Researching AI Legibility through Design

Abstract

Everyday interactions with computers are increasingly likely to involve elements of Artificial Intelligence (AI). Encompassing a broad spectrum of technologies and applications, AI poses many challenges for HCI and design. One such challenge is the need to make AI's role in a given system legible to the user in a meaningful way. In this paper we employ a Research through Design (RtD) approach to explore how this might be achieved. Building on contemporary concerns and a thorough exploration of related research, our RtD process reflects on designing imagery intended to help increase AI legibility for users. The paper makes three contributions. First, we thoroughly explore prior research in order to critically unpack the AI legibility problem space. Second, we respond with design proposals whose aim is to enhance the legibility, to users, of systems using AI. Third, we explore the role of design-led enquiry as a tool for critically exploring the intersection between HCI and AI research.

Design (Non) Fiction: Deconstructing/Reconstructing the Definitional Dualism of AI

“…In a bid to avoid a solution that communicated AI functions or operations in an unapproachable written form, we surveyed how AI is currently communicated via imagery. We found that AI imagery, or iconography, currently lacked the semantics to relay the operational remit of the working parameters of AI, with the majority of images flouting AI's Definitional Dualism (Lindley et al, 2020). This investigation emphasised the need to develop a visual language for AI legibility.…”

“…Despite this, any discussions of AI with non-AIexperts often ends up discussing the thinking machines of AGI. The dichotomy between these two views of AI has been defined as the 'Definitional Dualism of AI' (Lindley et al, 2020), highlighting the misconceptions between AI as materialised in film, media and advertisement campaigns and the actual AI we might experience in our everyday lives. This paper aims to establish a clearer ontology of AI to develop alternate approaches to design with and for AI technologies.…”

Human Data Interaction in Data-Driven Media Experiences : An Exploration of Data Sensitive Responses to the Socio-Technical Challenges of Personal Data Leverage

“…However, while such documents remain central in terms of need for increased legibility, there is the need to extend the view of research to look at legibility in services beyond EULA documents to an on-going relationship that is in sync with the service itself. While recent research has begun such exploration within the scope of AI [40] and even experimental media [55], such 'fledgeling' attempts only further reinforce the need and call for further intervention here.…”

The Process of Gaining an AI Legibility Mark

“…The IOAIL has established a visual language of labels and icons to make users' interactions with AI legible developed from historical research into communicating the inner workings of AI technology [7]. As well as icons communicating the functionalities of an AI system for a user, the IOAIL has created an overall class system as a quick indication for users of how much of the AI system is known (Figure 3), for instance if a product has a mark of IOAIL 1 than it would be a sign that this product and the owner has not disclosed many attributes of the AI present in the technology being scrutinised, an IOAIL 3 mark is a sign that the product and owners have disclosed all attributes and the AI is effectively legible.…”