Technology for Automating Data Science (AutoDS) consistently undervalues the role of human labor, resulting in tools that, at best, are ignored and, at worst, can actively mislead or even cause harm. Even if full and frictionless automation were possible, human oversight is still desired and required to review the outputs of AutoDS tooling and integrate them into decision-making processes. We propose a human-centered lens to AutoDS that emphasizes the collaborative relationships between humans and these automated processes and elevates the effects these interactions have on downstream decision-making. Our approach leverages a provenance framework that integrates user-, data-, and model-centric approaches to make AutoDS platforms observable and interrogable by humans.

Contrastive neural network models (i.e. CLIP) based image generation services (e.g. DALL-E2, MidJourney, Stable Diffusion) have shown that they can produce a huge range of flawless images, consistent with a user-provided image concept in text. While a lot of people have shared successful cases on the Internet, we still have very limited knowledge about whether such tools are helpful for daily design work. We conducted a preliminary observational study to investigate how designers create moodboards using DALL-E2. The results indicate that novice users would find it hard to find best prompts for creating and modifying generate images. The goal of this position paper is to propose potential research areas and ideas such as how to set guidelines for designing interactive image generation services for a specific purpose.

In this paper, we explore how different video augmentation techniques transition the representation learning of a dyad’s joint engagement. We evaluate state-of-the-art action recognition models (TimeSformer, X3D, I3D, and SlowFast) on parent-child interaction video dataset with joint engagement recognition task and demonstrate how the performance varies by applying different video augmentation techniques (General Aug, DeepFake, and CutOut). We also introduce a novel metric to objectively measure the quality of learned representations (Grad-CAM) and relate this with social cues (smiling, head angle, and body closeness) by conducting correlation analysis. Furthermore, we hope our method serves as a strong baseline for future human interaction analysis research.

How much are we to trust a decision made by an AI algorithm? Trusting an algorithm without cause may lead to abuse, and mistrusting it may similarly lead to disuse. Trust in an AI is only desirable if it is warranted; thus, calibrating trust is critical to ensuring appropriate use. In the name of calibrating trust appropriately, AI developers should provide contracts specifying use cases in which an algorithm can and cannot be trusted. Automated explanation of AI outputs is often touted as a method by which trust can be built in the algorithm. However, automated explanations arise from algorithms themselves, so trust in these explanations is similarly only desirable if it is warranted. Developers of algorithms explaining AI outputs (xAI algorithms) should provide similar contracts, which should specify use cases in which an explanation can and cannot be trusted.

Recommendation systems can be seen as one of the first successful paradigms of true human-AI collaboration. That is, the AI identifies what the user might want and provide this to them at the right time; and the user, implicitly or explicitly, gives feedback of whether they value said recommendations. However, to make the recommender a \emph{true companion} of users, amplifying and augmenting the capabilities of users to be more knowledgeable, healthy, and happy, requires a shift into the way this collaboration happens. In this position paper, we argue for an increasing focus into reflecting the user values into the design, evaluation, training objectives, and interaction paradigm of state-of-the-art recommendation.

Large-scale generative models are increasingly being used in tooling applications. As one prominent example, code generation models recommend code completions within an IDE to help programmers author software. However, since these models are imperfect, their erroneous recommendations can introduce bugs or even security vulnerabilities into a code base if not overridden by a human user. In order to override such errors, users must first detect them. One method of assisting this detection has been highlighting tokens with low generation probabilities. We also propose another method, predicting the tokens people are likely to edit in a generation. Through a mixed-methods, pre-registered study with N = 30 participants, we find that the edit model highlighting strategy results in significantly faster task completion time, significantly more localized edits, and was strongly preferred by participants.

We often construct AI systems that optimize over specified objectives serving as proxies for human values. Consider recommender systems on social media and entertainment streaming platforms, which maximize the time-on-application or other user-engagement metrics, as a proxy for providing entertaining content to users. The research community has also begun to study how optimizing systems influence human values (e.g., shifts political leanings or predictable induction into specific online communities). We are left with an obvious, yet overlooked framework: Consideration of values and optimizers as a highly intertwined and interactive system; one that constantly feeds into and transforms the other. This perspective is crucial for engineering safe and beneficial AI systems—ones which preserve diverse values across individuals and communities.

The current machine learning applications in the hiring process are prone to bias, especially due to poor quality and small quantity of data. The bias in hiring imposes potential societal and legal risks. Thus, it is important to evaluate ML applications' bias in the hiring context. To investigate the algorithmic bias, we use real-world employment data to train models for predicting job candidates' performance and retention. The result shows that ML algorithms make biased decisions toward a certain group of job candidates. This analysis motivates us to resort to an alternative method---AI-assisted hiring decision making. We plan to conduct an experiment with human subjects to evaluate the effectiveness of human-AI collaboration for algorithmic bias mitigation. In our designed study, we will systematically explore the role of human-AI teaming in enhancing the fairness of hiring in practice.

Value alignment has been widely argued to be one of the central safety problems in AI. While the problem itself arises from the way humans interact with the AI systems, most current solutions to value alignment tend to sideline the human or make unrealistic assumptions about possible human interactions. In this position paper, we propose a human-centered formalization of the value alignment problem that generalizes human-AI interaction frameworks that were originally developed for explainable AI. We see how such a human-aware formulation of the problem provides us with novel ways of addressing and understanding the problem.

Natural language generation has been one of the prime focuses of human-AI collaboration in recent years. We are specifically interested in exploring the idea of creativity in human-AI co-creation, most especially in the context of persona generation for the iterativehuman-centered design process. Collaborating with AIs to generate engaging personas may present opportunities to overcome the shortcomings of personas and how they’re currently used in the design process. We aim to study how collaborating with AIs might help designers and researchers to create engaging personas and narrative scenarios for their products, and by extension the implications of human-AI collaborative creative writing on fields like literature with character generation. The implications of such a study could be generalized beyond user-experience design and persona generation. The ability to create engaging personas is not dissimilar from the ability to generate characters as a whole, and the subsequent potential for natural language generation to assist in creative writing and thinking is implicit. In this paper, we will discuss the process and potential merits of iterating with AIs for creative content creation, as well as expand upon experiments we have conducted and the questions we hope to answer in our future research.

Algorithms used for decision-making in higher education promise cost-savings to institutions and personalized service for students, but at the same time, raise ethical challenges around surveillance, fairness, and interpretation of data. To address the lack of systematic understanding of how these algorithms are currently designed, we reviewed algorithms proposed by the research community for higher education. We explored the current trends in the use of computational methods, data types, and target outcomes, and analyzed the role of human-centered algorithm design approaches in their development. Our preliminary research suggests that the models are trending towards deep learning, increased use of student personal data and protected attributes, with the target scope expanding towards automated decisions. Despite the associated decrease in interpretability and explainability, current development predominantly fails to incorporate human-centered lenses.

There is a growing frustration amongst researchers and developers in Explainable AI (XAI) around the lack of consensus around what is meant by ‘explainability’. Do we need one definition of explainability to rule them all? In this paper, we argue why a singular definition of XAI is neither feasible nor desirable at this stage of XAI’s development. We view XAI through the lenses of Social Construction of Technology (SCOT) to explicate how diverse stakeholders (relevant social groups) have different interpretations (interpretative flexibility) that shape the meaning of XAI. Forcing a standardization (closure) on the pluralistic interpretations too early can stifle innovation and lead to premature conclusions. We share how we can leverage the pluralism to make progress in XAI without having to wait for a definitional consensus.

Artificial intelligence has a communications challenge. To create human-centric AI, it is important that XAI is able to adapt to different users. The SciCom field provides a mixed-methods approach that can provide better understanding of users’ framings so as to improve public engagement and expectations of AI systems, as well as help AI systems better adapt to their particular user.

/

Two years after publicly launching the AI Incident Database (AIID) as a collection of harms or near harms produced by AI in the world, a backlog of issues'' that do not meet its incident ingestion criteria have accumulated in its review queue. Despite not passing the database's current criteria for incidents, these issues advance human understanding of where AI presents the potential for harm. Similar to databases in aviation and computer security, the AIID proposes to adopt a two-tiered system for indexing AI incidents (i.e., a harm or near harm event) and issues (i.e., a risk of a harm event). Further, as some machine learning-based systems will sometimes produce a large number of incidents, the notion of an incidentvariant'' is introduced. These proposed changes mark the transition of the AIID to a new version in response to lessons learned from editing 1,800+ incident reports and additional reports that fall under the new category of ``issue.''

Human-Center eXplainable AI (HCXAI) literature identifies the need to address user needs. This paper examines how existing XAI research involves human users in designing and developing XAI systems and identifies limitations in current practices, especially regarding how researchers identify user requirements. Finally, we propose several suggestions on how to derive better user requirements by deeper engagement with user groups.

We argue that synthesizing insights from humans and ML models at the level of features is an important direction to explore to improve human-ML collaboration on decision-making problems. We show through an illustrative example that feature-level synthesis can produce correct predictions in a case where existing methods fail, then lay out directions for future exploration.

Oversight mechanisms, whereby the functioning and behaviour of AI systems are controlled to ensure that they are tuned to public benefit, are a core aspect of human-centered AI. They are especially important in public sector AI applications, where decisions on core public services such as education, benefits, and child welfare have significant impacts. Much current thinking on oversight mechanisms revolves around the idea of human decision makers being present ‘in the loop’ of decision making, such that they can insert expert judgment at critical moments and thus rein in the functioning of the machine. While welcome, we believe that the theory of human in the loop oversight has yet to fully engage with the idea that decision making, especially in high-stakes contexts, is often currently made by hierarchical teams rather than one individual. This raises the question of how such hierarchical structures can effectively engage with an AI system that is either supporting or making decisions. In this position paper, we outline some of the key contemporary elements of hierarchical decision making in contemporary public services and show how they relate to current thinking about AI oversight, thus sketching out future research directions for the field.

Algorithmic risk assessment tools are now commonplace in public sector domains such as criminal justice and human services. In this paper we argue that understanding how the deployment of such tools affect decision-making requires a considering of organizational factors and worker characteristics that may influence the take-up of algorithmic recommendations. We discuss some existing evaluations of real-world algorithms and show that labor force characteristics play a significant role in influencing these human-in-the-loop systems. We then discuss our findings from a real-world child abuse hotline screening use case, in which we investigate the role that worker experience plays in algorithm-assisted decision-making. We argue that system designers should consider ways of preserving institutional knowledge when introducing algorithms into settings with high employee turnover.

Algorithm aversion occurs when humans are reluctant to use algorithms despite their superior performance. Prior studies have shown that giving users outcome control'', the ability to appeal or modify model's predictions, can mitigate this aversion. This can be contrasted withprocess control'', which entails control over the development of the algorithmic tool. The effectiveness of process control is currently under-explored. To compare how various controls over algorithmic systems affect users' willingness to use the systems, we replicate a prior study on outcome control and conduct a novel experiment investigating process control. We find that involving users in the process does not always result in a higher reliance on the model. We find that process control in the form of choosing the training algorithm mitigates algorithm aversion, but changing inputs does not. Giving users both outcome and process control does not result in further mitigation than either outcome or process control alone. Having conducted the studies on both Amazon Mechanical Turk (MTurk) and Prolific, we also reflect on the challenges of replication for crowdsourcing studies of human-AI interaction.

We use the Rational Speech Act framework to examine AI explanations as a pragmatic inference process. This reveals fatal flaws in how we currently train and deploy AI explainers. To evolve from level-0 explanations to level-1, we present two proposals for data collection and training: learning from L1 feedback, and learning from S1 supervision.

Improving the performance of human-AI (artificial intelligence) collaborations tends to be narrowly scoped, with better prediction performance often considered the only metric of improvement. As a result, work on improving the collaboration usually focuses on improving the AI's accuracy. Here, we argue that such a focus is myopic, and instead, practitioners should take a more holistic view of measuring the performance of AI models, and human-AI collaboration more specifically. In particular, we argue that although some use cases merit optimizing for classification accuracy, for others, accuracy is less important and improvement on human-centered metrics should be valued instead.

The outputs of most Machine Learning (ML) systems are often riddled with uncertainties, biased from the training data, sometimes incorrect, and almost always inexplicable. However, in most cases, their user interfaces are oblivious to those shortcomings, creating many undesirable consequences, both practical and ethical. I propose that ML user interfaces should be designed to make clear those issues to the user by exposing uncertainty and bias, instilling distrust, and avoiding imposture. This is captured by the overall concept of Honesty, which I argue should be the most important guide for the design of ML interfaces.

In this short article, we focus on four dimensions of data that create layers of complexity using the data in the context of AI systems developed for medical applications collectively referred to as diagnostic medicine. These complexities, or ‘data dilemmas,’ share a core human element, making it clear why a human-centered approach is needed in understanding the relationship between medical data and AI systems.

Research on human-centered AI teammates has often worked to create AI teammates that adapt around humans, but humans have a remarkable and natural ability to adapt around their environment and teammates. This paper capitalizes on human adaptability by showcasing how humans actively adapt around their AI teammates even when those teammates change. In doing so, results of a mixed-methods experiment (N = 60) demonstrates that human adaptation is a critical and natural component of human-centered AI teammate design.

Card et al.’s classic paper "The Design Space of Input Devices" established the value of design spaces as a tool for HCI analysis and invention. We posit that developing design spaces for emerging pre-trained, general AI models is necessary for supporting their integration into human-centered systems and practices. We explore what it means to develop an AI model design space by proposing two design spaces relating to pre-trained AI models: the first considers how HCI can impact pre-trained models (i.e., interfaces for models) and the second considers how pre-trained models can impact HCI (i.e., models as an HCI prototyping material).

With the successful adoption of machine learning (ML) in decision making, there have been growing concerns around the transparency and fairness of ML models leading to significant advances in the field of eXplainable Artificial Intelligence (XAI). Generating explanations using existing techniques in XAI and merely reporting model bias, however, are insufficient to locate and mitigate sources of bias. In line with the data-centric AI movement, we posit that to mitigate bias, we must solve the myriad data errors and biases inherent in the data, and propose a human-machine framework that strengthens human engagement with data to remedy data errors and data biases toward building fair and trustworthy AI systems.

Biases in human-centric computer vision models are often attributed to a lack of sufficient data diversity, with many demographics insufficiently represented. However, auditing datasets for diversity can be difficult, due to an absence of ground-truth labels of relevant features. Few datasets contain self-identified demographic information, inferring demographic information risks introducing additional biases, and collecting and storing data on sensitive attributes can carry legal risks. Moreover, categorical demographic labels do not necessarily capture all the relevant dimensions of human diversity that are important for developing fair and robust models. We propose to implicitly learn a set of continuous face-varying dimensions, without ever asking an annotator to explicitly categorize a person. We uncover the dimensions by learning on a novel dataset of 638,180 human judgments of face similarity (FAX). We demonstrate the utility of our learned embedding space for predicting face similarity judgments, collecting continuous face attribute values, and comparative dataset diversity auditing. Moreover, using a novel conditional framework, we show that an annotator's demographics influences the importance they place on different attributes when judging similarity, underscoring the need for diverse annotator groups to avoid biases.

Metric Elicitation (ME) is a framework for eliciting classification metrics that better align with implicit user preferences based on the task and context. The existing ME strategy so far is based on the assumption that users can most easily provide preference feedback over classifier statistics such as confusion matrices. This work examines ME, by providing a first ever implementation of the ME strategy. Specifically, we create a web-based ME interface and conduct a user study that elicits users' preferred metrics in a binary classification setting. We discuss the study findings and present guidelines for future research in this direction.

Multiplayer gaming yields social benefits, but can cause harm through toxicity—particularly as directed toward women, players of color, and 2SLGBTQ+ players. Detecting toxicity is challenging, but is necessary for intervention. We present three challenges to automated toxicity detection, and share potential solutions so that researchers can develop HCAI models that detect toxic game communication.

In this position paper, we reflect on fictional stories dealing with the infinite and how they connect with the current, fast-evolving field of image generation models. We draw attention to how some of these literary constructs can serve as powerful metaphors for guiding human-centered design and technical thinking in the space of these emerging technologies and the experiences we build around them. We hope our provocations seed conversations about current and yet to be developed interactions with these emerging models in ways that may amplify human agency.

Motivated by mitigating potentially harmful impacts of technologies, the AI community has formulated and accepted mathematical definitions for certain pillars of accountability: e.g. privacy, fairness, and model transparency. Yet, we argue this is fundamentally misguided because these definitions are imperfect, siloed constructions of the human values they hope to proxy, while giving the guise that those values are sufficiently embedded in our technologies. Under popularized techniques, tensions arise when practitioners attempt to achieve each pillar of fairness, privacy, and transparency in isolation or simultaneously. In this position paper, we argue that the AI community needs to consider alternative formulations of these pillars based on the context in which technology is situated. By leaning on sociotechnical systems research, we can formulate more compatible, domain-specific definitions of our human values for building more ethical systems.