Governments are the primary providers of essential public services and are responsible for delivering them effectively. In high-stakes decision-making domains such as child welfare (CW), agencies must protect children without unnecessarily prolonging a family’s engagement with the system. With growing optimism around AI, governments are pushing for its integration but concerns regarding feasibility and harms remain. Through collaborations with a large Canadian CW agency, we examined how LocalLLM and BERTopic models can track CW case progress. We demonstrate how the tools can potentially assist workers in opportunistically addressing gaps in their work by signaling case progress/deviations. And yet, we also show how they fail to detect case trajectories that require discretionary judgments grounded in social work training, areas where practitioners would actually want support to pre-emptively address substantive case concerns. We also provide a roadmap of future participatory directions to co-design language tools for/with the public sector.

Public sector agencies increasingly rely on data and information systems to demonstrate that they “support” families. Yet, the metrics that stand in for support are often misaligned with how they are lived. We draw on interviews with 75 parents involved in the child welfare system (CWS)—an entry point into the broader public safety net—and use an interpretive computational workflow that combines thematic coding, a small language model, and multidimensional scaling to examine parents’ accounts of support and unmet needs. We find anad hoc safety net in which formal services, public assistance, and caseworkers’ efforts are braided together with kin, peers, and employers, but in fragmented and conditional ways. Parents with robust informal networks are better positioned to appear engaged and to receive additional help, while those with few informal supports are more likely to be documented as non-compliant and experience further neglect. These patterns reveal information gaps between sociocentric metrics (e.g., referrals, completions) and parents’ egocentric outcomes (timeliness, trust, feasibility). We discuss implications for designing collaborative information systems that use narrative-rich qualitative accounts to uncover latent patterns in how support and neglect are experienced by parents in the public safety net.

AI projects often fail due to financial, technical, ethical, or user acceptance challenges-failures frequently rooted in early-stage decisions.While HCI and Responsible AI (RAI) research emphasize this, practical approaches for identifying promising concepts early remain limited.Drawing on Research through Design, this paper investigates how early-stage AI concept sorting in commercial settings can reflect RAI principles.Through three design experiments-including a probe study with industry practitioners-we explored methods for evaluating risks and benefits using multidisciplinary collaboration.Participants demonstrated strong receptivity to addressing RAI concerns early in the process and effectively identified low-risk, high-benefit AI concepts.Our findings highlight the potential of a design-led approach to embed ethical and service design thinking at the front end of AI innovation.By examining how practitioners reason about AI concepts, our study invites HCI and RAI communities to see early-stage innovation as a critical space for engaging ethical and commercial considerations together.

Homelessness systems in North America adopt coordinated data-driven approaches to efficiently match support services to clients based on their assessed needs and available resources. AI tools are increasingly being implemented to allocate resources, reduce costs and predict risks in this space. In this study, we conducted an ethnographic case study on the City of Toronto's homelessness system's data practices across different critical points. We show how the City's data practices offer standardized processes for client care but frontline workers also engage in heuristic decision-making in their work to navigate uncertainties, client resistance to sharing information, and resource constraints. From these findings, we show the temporality of client data which constrain the validity of predictive AI models. Additionally, we highlight how the City adopts an iterative and holistic client assessment approach which contrasts to commonly used risk assessment tools in homelessness, providing future directions to design holistic decision-making tools for homelessness.

Data scientists often formulate predictive modeling tasks involving fuzzy, hard-to-define concepts, such as the ''authenticity'' of student writing or the ''healthcare need'' of a patient. Yet the process by which data scientists translate fuzzy concepts into a concrete, proxy target variable remains poorly understood. We interview fifteen data scientists in education (N=8) and healthcare (N=7) to understand how they construct target variables for predictive modeling tasks. Our findings suggest that data scientists construct target variables through a bricolage process, in which they use creative and pragmatic approaches to make do with the limited data at hand. Data scientists attempt to satisfy five major criteria for a target variable through bricolage: validity, simplicity, predictability, portability, and resource requirements. To achieve this, data scientists adaptively apply problem (re)formulation strategies, such as swapping out one candidate target variable for another when the first fails to meet certain criteria (e.g., predictability), or composing multiple outcomes into a single target variable to capture a more holistic set of modeling objectives. Based on our findings, we present opportunities for future HCI, CSCW, and ML research to better support the art and science of target variable construction.

Research into recidivism risk prediction in the criminal justice system has garnered significant attention from HCI, critical algorithm studies, and the emerging field of human-AI decision-making. This study focuses on algorithmic crime mapping, a prevalent yet underexplored form of algorithmic decision support (ADS) in this context. We conducted experiments and follow-up interviews with 60 participants, including community members, technical experts, and law enforcement agents (LEAs), to explore how lived experiences, technical knowledge, and domain expertise shape interactions with the ADS, impacting human-AI decision-making. Surprisingly, we found that domain experts (LEAs) often exhibited anchoring bias, readily accepting and engaging with the first crime map presented to them. Conversely, community members and technical experts were more inclined to engage with the tool, adjust controls, and generate different maps. Our findings highlight that all three stakeholders were able to provide critical feedback regarding AI design and use - community members questioned the core motivation of the tool, technical experts drew attention to the elastic nature of data science practice, and LEAs suggested redesign pathways such that the tool could complement their domain expertise.

Caseworkers in the child welfare (CW) sector use predictive decision-making algorithms built on risk assessment (RA) data to guide and support CW decisions. Researchers have highlighted that RAs can contain biased signals which flatten CW case complexities and that the algorithms may benefit from incorporating contextually rich case narratives, i.e. - the casenotes written by caseworkers. To investigate this hypothesized improvement, we quantitatively deconstructed two commonly used RAs from a United States CW agency. We trained classifier models to compare the predictive validity of RAs with and without casenote narratives and applied computational text analysis on casenotes to highlight topics uncovered in the casenotes. Our study finds that common risk metrics used to assess families and build CWS predictive risk models (PRMs) are unable to predict discharge outcomes for children who are not reunified with their birth parent(s). We also find that although casenotes cannot predict discharge outcomes, they contain contextual case signals. Given the lack of predictive validity of RA scores and casenotes, we propose moving beyond quantitative risk assessments for public sector algorithms and towards using contextual sources of information such as narratives to study public sociotechnical systems.

Smartphone enhances healthcare support for everyone, from local to remote patients. Recent advancements in smartphone sensors redefine their usage and the prospect of remote point-of-care tools (e.g., blood diagnostic devices), especially for low-resource settings. This paper studies the sufferings of rural people due to the limited healthcare facilities and figures out the implications. The proliferation of smartphone users suggests converting many smartphones into point-of-care diagnosis devices would be a life-saving decision. Previous studies showed smartphone’s built-in camera captures physiological features (e.g., hemoglobin) from fingertip videos captured under different lights. So, we created a mobile application and attachments (light sources) to record fingertip videos for hemoglobin level calculation. Then we collected feedback on how the rural users interacted with the application. Finally, we applied qualitative and quantitative analysis to investigate their answers. Their invaluable feedback reflected the implications of various aspects of a smartphone-based point-of-care tool. The findings unveil how rural-area people can receive a smartphone's blood diagnostic services. Our results will facilitate mobile health application designers and developers to build a smartphone-based point-of-care tool for any rural area people.

Risk assessment algorithms are being adopted by public sector agencies to make high-stakes decisions about human lives. Algorithms model “risk” based on individual client characteristics to identify clients most in need. However, this understanding of risk is primarily based on easily quantifiable risk factors that present an incomplete and biased perspective of clients. We conducted a computational narrative analysis of child-welfare casenotes and draw attention to deeper systemic risk factors that are hard to quantify but directly impact families and street-level decision-making. We found that beyond individual risk factors, the system itself poses a significant amount of risk where parents are over-surveilled by caseworkers and lack agency in decision-making processes. We also problematize the notion of risk as a static construct by highlighting the temporality and mediating effects of different risk, protective, systemic, and procedural factors. Finally, we draw caution against using casenotes in NLP-based systems by unpacking their limitations and biases embedded within them.

Caseworkers are trained to write detailed narratives about families in Child-Welfare (CW) which informs collaborative high-stakes decision-making. Unlike other administrative data, these narratives offer a more credible source of information with respect to workers’ interactions with families as well as underscore the role of systemic factors in decision-making. SIGCHI researchers have emphasized the need to understand human discretion at the street-level to be able to design human-centered algorithms for the public sector. In this study, we conducted computational text analysis of casenotes at a child-welfare agency in the midwestern United States and highlight patterns of invisible street-level discretionary work and latent power structures that have direct implications for algorithm design. Casenotes offer a unique lens for policymakers and CW leadership towards understanding the experiences of on-the-ground caseworkers. As a result of this study, we highlight how street-level discretionary work needs to be supported by sociotechnical systems developed through worker-centered design. This study offers the first computational inspection of casenotes and introduces them to the SIGCHI community as a critical data source for studying complex sociotechnical systems.

Algorithms have permeated throughout civil government and society, where they are being used to make high-stakes decisions about human lives. In this paper, we first develop a cohesive framework of algorithmic decision-making adapted for the public sector (ADMAPS) that reflects the complex socio-technical interactions between human discretion, bureaucratic processes, and algorithmic decision-making by synthesizing disparate bodies of work in the fields of Human-Computer Interaction (HCI), Science and Technology Studies (STS), and Public Administration (PA). We then applied the ADMAPS framework to conduct a qualitative analysis of an in-depth, eight-month ethnographic case study of algorithms in daily use within a child-welfare agency that serves approximately 900 families and 1300 children in the mid-western United States. Overall, we found that there is a need to focus on strength-based algorithmic outcomes centered in social ecological frameworks. In addition, algorithmic systems need to support existing bureaucratic processes and augment human discretion, rather than replace it. Finally, collective buy-in in algorithmic systems requires trust in the target outcomes at both the practitioner and bureaucratic levels. As a result of our study, we propose guidelines for the design of high-stakes algorithmic decision-making tools in the child-welfare system, and more generally, in the public sector. We empirically validate the theoretically derived ADMAPS framework to demonstrate how it can be useful for systematically making pragmatic decisions about the design of algorithms for the public sector.

This paper examines volitionality of Facebook usage, that is, which individuals feel they have a choice about whether or not to use the site. It analyzes data from two large surveys, conducted three years apart. Across the two surveys, a variety of factors impacted whether or not respondents saw their Facebook usage as a matter of their own choice, such as engaging in non-use behaviors, measures of Facebook addiction, a sense of their own agency, and, across both studies, level of education. These results expand on prior literature around technology use and non-use, especially in terms of which populations may feel obligated to use, or be unwillingly prevented from using, social media such as Facebook. Furthermore, they provide potential implications both for future work and for technology policy.

The U.S. Child Welfare System (CWS) is charged with improving outcomes for foster youth; yet, they are overburdened and underfunded. To overcome this limitation, several states have turned towards algorithmic decision-making systems to reduce costs and determine better processes for improving CWS outcomes. Using a human-centered algorithmic design approach, we synthesize 50 peer-reviewed publications on computational systems used in CWS to assess how they were being developed, common characteristics of predictors used, as well as the target outcomes. We found that most of the literature has focused on risk assessment models but does not consider theoretical approaches (e.g., child-foster parent matching) nor the perspectives of caseworkers (e.g., case notes). Therefore, future algorithms should strive to be context-aware and theoretically robust by incorporating salient factors identified by past research. We provide the HCI community with research avenues for developing human-centered algorithms that redirect attention towards more equitable outcomes for CWS.

The human affective state is a product of complex biological processes and environmental stimuli. Situation aware systems aim at identifying the affective state of an individual using data from a gamut of connected devices. The bottle necks for such systems include continuous data streams, mobility of the data collection apparatus, device ubiquity and the device cost. While there is research done using physiological sensors that can overcome these challenges, their accuracy is often dismal and the results are not granular, i.e. the affective state is singular. In this paper we present results from an experiment that enabled us to generate models to identify an individuals affective state as a mixture of emotional states and their respective activation's. Secondly, we show that the affective state of an individual is actually a mixture of emotional states ( amusement, anger, neutral, sad, fear and disgust). During an experimental study, 85 participants were induced with specific emotions using audio-visual stimulus. Physiological data including heart rate, blood volume pressure (BVP), inter beat interval(IBI) and electrodermal activity(EDA) along with a self-report indicating the levels of 6 emotional states that include Amusement, Anger, Sad, Disgust, Fear and Neutral was recorded. Additionally, we recorded a self-reported score for Anxiety. The videos used to induce emotions were validated in a recently published study in Psychology. The data collected was used to create models that identify the dominant emotional state and the emotional spectrum (activation levels of all emotional states) for an individual. We create a map between the physiological data and the dominant emotional state and also between physiological data and the self-report scores. In addition, we identify often overlooked characteristics of human emotion such as variability in perception and overlap of emotional states and finally create a topological map of emotional states based on physiological data.

Prior studies of technology non-use demonstrate the need for approaches that go beyond a simple binary distinction between users and non-users. This paper proposes a set of two different methods by which researchers can identify types of non/use relevant to the particular sociotechnical settings they are studying. These methods are demonstrated by applying them to survey data about Facebook non/use. The results demonstrate that the different methods proposed here identify fairly comparable types of non/use. They also illustrate how the two methods make different trade offs between the granularity of the resulting typology and the total sample size. The paper also demonstrates how the different typologies resulting from these methods can be used in predictive modeling, allowing for the two methods to corroborate or disconfirm results from one another. The discussion considers implications and applications of these methods, both for research on technology non/use and for studying social computing more broadly.

ASD (Autism Spectrum Disorder) is a physiological condition that inhibits individuals from functioning in society. Such individuals suffer from high anxiety levels. This leads to both verbal and non-verbal communicative impairments and impedes self-expression. While there is no cure, there are structured interventions that teach children with ASD to cope with anxiety and function in society. However, these interventions do not account for the variability within the population with ASD. As a consequence, certain interventions fail to make a meaningful impact and, in certain scenarios, prove to be detrimental to the mental health of the participant. While there are screening measures, such as surveys conducted to avoid such circumstances, their effectiveness in identifying prospective successful candidates is poor. In this paper we propose a data driven intervention screening method that would enable Autism clinics to screen individuals most likely to benefit from the intervention and, more importantly, identify individuals that would be negatively affected.

Computer Vision has become the poster child for Deep Learning. The image classification accuracy of convolutional neural nets on benchmark data sets has increased every year since their inception. This has been aided with advances in feature fusion. The increase in the availability of imagetext occurrence has lead to text augmented feature spaces that have lead to higher accuracy in in image classification tasks. However, these works are limited to instances where text is readily available. This study presents an approach to featurize text within natural images with the goal of augmenting image features for image classification tasks. Text extraction and featurization in natural images is a challenging task due to challenges in reliable text localization and OCR results, both being impeded by the variability in image text and errors in OCR. We overcome these challenges by implementing a novel bounding box concatenation algorithm and a novel feature boosting algorithm. The result is a pipeline that encodes an image into a text feature space. Classifiers trained on the text based feature space have comparable accuracy to the state of the art Convolutional Neural Nets (CNN's) while being significantly inexpensive computationally. Moreover, the augmentation of text features to image features generates a hybrid feature space with a higher information content for a classification problem when compared to a feature space comprised exclusively of image features. Thus, we see a rise in classification accuracy across all state of the art machine learning algorithms.