In the mid-2010s, media artists began developing practices using machine learning (ML) as an artistic medium. Since 2022, the rise of large generative models, the mainstreaming of AI as consumer products, and intensifying ethical disputes have reconfigured the conditions of their artistic practice. This paper aims to understand how artists working with ML over the past decade respond to these shifts, shedding light on how practices, tools, and culture co-evolve. We address this question through thematic analysis of semi-structured interviews with 30 artists active before 2020. Our findings show how artists experience narrowing aesthetics and reduced malleability of post-2020 ML systems, have diverging views on where to locate moral responsibility with large AI models, and face shifting cultural reception that challenges the legibility of their work. We map how artists envision their practice going forward and discuss those orientations with respect to HCI conversations on design and creativity.

The notion of mental model has long been used in HCI to capture people’s understanding and reasoning about computing systems. Eliciting users’ mental models can explain their behaviors and attitudes toward a system—why and how they use, rely on, trust, or reject it. However, its use remains conceptually fragmented and methodologically diverse and has not been revisited in light of modern AI systems, whose opacity and newfound abilities may challenge human understanding. To address this gap, we systematically review 88 empirical studies that elicit humans’ mental models of AI systems. We extracted and analyzed how studies define and elicit mental models, the type of mental model their method presupposes, and how these vary across AI system types. Drawing from the mental model’s framing in cognitive psychology and HCI, and based on descriptive and relational analysis between the variables extracted, we find that (1) mental model elicitations’ goal bifurcates between system-specific evaluation and class-level probes surfacing lay theories; (2) epistemic assumptions exceed the classic functional-structural lens (how the system behaves / how it works internally) with analogical and anthropomorphic framings of AI systems; (3) elicitation methods are shaped more by system characteristics and community-specific practices than theoretical commitments, particularly for predictive and explainable AI systems and autonomous or driver-assist vehicles. We derive 9 practical guidelines to support more deliberate and reflective methods for eliciting mental models of AI systems. In doing so, we aim to reestablish continuity between the cognitive theory of mental models and their empirical use in HCI, improving the transparency and comparability of research surrounding the concept.

Pattern-recognition-based arm prostheses rely on recognizing muscle activation to trigger movements. The effectiveness of this approach depends not only on the performance of the machine learner but also on the user’s understanding of its recognition capabilities, allowing them to adapt and work around recognition failures. We investigate how different model training strategies to select gesture classes and record respective muscle contractions impact model accuracy and user comprehension. We report on a lab experiment where participants performed hand gestures to train a classifier under three conditions: (1) the system cues gesture classes randomly (control), (2) the user selects gesture classes (teacher-led), (3) the system queries gesture classes based on their separability (learner-led). After training, we compare the models’ accuracy and test participants’ predictive understanding of the prosthesis’ behavior. We found that teacher-led and learner-led strategies yield faster and greater performance increases, respectively. Combining two evaluation methods, we found that participants developed a more accurate mental model when the system queried the least separable gesture class (learner-led). Our results conclude that, in the context of machine learning-based myoelectric prosthesis control, guiding the user to focus on class separability during training can improve recognition performances and support users’ mental models about the system’s behavior. We discuss our results in light of several research fields : myoelectric prosthesis control, motor learning, human-robot interaction, and interactive machine teaching.

Image generation gained popularity with machine learning (ML) models generating images from text, fuelling new online communities of practices. This work explores the sociology, motivations, and usages of AI art hobbyists. We analyzed an online questionnaire answered by 64 practitioners and a dataset of user prompts sent to the Stable Diffusion generative model. Our findings suggest that TTI generation is a recreational activity mainly conducted by narrow socio-demographic groups who use auxiliary techniques across platforms and beyond request-response interactions. Inherent model limitations and finding suitable prompt formulation are the main obstacles practitioners face. A taxonomy and a corresponding ML model capable of recognizing the semantic content of unseen prompts were created to conduct the user prompt analysis. The prompt analysis revealed that artist names are the main specifier used beside the main subject, often in sequences. We finally discuss the design and socio-technical implications of our work for creativity support.

Machine Learning models can output confident but incorrect predictions. To address this problem, ML researchers use various techniques to reliably estimate ML uncertainty, usually performed on controlled benchmarks once the model has been trained. We explore how the two types of uncertainty—aleatoric and epistemic—can help non-expert users understand the strengths and weaknesses of a classifier in an interactive setting. We are interested in users’ perception of the difference between aleatoric and epistemic uncertainty and their use to teach and understand the classifier. We conducted an experiment where non-experts train a classifier to recognize card images, and are tested on their ability to predict classifier outcomes. Participants who used either larger or more varied training sets significantly improved their understanding of uncertainty, both epistemic or aleatoric. However, participants who relied on the uncertainty measure to guide their choice of training data did not significantly improve classifier training, nor were they better able to guess the classifier outcome. We identified three specific situations where participants successfully identified the difference between aleatoric and epistemic uncertainty: placing a card in the exact same position as a training card; placing different cards next to each other; and placing a non-card, such as their hand, next to or on top of a card. We discuss our methodology for estimating uncertainty for Interactive Machine Learning systems and question the need for two-level uncertainty in Machine Teaching.

Human-centered approaches to machine learning have established theoretical foundations, design principles and interaction techniques to facilitate end-user interaction with machine learning systems. Yet, general-purpose toolkits supporting the design of interactive machine learning systems are still missing, despite their potential to foster reuse, appropriation and collaboration between different stakeholders including developers, machine learning experts, designers and end users. In this paper, we present an architectural model for toolkits dedicated to the design of human interactions with machine learning. The architecture is built upon a modular collection of interactive components that can be composed to build interactive machine learning workflows, using reactive pipelines and composable user interfaces. We introduce Marcelle, a toolkit for the design of human interactions with machine learning that implements this model. We illustrate Marcelle with two implemented case studies: (1) a HCI researcher conducts user studies to understand novice interaction with machine learning, and (2) a machine learning expert and a clinician collaborate to develop a skin cancer diagnosis system. Finally, we discuss our experience with the toolkit, along with its limitation and perspectives.

Machine learning systems became pervasive in modern interactive technology but provide users with little, if any, agency with respect to how their models are trained from data. In this paper, we are interested in the way novices handle learning algorithms, what they understand from their behavior and what strategy they may use to "make it work". We developed a web-based sketch recognition algorithm based on Deep Neural Network (DNN), called Marcelle-Sketch, that end-users can train incrementally. We present an experimental study that investigate people's strategies and (mis)understandings in a realistic algorithm-teaching task. Our study involved 12 participants who performed individual teaching sessions using a think-aloud protocol. Our results show that participants adopted heterogeneous strategies in which variability affected the model performances. We highlighted the importance of sketch sequencing, particularly at the early stage of the teaching task. We also found that users' understanding is facilitated by simple operations on drawings, while confusions are caused by certain inherent properties of DNN. From these findings, we propose implications for design of IML systems dedicated to novices and discuss the socio-cultural aspect of this research.