analysis of human-technology-world relations unfolding from the diverse artefacts, scenarios and fictions in existing design research projects, my co-authors and I derived three provocative shorthands as an initial vocabulary for ML uncertainty as a design material: thingly uncertainty, pattern leakage and futures creep. Further details below the diagrams showing analyses conducted on design research projects that involved ML technologies.
Benjamin, Jesse Josua,Arne Berger, Nick Merrill, and James Pierce. 2021a. ‘Machine Learning Uncertainty as a Design Material: A Post-Phenomenological Inquiry’. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 1–14. CHI ’21. New York, NY, USA: Association for Computing Machinery. https://doi.org/10.1145/3411764.3445481.
ML uncertainty may not only lead to data leaking from the outside-in to smart home security systems, but patterns leaking from the inside-out.
Data uncertainty can lead to pattern leakage of false positives into everyday experience, affecting human-technology relations beyond the original domain of data sourcing.
Morse Things (IoT devices) embedded with reinforcement learning interact based on the amount of communication during prior cycles; with adjustments after every activity. These probabilistic ‘futures’ creep into human experience of time.
The Highwater Pants (smart wearable) correlate the GPS location of the wearer with a dataset for sealevel-rise prediction, and adapts its activity to the relevant prediction. Here, futures creep shapes the Highwater Pants’ mediation of how bikers relate to possible future states of their environment.
The hypothesis for this project was that, insofar as that design research can be seen as philosophy by other means, design research projects may hold latent concepts, arguments, and hypotheses concerning the attributes of ML technologies as a design material. Deriving concepts through analysis, once made ready in the form of provocative shorthands, can then inform design research practitioners to approach ML technologies from a new, generative vantage point.
Taking the hypothesis that uncertainty could, due to the use of probabilistic techniques, be a suitable way to grasp ML technologies as a design material, I investigated four distinct design research projects (for details see Benjamin et al., 2021). From the analysis of human-technology-world relations unfolding from the diverse artefacts, scenarios and fictions in these projects, my co-authors and I then derived three provocative shorthands as an initial vocabulary for ML uncertainty as a design material.
First, thingly uncertainty functions as a general concept for what is qualitatively distinct about ML-driven design artefacts: Rather than embodying fixed, “scripted” readings of the world, ML-driven artefacts can be much looser and uncertain about the world. As such they act and adapt within a continuum of relations to their environment and the humans that experience them. Second, and as a subconcept, pattern leakage describes the propensity of probabilistic patterns (e.g., of features inferred in image recognition) to shape the world they are deployed to represent. My co-authors and I encourage design researchers to investigate pattern leakage by, for example, exaggerating uncertainty thresholds to exploit the loose access to the world of thingly uncertainty, and thereby provoking patterns to leak. A further use can be, in for example design fiction, to probe for unintended consequences due to intrinsic pattern leakage of ML technologies. Third, and as a further subconcept, we present futures creep to denote how probabilistic, uncertain predictions shape human experience of temporality. An example to make use of this concept is to design intermediary artefacts that ‘keep time’ on ML technologies, or to exaggerate the perceived agency of artefacts by, for example, expanding the constraints on response time.
The Thingly Uncertainty project, then, made (1) a methodological contribution by using post-phenomenological analysis on par with design research practice, and (2) actionable concepts for future work in design research in the form of the provided provocative shorthands.