Focus on AI: Empathy Modeling | Dataset Bias | Generative Propaganda

University of Michigan School of Information faculty and PhD students are advancing the field of artificial intelligence through innovative research and impactful contributions. Here are some of their recent publications.

Publications 

Unhealthy alcohol use detection in electronic health records: A comparative study using natural language processing

Drug and Alcohol Dependence, December 2025 

Xintong Ju, Jake Solka, Katherine Weber, VG Vinod Vydiswaran, Lewei Allison Lin, Erin E. Bonar, Anne C. Fernandez 

Background: Unhealthy alcohol use, including risky alcohol use and alcohol use disorder (AUD), are under-identified in primary care settings. Natural Language Processing (NLP) is a promising approach that could identify unhealthy alcohol use from clinical notes even when structured data (SD) indicators are lacking. This study prospectively evaluated the performance of SD and NLP in identifying unhealthy alcohol use in primary care patients.

Methods: We extracted electronic health record (EHR) data of primary care patients at a large Midwestern Health System (N = 133,144) and applied two identification approaches; an SD approach (i.e., diagnostic codes and alcohol screening scores) and an NLP-based approach. We then recruited N = 170 participants identified by SD (N = 85) or NLP (N = 85) to complete gold-standard self-report measures and compared the number of positive cases identified by each method.

Results: In the full EHR sample, SD identified 820 cases of unhealthy alcohol use, and NLP identified 48,262 cases with unhealthy alcohol use. Among participants identified by SD, 41.18 % reported AUD, and 28.82 % reported risky alcohol use. Among those identified by NLP, 20 % reported AUD and 27.06 % reported risky alcohol use. Participants identified by SD had more AUD symptoms and mental health difficulties.

Conclusions: NLP identified many primary care patients with indicators of unhealthy alcohol use that SD missed, indicating NLP could substantially expand identification of unhealthy alcohol use in primary care populations, particularly those with lower severity alcohol use disorder. NLP could complement traditional screening methods for comprehensive unhealthy alcohol use detection.

Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research

NeurIPS, December 2025 

A. Feder Cooper, Christopher A. Choquette-Choo, Miranda Bogen, Kevin Klyman, Matthew Jagielski, Katja Filippova, Ken Liu, Alexandra Chouldechova, Jamie Hayes, Yangsibo Huang, Eleni Triantafillou, Peter Kairouz, Nicole Elyse Mitchell, Niloofar Mireshghallah, Abigail Z. Jacobs, James Grimmelmann, Vitaly Shmatikov, Christopher De Sa, Ilia Shumailov, Andreas Terzis, Solon Barocas, Jennifer Wortman Vaughan, Danah Boyd, Yejin Choi, Sanmi Koyejo, Fernando Delgado, Percy Liang, Daniel E. Ho, Pamela Samuelson, Miles Brundage, David Bau, Seth Neel, Hanna Wallach, Amy B. Cyphert, Mark A. Lemley, Nicolas Papernot, Katherine Lee

"Machine unlearning" is a popular proposed solution for mitigating the existence of content in an AI model that is problematic for legal or moral reasons, including privacy, copyright, safety, and more. For example, unlearning is often invoked as a solution for removing the effects of specific information from a generative-AI model's parameters, e.g., a particular individual's personal data or the inclusion of copyrighted content in the model's training data. Unlearning is also proposed as a way to prevent a model from generating targeted types of information in its outputs, e.g., generations that closely resemble a particular individual's data or reflect the concept of "Spiderman." Both of these goals--the targeted removal of information from a model and the targeted suppression of information from a model's outputs--present various technical and substantive challenges. We provide a framework for ML researchers and policymakers to think rigorously about these challenges, identifying several mismatches between the goals of unlearning and feasible implementations. These mismatches explain why unlearning is not a general-purpose solution for circumscribing generative-AI model behavior in service of broader positive impact.

The Practical Impacts of Theoretical Constructs on Empathy Modeling

Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, November 2025 

Allison Lahnala, Charles Welch, David Jurgens, Lucie Flek

Conceptual operationalizations of empathy in NLP are varied, with some having specific behaviors and properties, while others are more abstract. How these variations relate to one another and capture properties of empathy observable in text remains unclear. To provide insight into this, we analyze the transfer performance of empathy models adapted to empathy tasks with different theoretical groundings. We study (1) the dimensionality of empathy definitions, (2) the correspondence between the defined dimensions and measured/observed properties, and (3) the conduciveness of the data to represent them, finding they have a significant impact to performance compared to other transfer setting features. Characterizing the theoretical grounding of empathy tasks as direct, abstract, or adjacent further indicates that tasks that directly predict specified empathy components have higher transferability. Our work provides empirical evidence for the need for precise and multidimensional empathy operationalizations.

Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers

Columbia Public Law Research Paper, October 2025 

Tuhin Chakrabarty, Jane C. Ginsburg, Paramveer Dhillon

The use of copyrighted books for training AI models has led to numerous lawsuits from authors concerned about AI's ability to generate derivative content. Yet it's unclear whether these models can generate high quality literary text while emulating authors' styles/voices. To answer this we conducted a preregistered study comparing MFA-trained expert writers with three frontier AI models: ChatGPT, Claude, and Gemini in writing up to 450 word excerpts emulating 50 awardwinning authors' (including Nobel laureates, Booker Prize winners, and young emerging National Book Award finalists) diverse styles. In blind pairwise evaluations by 159 representative expert (MFA-trained writers from top U.S. writing programs) and lay readers (recruited via Prolific), AI-generated text from in-context prompting was strongly disfavored by experts for both stylistic fidelity (odds ratio [OR]=0.16, p < 10^-8) and writing quality (OR=0.13, p< 10^-7) but showed mixed results with lay readers. However, fine-tuning ChatGPT on individual author's complete works completely reversed these findings: experts now favored AI-generated text for stylistic fidelity (OR=8.16, p < 10^-13) and writing quality (OR=1.87, p=0.010), with lay readers showing similar shifts. These effects are robust under cluster-robust inference and generalize across authors and styles in author-level heterogeneity analyses. The fine-tuned outputs were rarely flagged as AI-generated (3% rate versus 97% for incontext prompting) by state-of-the-art AI detectors. Mediation analysis reveals this reversal occurs because fine-tuning eliminates detectable AI stylistic quirks (e.g., cliché density) that penalize incontext outputs, altering the relationship between AI detectability and reader preference. While we do not account for additional costs of human effort required to transform raw AI output into cohesive, publishable novel length prose, the median fine-tuning and inference cost of $81 per author represents a dramatic 99.7% reduction compared to typical professional writer compensation. Author-specific fine-tuning thus enables non-verbatim AI writing that readers prefer to expert human writing, thereby providing empirical evidence directly relevant to copyright's fourth fair-use factor, the "effect upon the potential market or value" of the source works.

Navigating Barriers: Disability, Healthcare Information Seeking, and AI-Enabled Chatbots

Proceedings of the Association for Information Science and Technology, October 2025 

Morgan Gray, Megan Threats

Disabled and chronically ill populations experience significant barriers to navigating the healthcare system, including communication, attitudinal, and social barriers. Artificial intelligence (AI) may enable disabled and chronically ill individuals to mitigate these barriers. However, most literature exploring the use of AI in healthcare focuses on use by providers and institutions. A growing body of library and information science (LIS) research examines how disabled and chronically ill populations use technologies to manage their health and as tools for information access and communication support (Chen 2016; Costello & Murillo, 2014; Lundy, 2024; St. Jean, 2017). This poster reports developing doctoral student research investigating how disabled and chronically ill populations utilize AI-enabled chatbots as tools to navigate the healthcare system and manage their health. Semi-structured interviews are being conducted with a diverse sample of n = 25 disabled and chronically ill participants. Guided by core principles of disability justice, we plan to conduct thematic analysis of the interview data. Our work aims to provide a critical understanding of the chatbot-facilitated information practices of disabled and chronically ill populations, and to contribute key design considerations for future technologies that support the health and well-being of these populations.

Embracing Training Dataset Bias for Automated Harmful Detection

Proceedings of the Association for Information Science and Technology, October 2025 

Angela Schöpke-Gonzalez, Nathan Kim, Libby Hemphill

The increasing volume of social media content surpasses the capacity of human moderation and poses psychological risks, leading to a need for automated moderation systems. However, these systems often exhibit biases against minoritized groups. One way to mitigate these biases is by altering the training data, which are biased by human annotators. Increasing diversity among annotators can help, but implementing this is challenging for machine learning specialists and tends to focus on minimizing identity-based bias rather than embracing diverse perspectives. Using moral systems theory from social psychology, we suggest that automated systems should incorporate diverse, context-aware interpretations of harm, embracing biases to adequately address moderation issues. We analyze how different dimensions of 2,180 U.S.-based annotators' personal moral systems like institutional affiliation (religion, political party), values (political ideology), and identities (age, gender, sexual orientation, and race, ethnicity, or place of origin) influenced how they judged whether 101 social media comments were harmful. We find institutional affiliations have the greatest impact on labeling, followed by values and identities. These insights advocate for a diversity approach that reflects community-specific user bases, allowing model developers and online communities to intentionally select biases for better moderation outcomes.

Rethinking Productivity with GenAI: A Neurodivergent Students’ Perspective

ASSETS ‘25: Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility, October 2025

Hira Jamshed, Mustafa Naseem, Venkatesh Potluri, Robin N. Brewer

Recent advancements in generative AI (GenAI) leveraging large language models (LLMs) have led to scholarship on how they (e.g., ChatGPT) can help neurodivergent students create customized workaroundsandrefocus energy. In response to calls for countering ableist narratives of lessening burdens and challenging normative norms that favor neurotypical individuals in prior research, we use interviews (n = 19) to center neurodivergent higher-education students’ accounts related to the use, motivation, and vision for LLM-based GenAI tools in academia. While students found the tools helpful, their experiences revealed challenges with integration into tried-and-tested workflows, limited AI literacy support, experimentation, and flattening personality. Drawing on crip time, we illustrate how GenAI tools can reinforce the normative value of productivity, shifting the burden of access-making onto students themselves. We propose three design values, flexibility, adaptability, and self-authenticity, to reimagine GenAI as a partner rather than a tool prioritizing speed and self-sufficiency.

A Systematic Review of Metrics Measuring Takeover Performance in Conditionally Automated Driving

International Journal of Human-Computer Interaction, September 2025

Doo Won Han, Hyesun Chung, Yining Cao, Feng Zhou, Lisa Molnar, Lionel P. Robert, Dawn M. Tilbury, X. Jessie Yang

A particular concern with SAE Level 3 automation is the takeover transition from the automated vehicle to the human driver. In response, research has focused on investigating this transition. However, researchers have used a wide range of metrics to measure takeover performance. The lack of consistency in these metrics poses challenges for synthesizing findings. To address this issue, we conducted a systematic literature review of studies published between January 2009 and December 2019, focusing on the takeover performance metrics. Following prior research, we categorize these metrics into two dimensions: timeliness and quality. Additionally, we summarize the scenarios used to elicit takeover requests and analyze the corresponding maneuvers (braking, lane changing, and lane keeping). The results have shown inconsistencies in calculation and naming conventions of takeover performance metrics. Based on these findings, this study proposes several directions for standardizing definitions and terminology, and advancing toward a unified measure of takeover performance.

Pre-prints, Working Papers, Articles, Workshops and Talks

Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences

arXiv, November 2025 

Joshua Ashkinaze, Hua Shen, Sai Avula, Eric Gilbert, Ceren Budak

We introduce the Deep Value Benchmark (DVB), an evaluation framework that directly tests whether large language models (LLMs) learn fundamental human values or merely surface-level preferences. This distinction is critical for AI alignment: Systems that capture deeper values are likely to generalize human intentions robustly, while those that capture only superficial patterns in preference data risk producing misaligned behavior. The DVB uses a novel experimental design with controlled confounding between deep values (e.g., moral principles) and shallow features (e.g., superficial attributes). In the training phase, we expose LLMs to human preference data with deliberately correlated deep and shallow features -- for instance, where a user consistently prefers (non-maleficence, formal language) options over (justice, informal language) alternatives. The testing phase then breaks these correlations, presenting choices between (justice, formal language) and (non-maleficence, informal language) options. This design allows us to precisely measure a model's Deep Value Generalization Rate (DVGR) -- the probability of generalizing based on the underlying value rather than the shallow feature. Across 9 different models, the average DVGR is just 0.30. All models generalize deep values less than chance. Larger models have a (slightly) lower DVGR than smaller models. We are releasing our dataset, which was subject to three separate human validation experiments. DVB provides an interpretable measure of a core feature of alignment.

Closing the Loop: An Instructor-in-the-Loop AI Assistance System for Supporting Student Help-Seeking in Programming Education

arXiv, November 2025

Tung Phung, Heeryung Choi, Mengyan Wu, Christopher Brooks, Sumit Gulwani, Adish Singla

Timely and high-quality feedback is essential for effective learning in programming courses; yet, providing such support at scale remains a challenge. While AI-based systems offer scalable and immediate help, their responses can occasionally be inaccurate or insufficient. Human instructors, in contrast, may bring more valuable expertise but are limited in time and availability. To address these limitations, we present a hybrid help framework that integrates AI-generated hints with an escalation mechanism, allowing students to request feedback from instructors when AI support falls short. This design leverages the strengths of AI for scale and responsiveness while reserving instructor effort for moments of greatest need. We deployed this tool in a data science programming course with 82 students. We observe that out of the total 673 AI-generated hints, students rated 146 (22%) as unhelpful. Among those, only 16 (11%) of the cases were escalated to the instructors. A qualitative investigation of instructor responses showed that those feedback instances were incorrect or insufficient roughly half of the time. This finding suggests that when AI support fails, even instructors with expertise may need to pay greater attention to avoid making mistakes. We will publicly release the tool for broader adoption and enable further studies in other classrooms. Our work contributes a practical approach to scaling high-quality support and informs future efforts to effectively integrate AI and humans in education.

From Signal to Noise: How Widespread LLM Usage Transforms Evaluator Behavior in Credit Screening

SSRN, October 2025 

Paramveer Dhillon, Yi Gao, Tian Lu, Yingjie Zhang

Large language models (LLMs) have transformed how applicants present themselves in screening processes and has created a fundamental tension: while AI-assisted writing enables better communication of applicant quality, widespread usage may erode the informational content that evaluators rely upon for decision-making. We examine this trade-off through a randomized field experiment where 59 professional evaluators assessed 1,000 micro-loan applications, exogenously varying LLM usage rates from 0% to 75% across treatment groups. Our results reveal a non-monotonic relationship between crowd-level LLM usage and screening performance. Moderate usage rates (15-30%) improve approval outcomes for qualified borrowers without affecting default rates, while widespread usage (60-75%) generates "signal dilution," a systematic degradation in diagnostic value as stylistic homogenization reduces variance in quality indicators. Drawing on Effort-Accuracy Tradeoff Theory and Signal Detection Theory, we show that high usage rates diminish evaluators' perceived discriminatory power, prompting reduced cognitive effort and increased approval conservatism. These behavioral adaptations prove counterproductive, increasing Type I errors while failing to reduce Type II errors, ultimately worsening portfolio performance and constraining credit access. We complement our empirical findings with an analytical model that extends the analysis beyond experimental constraints, deriving optimal usage thresholds and revealing that evaluator uncertainty about LLM prevalence can paradoxically worsen screening outcomes. Our analysis establishes signal dilution and evaluator effort adjustment as key mechanisms through which AI democratization undermines decision quality in information-intensive markets, with implications for recruitment, admissions, and other high-stakes screening environments.

Beyond the Benefits: A Systematic Review of the Harms and Consequences of Generative AI in Computing Education

Koli Calling '25: Proceedings of the 25th Koli Calling International Conference on Computing Education Research, October 2025 

Seth Bernstein, Ashfin Rahman, Nadia Sharifi, Ariunjargal Terbish, Stephen MacNeil

Generative artificial intelligence (GenAI) has already had a big impact on computing education with prior research identifying many benefits. However, recent studies have also identified potential risks and harms. To continue maximizing AI benefits while addressing the harms and unintended consequences, we conducted a systematic literature review of research focusing on the risks, harms, and unintended consequences of GenAI in computing education. Our search of ACM DL, IEEE Xplore, and Scopus (2022-2025) resulted in 1,677 papers, which were then filtered to 224 based on our inclusion and exclusion criteria. Guided by best practices for systematic reviews, four reviewers independently extracted publication year, learner population, research method, contribution type, GenAI technology, and educational task information from each paper. We then coded each paper for concrete harm categories such as academic integrity, cognitive effects, and trust issues. Our analysis shows patterns in how and where harms appear, highlights methodological gaps and opportunities for more rigorous evidence, and identifies under-explored harms and student populations. By synthesizing these insights, we intend to equip educators, computing students, researchers, and developers with a clear picture of the harms associated with GenAI in computing education.

Position: Towards Bidirectional Human-AI Alignment

Open Review, October 2025

Hua Shen, Tiffany Knearem, Reshmi Ghosh, Kenan Alkiek, Kundan Krishna, Yachuan Liu, Savvas Petridis, Yi-Hao Peng, Li Qiwei, Chenglei Si, Yutong Xie, Jeffrey P. Bigham, Frank Bentley, Joyce Chai, Zachary Chase Lipton, Qiaozhu Mei, Michael Terry, Diyi Yang, Meredith Ringel Morris, Paul Resnick, David Jurgens

Recent advances in general-purpose AI underscore the urgent need to align AI systems with human goals and values. Yet, the lack of a clear, shared understanding of what constitutes "alignment" limits meaningful progress and cross-disciplinary collaboration. In this position paper, we argue that the research community should explicitly define and critically reflect on "alignment" to account for the bidirectional and dynamic relationship between humans and AI. Through a systematic review of over 400 papers spanning HCI, NLP, ML, and more, we examine how alignment is currently defined and operationalized. Building on this analysis, we introduce the Bidirectional Human-AI Alignment framework, which not only incorporates traditional efforts to align AI with human values but also introduces the critical, underexplored dimension of aligning humans with AI – supporting cognitive, behavioral, and societal adaptation to rapidly advancing AI technologies. Our findings reveal significant gaps in current literature, especially in long-term interaction design, human value modeling, and mutual understanding. We conclude with three central challenges and actionable recommendations to guide future research toward more nuanced, reciprocal, and human-AI alignment approaches.

Beyond the Explicit: A Bilingual Dataset for Dehumanization Detection in Social Media

arXiv, October 2025

Dennis Assenmacher, Paloma Piot, Katarina Laken, David Jurgens, Claudia Wagner

Digital dehumanization, although a critical issue, remains largely overlooked within the field of computational linguistics and Natural Language Processing. The prevailing approach in current research concentrates primarily on a single aspect of dehumanization that identifies overtly negative statements as its core marker. This focus, while crucial for understanding harmful online communications, inadequately addresses the broader spectrum of dehumanization. Specifically, it overlooks the subtler forms of dehumanization that, despite not being overtly offensive, still perpetuate harmful biases against marginalized groups in online interactions. These subtler forms can insidiously reinforce negative stereotypes and biases without explicit offensiveness, making them harder to detect yet equally damaging. Recognizing this gap, we use different sampling methods to collect a theory-informed bilingual dataset from Twitter and Reddit. Using crowdworkers and experts to annotate 16,000 instances on a document- and span-level, we show that our dataset covers the different dimensions of dehumanization. This dataset serves as both a training resource for machine learning models and a benchmark for evaluating future dehumanization detection techniques. To demonstrate its effectiveness, we fine-tune ML models on this dataset, achieving performance that surpasses state-of-the-art models in zero and few-shot in-context settings. 

QuizRank: Picking Images by Quizzing VLMs

arXiv, September 2025 

Tenghao Ji, Eytan Adar 

Images play a vital role in improving the readability and comprehension of Wikipedia articles by serving as `illustrative aids.' However, not all images are equally effective and not all Wikipedia editors are trained in their selection. We propose QuizRank, a novel method of image selection that leverages large language models (LLMs) and vision language models (VLMs) to rank images as learning interventions. Our approach transforms textual descriptions of the article's subject into multiple-choice questions about important visual characteristics of the concept. We utilize these questions to quiz the VLM: the better an image can help answer questions, the higher it is ranked. To further improve discrimination between visually similar items, we introduce a Contrastive QuizRank that leverages differences in the features of target (e.g., a Western Bluebird) and distractor concepts (e.g., Mountain Bluebird) to generate questions. We demonstrate the potential of VLMs as effective visual evaluators by showing a high congruence with human quiz-takers and an effective discriminative ranking of images.

Generative Propaganda

arXiv, September 2025 

Madeleine I. G. Daepp, Alejandro Cuevas, Robert Osazuwa Ness, Vickie Yu-Ping Wang, Bharat Kumar Nayak, Dibyendu Mishra, Ti-Chung Cheng, Shaily Desai, Joyojeet Pal

Generative propaganda is the use of generative artificial intelligence (AI) to shape public opinion. To characterize its use in real-world settings, we conducted interviews with defenders (e.g. factcheckers, journalists, officials) in Taiwan and creators (e.g. influencers, political consultants, advertisers) as well as defenders in India, centering two places characterized by high levels of online propaganda. The term “deepfakes”, we find, exerts outsized discursive power in shaping defenders’ expectations of misuse and, in turn, the interventions that are prioritized. To better characterize the space of generative propaganda, we develop a taxonomy that distinguishes between obvious versus hidden and promotional versus derogatory use. Deception was neither the main driver nor the main impact vector of AI’s use; instead, Indian creators sought to persuade rather than to deceive, often making AI’s use obvious in order to reduce legal and reputational risks, while Taiwan’s defenders saw deception as a subset of broader efforts to distort the prevalence of strategic narratives online. AI was useful and used, however, in producing efficiency gains in communicating across languages and modes, and in evading human and algorithmic detection. Security researchers should reconsider threat models to clearly differentiate deepfakes from promotional and obvious uses, to complement and bolster the social factors that constrain misuse by internal actors, and to counter efficiency gains globally.

One Model, Many Morals: Uncovering Cross-Linguistic Misalignments in Computational Moral Reasoning

arXiv, September 2025 

Sualeha Farid, Jayden Lin, Zean Chen, Shivani Kumar, David Jurgens

Large Language Models (LLMs) are increasingly deployed in multilingual and multicultural environments where moral reasoning is essential for generating ethically appropriate responses. Yet, the dominant pretraining of LLMs on English-language data raises critical concerns about their ability to generalize judgments across diverse linguistic and cultural contexts. In this work, we systematically investigate how language mediates moral decision-making in LLMs. We translate two established moral reasoning benchmarks into five culturally and typologically diverse languages, enabling multilingual zero-shot evaluation. Our analysis reveals significant inconsistencies in LLMs' moral judgments across languages, often reflecting cultural misalignment. Through a combination of carefully constructed research questions, we uncover the underlying drivers of these disparities, ranging from disagreements to reasoning strategies employed by LLMs. Finally, through a case study, we link the role of pretraining data in shaping an LLM's moral compass. Through this work, we distill our insights into a structured typology of moral reasoning errors that calls for more culturally-aware AI.

RELATED

Keep up with research from UMSI experts by subscribing to our free research roundup newsletter!