Human-Centered AI at IUI 2023
by Zahra Ashktorab (IBM Research, US) and Ziang Xiao (Johns Hopkins University, US)
The ACM Intelligent User Interfaces conference provides a platform for researchers to showcase their work on the intersection of Human-Computer Interaction (HCI) and Artificial Intelligence (AI). At IUI 2023, researchers from various academic and industry labs presented papers that focused on the potential of Human-Centered AI to improve various aspects of work and decision-making. In this article, we identify three emergent themes from several of the notable HCAI papers at IUI.
Human-AI Collaboration
A fundamental idea in human-centered AI is that people ought to work together with — to collaborate with — AI systems. Several papers explored how human-AI collaboration can enhance work and decision-making.
Research by Hemmer et al. [1] examined the effectiveness of delegating image classification tasks to AI models. They found that delegation improved peoples’ performance and satisfaction and they concluded that giving people control over which tasks they perform versus which tasks the AI performs can be a valuable way for people to collaborate with AI systems in a human-centered fashion. Work by Ou et al. [4] examined the impact of user expertise on decision-making tasks and found that domain experts outperformed non-experts when making AI-assisted decisions.
Conversational interactions with large language models (LLMs) is a hot topic. Ross et al. [3] examined the use of a conversational assistant for software engineering. They found that conversation provides an effective user experience that uncovers the emergent capabilities of an LLM in ways not exposed via direct manipulation interfaces. Xiao et al. [9] also examined the use of conversational agents in a high-stakes information-seeking task: answering questions about COVID-19. They found that, even though the natural language understanding capability of an AI agent still has room for improvement, participants preferred the AI chatbot’s delivery of information in natural language over web search results.
In-context learning allows people to use an LLM for a wide range of tasks by curating a demonstration set of examples in the prompt. However, the performance of the model is highly dependent on the selected examples, and a better set of curated examples may drive significant performance gains. Wu et al. [12] presented ScatterShot, an interactive system for building high-quality example sets for in-context learning. The system first slices data into categories based on task-specific patterns. Then, it samples informative inputs from underexplored or not-yet-saturated slices in an active learning manner to help people curate a more effective set of examples for in-context learning. Through a simulation study, the authors found that ScatterShot helps people effectively and efficiently identify the example set for in-context learning and improve the LLM’s performance. This work received an Honorable Mention award.
Lastly, work by Bhat et al. [11] examined how people use an LLM’s suggestions when writing. They proposed a model of writer-suggestion interaction to understand peoples’ co-creative process when writing with LLMs. The authors conducted a qualitative study to understand writers’ cognitive processes while writing with LLMs. They found that writers utilized an LLM’s suggestions both directly and indirectly by abstracting and extracting parts of the suggestions. However, the need to constantly read and evaluate an LLM’s suggestions led to a higher level of cognitive load that impeded effective writing. This work received an Honorable Mention award.
AI and Decision Making
Several papers investigated the impact of AI on decision-making processes. Kahr et al. [5] examined the effects of a model’s accuracy and the types of explanations it provided on peoples’ trust in the model. They compared explanations that were presented in an abstract manner represented as a simple series of words, with explanations that were presented in a human-like manner, as a full-sentence text with human-like reasoning. In their evaluative study, they found no differences in how each of these types of explanations impacted trust, except in cases in which the model had a higher degree of accuracy. In those cases, they found that human-like explanations elevated trust more than abstract explanations.In human-AI decision making, reliance is the idea that people accept the recommendation of the AI when it is presented to them before they have made their own judgment. Overreliance happens when people accept an AI’s recommendation even when it isn’t correct. Work by Schemmer et al. [2] addresses the lack of consistency in how appropriate reliance (AR) is defined in current research. AR has sometimes been defined as a binary target state (e.g. “appropriate” or “not appropriate”), and other times as a metric indicating the degree of appropriateness. To address this inconsistency, Schemmer et al. introduce a two-dimensional metric called the appropriateness of reliance (AoR) scale. The AoR scale assesses two behaviors: when a human overrides an incorrect AI suggestion (called “self-reliance”) and when a human follows a correct AI suggestion (called “AI reliance”). Their work, which won a Best Paper Award, offers new insights into conceptualizing and measuring the idea of appropriate reliance in human-centered AI systems.
Weidele et al. [7] demonstrated an interactive user interface for automated decision optimization, called AutoDOViz. Decision optimization is the process of using mathematical models and algorithms to solve decision problems, such as how much inventory a store should stock, given a set of constraints and objectives. AutoDOViz uses reinforcement learning (RL) to make decision optimization more accessible to data scientists. In a user study, the researchers found that AutoDOViz increased participants’ trust in the decision optimization models and improved participants’ understanding of the automated training and evaluation process.
Prabhudesai et al. [8] explored how and why decision makers utilize the uncertainty of a model’s predictions in their decision-making process. Their findings suggest that people may interpret visualizations of uncertainty differently in a high-stakes decision making process. Further, peoples’ interpretations vary based on their levels of expertise. This study proposes several novel uncertainty-based designs to aid AI-assisted decision-making.
Explanations and Mental Models
Brachman et al. [6] examined how explanations could support people in authoring effective natural language utterances for a system that generates small programs. They conducted an online study to compare two types of explanations: system-focused, which provide information about how the system processes utterances and matches terms to a knowledge base, and social recommendations, which provides information about how other people have successfully interacted with the system. They found that social suggestions, in the form of text snippets used by other users, helped users generate small programs more than system-focused explanations or social recommendations.
Ooge et al. [10] argued that XAI researchers should focus more on adolescents due to their unique developmental stage and experience with AI systems. They examined how adolescents interact with and react to AI controls and explanation methods in an e-learning platform. They found that adolescents appreciated the ability to control which learning exercises were recommended to them, which prompted them to spend more time reflecting on their mastery level of the material. Self-reflection is a critical practice in self-regulated learning. Furthermore, visualizing the impact of the AI controls increased students’ trust and perceived understanding.
Conclusion
The papers presented at IUI 2023 highlighted the potential of Human-Centered AI to improve various aspects of work and decision-making. They provided insights into the unique challenges and opportunities in human-AI collaboration, AI-assisted decision making, and explainable AI. This work highlights the importance for researchers and practitioners to design more effective and inclusive AI systems.
References
The full list of papers published at IUI 2023 can be found on the conference website: https://iui.acm.org/2023/toc.html.
[1] Patrick Hemmer, Monika Westphal, Max Schemmer, Sebastian Vetter, Michael Vössing, and Gerhard Satzger. 2023. Human-AI Collaboration: The Effect of AI Delegation on Human Task Performance and Task Satisfaction. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 453–463. https://doi.org/10.1145/3581641.3584052
[2] Max Schemmer, Niklas Kuehl, Carina Benz, Andrea Bartos, and Gerhard Satzger. 2023. Appropriate Reliance on AI Advice: Conceptualization and the Effect of Explanations. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 410–422. https://doi.org/10.1145/3581641.3584066
[3] Steven I. Ross, Fernando Martinez, Stephanie Houde, Michael Muller, and Justin D. Weisz. 2023. The Programmer’s Assistant: Conversational Interaction with a Large Language Model for Software Development. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 491–514. https://doi.org/10.1145/3581641.3584037
[4] Changkun Ou, Sven Mayer, and Andreas Martin Butz. 2023. The Impact of Expertise in the Loop for Exploring Machine Rationality. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 307–321. https://doi.org/10.1145/3581641.3584040
[5] Patricia K. Kahr, Gerrit Rooks, Martijn C. Willemsen, and Chris C.P. Snijders. 2023. It Seems Smart, but It Acts Stupid: Development of Trust in AI Advice in a Repeated Legal Decision-Making Task. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 528–539. https://doi.org/10.1145/3581641.3584058
[6] Michelle Brachman, Qian Pan, Hyo Jin Do, Casey Dugan, Arunima Chaudhary, James M. Johnson, Priyanshu Rai, Tathagata Chakraborti, Thomas Gschwind, Jim A Laredo, Christoph Miksovic, Paolo Scotton, Kartik Talamadupula, and Gegi Thomas. 2023. Follow the Successful Herd: Towards Explanations for Improved Use and Mental Models of Natural Language Systems. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 220–239. https://doi.org/10.1145/3581641.3584088
[7] Daniel Karl I. Weidele, Shazia Afzal, Abel N. Valente, Cole Makuch, Owen Cornec, Long Vu, Dharmashankar Subramanian, Werner Geyer, Rahul Nair, Inge Vejsbjerg, Radu Marinescu, Paulito Palmes, Elizabeth M. Daly, Loraine Franke, and Daniel Haehn. 2023. AutoDOViz: Human-Centered Automation for Decision Optimization. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 664–680. https://doi.org/10.1145/3581641.3584094
[8] Snehal Prabhudesai, Leyao Yang, Sumit Asthana, Xun Huan, Q. Vera Liao, and Nikola Banovic. 2023. Understanding Uncertainty: How Lay Decision-makers Perceive and Interpret Uncertainty in Human-AI Decision Making. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 379–396. https://doi.org/10.1145/3581641.3584033
[9] Ziang Xiao, Q. Vera Liao, Michelle Zhou, Tyrone Grandison, and Yunyao Li. 2023. Powering an AI Chatbot with Expert Sourcing to Support Credible Health Information Access. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 2–18. https://doi.org/10.1145/3581641.3584031
[10] Jeroen Ooge, Leen Dereu, and Katrien Verbert. 2023. Steering Recommendations and Visualising Its Impact: Effects on Adolescents’ Trust in E-Learning Platforms. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 156–170. https://doi.org/10.1145/3581641.3584046
[11] Advait Bhat, Saaket Agashe, Parth Oberoi, Niharika Mohile, Ravi Jangir, and Anirudha Joshi. 2023. Interacting with Next-Phrase Suggestions: How Suggestion Systems Aid and Influence the Cognitive Processes of Writing. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 436–452. https://doi.org/10.1145/3581641.3584060
[12] Sherry Wu, Hua Shen, Daniel S Weld, Jeffrey Heer, and Marco Tulio Ribeiro. 2023. ScatterShot: Interactive In-context Example Curation for Text Transformation. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI ‘23). Association for Computing Machinery, New York, NY, USA, 353–367. https://doi.org/10.1145/3581641.3584059
