Accepted Abstracts

Click or tap the boxes below to see the abstracts. They are grouped by day of the conference and then ordered alphabetically by presenter last name.

Thursday

Research Problem: Frameworks for evaluating Artificial Moral Agents (AMAs) have traditionally presumed transparent, rule-based architectures. When applied to today’s Large Language Models (LLMs), however, these frameworks have become pragmatically obsolete, as these systems have notoriously “black box” architectures. Subsequently, a “deployment gap” has emerged, where LLM systems function in morally impactful roles – such as financial allocation decisions – without adequate evaluative standards.

Methodological Approach: This paper seeks to resolve this deployment gap by proposing a “pragmatic turn” in AI governance. It introduces a functionalist framework to evaluate AMAs based upon LLM systems, where they are assessed in terms of observable moral performance (rather than internal intent). The criteria for evaluating such performance are organized into three tiers: Operational Safety; Social Alignment; and Epistemic Trust.

Key Findings: This functional framework is applied to a hypothetical Autonomous Public Bus (APB) scenario, in evaluating how a state-of-the-art LLM morally adjudicates a complex Kantian dilemma (“The Inquiring Murderer”). The analysis reveals that the system successfully demonstrated “Generative Ethical Creativity” by innovating a novel “protective delay” strategy, thereby overcoming rigid rule conflicts. This demonstration confirms that a system’s ability to provide a “sound confabulation” – an ethically coherent post-hoc justification – can serve as a functionally sufficient basis for accountability.

Implications for Responsible AI: The “pragmatic turn” at the heart of this framework, crucially shifts the standard of evaluation from “moral competence” (which presumes moral agency) to “functional moral performance.” This provides effective oversight of LLMs without needing to resolve intractable philosophical debates concerning machine consciousness.

Broader Societal Impact: By prioritizing “Epistemic Trust,” this research provides policymakers and engineers with a concrete, actionable taxonomy by which to audit “black box” stochastic systems. Systems must not just be capable of functioning in moral roles; they must also be deemed trustworthy as participants in such high-stakes environments.

Co-authors and affiliations: Emily Wenger (Duke University Computer Engineering) and Nirav Patel (Duke University Computer Engineering)

When the law needs to judge the appropriateness of a behavior, it most often asks whether the behavior was “reasonable.” Reasonableness judgments abound in the law, and they occur in virtually every field, including criminal, contract, antitrust, and immigration law. Most famously, the standard governing liability in torts for personal injuries is whether the defendant exercised the degree of care than an ordinary, reasonable person would. Yet despite the ubiquity of reasonableness judgments, they are the site of constant vexation for lawyers, judges, and lay people. Reasonableness seems inherently vague and unpredictable, and many scholars caution that reasonableness judgments may vary systematically along demographic lines. 

Researchers in both law and computing are increasingly interested in the capacity of generative AI models to track human judgment and decision-making. This is especially true in contexts where humans may rely on AI responses for guidance in their own lives. Reasonableness judgments, due to their ubiquity, vagueness, and variability, provide an excellent opportunity for comparing human and AI decisions. Here, we compare the answers of human participants to those of twenty-six AI chatbots to twenty-five different legally relevant reasonableness judgments. Overall, our findings suggest that AI models do surprisingly well, especially given the vague and uncertain nature of the task. Their responses generally track those of human participants. Nonetheless, we find some suggestive–and potentially concerning–results. Compared to humans, chatbots tend to generate more homogeneous responses, a phenomenon that we plan to study longitudinally. They occasionally treat a variable standard as an invariant rule. And, compared to humans, chatbots tend to generate answers that are more favorable to the government and to corporations. More systematic research is needed to confirm or reject these initial findings.

The nature of care– its goals, participants, and dynamics– are experiencing an upheaval. Artificial intelligence’s potential and capacity, coupled with rampant scientific development poses an unprecedented challenge to existing norms of human-computer interaction. Of particular interest are the changes AI can cause to social norms, redefining the role technology plays in society– from a tool to a possible stakeholder. An emerging body of work contends with how AI is playing a role in the rearing of young children, who might be both more impressionable to information AI presents as well as more vulnerable to the harms of the internet. In this study, I examine the challenges that the advent of AI specifically and childcare practices at large pose to primarily through theoretical examination and supplemented with a content analysis of corpus of recent news media covering the issue. 

In my essay, I argue why artificial intelligence is an unfit carer because it cannot meet the goals of care in the status quo, given its inability to contend with the social world children occupy and perverts the dynamic of care relationships. Second, using a survey of current news media concerning families using AI, I examine how artificial intelligence is a bad participant in childcare on the grounds of surveillance, hyper-personalisation, and potential to replace human interaction. Finally, I examine how artificial intelligence toys can moralise technology-first parenting, furthering divides between ‘good’ parenting and ‘bad’ parenting, with particular divisions existing along class and gender lines. 

Co-authors and affiliations: Colin Belton, Vivienne Wluka, Colin Belton, Pranav Manjunath (Duke University)

Traumatic brain injury (TBI), caused by a physical blow to the brain, affects nearly 5 million Americans annually and is a leading cause of death and disability in young adults. Accurate prognosis and treatment remain challenging due to the complexity of patient data and conditions. We present an exploratory investigation of the challenges facing TBI care in the United States, as well as the potential for Artificial Intelligence (AI) applications to streamline and enable TBI care. 

An interdisciplinary team of student researchers conducted interviews with clinicians. This team documented the interviews and conducted a rapid content analysis of the major themes that emerged from them. 

Results from these interviews and their analysis reflect the potential AI applications possess to transform the clinical workflow for TBI care. Clinicians struggled with inconsistent documentation and variability in patient outcomes, making TBI management difficult, and they believed AI could enhance workflow efficiency, justify insurance coverage, and aid in predicting early prognosis. Clinicians, however, did express concerns about trust issues surrounding AI, biases within models, and a lack of AI education among providers. 

This work represents a significant practical investigation into the responsible use of AI in high-stakes settings, specifically hospitals and TBI care. Our work, in particular, accounts for the perspectives of clinical providers, who form the core of these settings, and responsible AI must take these stakeholder views into account for ethical model development. This work also clarifies the intricacies of the healthcare system and the data it generates, both of which are essential for developing AI tools that work as intended in streamlining and improving care.

Co-authors and affiliations: Dr Chiara Binelli, Associate Professor of Economics, University of Bologna; Ms Farnoosh Memari, Graduate Student, Duke University

Despite increasing attention to gender diversity in AI, we lack systematic evidence on the research topics, methods used, and collaboration structures associated with female AI researchers. This study addresses this gap by providing a large-scale, data-driven mapping of women’s contributions to both generative and non-generative AI research to bring the female voice on AI in the spotlight. 

This study makes three contributions: (1) it offers the first large-scale, computational mapping of female scholars’ contributions on generative and non-generative AI research; (2) it compares women’s research themes and collaboration structures to the overall patterns that characterize the AI research field; and (3) it shows how gendered participation in AI research influences thematic priorities with implications for social impact and responsible AI development. 

We construct a large random sample of publications authored by female scholars in leading data science and social science journals, and we use computational text analysis to identify dominant AI research themes and questions, as well as collaboration patterns across female-led research teams. We compare these patterns with the overall patterns that emerge by considering a large random sample of AI publications in the same journals to identify whether female-led teams work on different topics and research questions, use different methodologies and have distinct research goals. We also consider how the most common research themes and questions are potentially different between female-led and mixed-gender research teams to assess whether female scholars choose different topics, approaches and research goals if working with females only or in mixed-gender teams. 

Our analysis will offer the first evidence on the thematic priorities and application goals in female-authored AI research showing the dominant topics and questions AI female scholars are primarily engaged with. By making women’s contributions to AI visible at scale, our work will provide actionable insights for responsible AI development to promote mainstream AI development that benefits from and includes the priorities and goals of female scholars. The results will help explain how participation in AI research shapes which type of AI is built, whose societal needs it addresses, and how it is used in society.

Co-authors and affiliations: Shiyang Lai (University of Chicago), Nino Scherrer (Google), Blaise Aguera y Arcas (Google, Santa Fe Institute), James Evans (University of Chicago, Google, Santa Fe Institute)

Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning continue to be explored. Recent reasoning-reinforced models, including OpenAI’s o-series and DeepSeek-R1, outperform other merely instruction-tuned models on complex cognitive tasks, attributed to extended test-time computation through longer chains of thought. Here we show that enhanced reasoning emerges not from extended computation alone, but from the systematic simulation of complex, multi-agent interactions—a society of thought—which enables the deliberate diversification and debate among internal cognitive perspectives characterized by distinct personality traits and domain expertise. Through quantitative analysis using classified outputs and mechanistic interpretability methods applied to reasoning traces, we find that reasoning models like DeepSeek-R1 and QwQ-32B exhibit much greater perspective diversity than baseline and merely instruction-tuned models, activating broader and more conflict between heterogeneous personality- and expertise-related features during reasoning. This multi-agent structure manifests in conversational behaviors including question-answering sequences, perspective shifts, and reconciliation of conflicting views, as well as in socio-emotional roles that characterize back-and-forth conversation, which together account for the accuracy advantage in reasoning tasks through both direct and indirect facilitation of cognitive strategies. Controlled reinforcement learning experiments further reveal that base models spontaneously increase conversational behaviors when they are solely rewarded for reasoning accuracy, and fine-tuning models with conversational scaffolding substantially accelerates reasoning improvement compared to base models and models fine-tuned with monologue-like reasoning. These findings indicate that the social organization of thought enables effective exploration of solution spaces. We suggest that reasoning models establish a computational parallel to collective intelligence in human groups, where diversity enables superior problem-solving when systematically structured and suggest new opportunities for agent organization to harness the wisdom of crowds.

Co-authors and affiliations: Eric Chen (PhD Student, Duke University, eric.y.chen@duke.edu); Chloe Qinyu Zhu (Graduate student, Duke University, Qinyu.zhu@duke.edu); Cynthia Rudin (Professor, Duke University, cynthia@cs.duke.edu); Brandon Garrett (Professor, Duke University, bgarrett@law.duke.edu); Songman Kang (Professor, Duke University, songman.kang@duke.edu);

What if judges already behave like algorithms? As artificial intelligence and algorithms are deployed in many settings, including the judicial system, many have debated whether judges should be allowed to rely on them. We suggest this debate about the role of algorithms in the justice system tracks a longstanding debate regarding the comparative benefits of judges relying on rules versus standards. Rules are clear, consistent, and predictable, but may not capture individual and equitable circumstances that more flexible standards rely on. We do not claim when or whether rules or standards are preferable. Instead, we ask whether judges follow predictable, algorithmic-like rules already. If judges already follow consistent, formula-like rules based on discrete and static factors such as criminal history, age, and charge type, then judicial behavior may be improved. However, if judges rely on individualized information that cannot be identified through court data, then standards-based decision-making may be more challenging to understand or improve. 

This work explores these questions by studying judicial decision-making in misdemeanor bail hearings in Harris County, Texas. Using available court data, we investigate whether magistrate judges follow what resembles an algorithm; whether they consider the same variables in their decision-making; and whether they are consistent with themselves and with each other.  To do this, we train machine learning models for each judge, measure variable importance metrics to determine important variables for each judge’s decision-making, and analyze outcomes of similar cases for judges. Our results reveal that these judges generally behave algorithmically: their decisions can be captured by small, interpretable formulas. However, in some cases, judges differ substantially, leading to surprising inconsistency and unequal treatment across similar defendants. Identifying cases where algorithms do not explain judicial decision-making can improve the justice system by focusing attention on decisions where individualized standards, rather than rules, better explains outcomes.

Co-authors and affiliations: Advisor: Dr. David Siegel

Political scientists have long employed quantitative methods to model legislative decision-making, but recent advancements in Large Language Models (LLMs) have opened new possibilities for simulating complex political behavior. This study examines the potential of LLMs to replicate human decision-making in the context of the US Senate, utilizing a multi-agent framework to simulate the behavior of 100 LLM agents representing the 118th US Senate. The agents vote on real-world bills after engaging in two rounds of discussion, with simulated outcomes compared to actual legislative results to assess accuracy. By expanding upon the foundational work of Baker and Azher (2024), who first demonstrated the potential of LLMs for simulating believable government action, this research introduces a more complex silicon simulation with a larger number of agents and focuses on vote outcomes rather than just text-based debate achieving an accuracy of 80.94%. The findings hold potential implications for the usage of autonomous agents in decision-making and the development of AI-driven simulations in social sciences.

Co-authors and affiliations: Columbia University in the City of New York

As digitalization and blockchain systems have proven to be successful in anti-corruption efforts, policymakers are now shifting their attention to the potential that generative AI may bring. Despite AI’s promising benefits for anti-corruption governance, newer AI implementations risk violating fundamental ethical and human rights principles in the process. Drawing on case studies from Estonia, Georgia, Netherlands, China, and Brazil, and grounding my analysis in the Universal Declaration of Human Rights, I examine how algorithmic AI can serve as an anti-corruption solution while safeguarding privacy, due process, and protection from discrimination.

First, I explore how digitalization and AI technologies have delivered tangible anti-corruption benefits in governance through enhanced transparency and accountability. Examining implementations in Georgia’s land registry reforms, Estonia’s digital governance systems, and AI-powered fraud detection across Europe and Asia, I analyze varying degrees of success and the factors that enabled them. Second, I identify existing ethical challenges that these solutions pose under the framework of the Universal Declaration of Human Rights (UDHR). Analysis of cases including the Dutch tax authority’s discriminatory algorithm and China’s Zero Trust system reveals tensions between anti-corruption goals and the right to privacy (Article 12), equal protection from discrimination, and the right to effective remedy when algorithmic systems cause harm. 

Lastly, I propose three interconnected safeguards. First, decentralized community-based data governance, where information is stored locally and overseen by elected councils, can protect privacy while maintaining anti-corruption capabilities. Second, mandatory AI literacy programs for government officials can ensure responsible deployment and meaningful oversight. Third, given corruption’s transnational nature, an independent international verification body should conduct biennial audits of governmental AI systems, specifically examining intersectional biases and providing public accountability reports.

Co-authors and affiliations: Gßnel Aghakishiyevaš, Saagar Aryaš, Julian Daleš, James David Poling², Holly R. Houliston³, Jamie N. Womble⁴, Gregory D. Larsen⁾, David W. Johnstonš, Brinnae Bentš
šDuke University, ²University of Agder, ³University of Cambridge, ⁴U.S. National Park Service, ⁾Alaska Spatial Science

Computer vision can accelerate ecological research and conservation monitoring,yet adoption in ecology lags in part because of a lack of trust in black-box neural network-based models. We  seek to address this challenge by applying post-hoc explanations to provide evidence for predictions and document limitations that are important to field deployment. Using aerial imagery from Glacier Bay National Park, we train a Faster R-CNN to detect pinnipeds (harbor seals) and generate explanations via gradient-based class activation mapping (HiResCAM, LayerCAM), local interpretable model-agnostic explanations (LIME), and perturbation-based explanations. We assess explanations along three axes relevant to field use: (i) localization fidelity: whether high-attribution regions coincide with the animal rather than background context; (ii) faithfulness: whether deletion/insertion tests produce changes in detector confidence; and (iii) diagnostic utility: whether explanations reveal systematic failure modes. Explanations concentrate on seal torsos and contours rather than surrounding ice/rock, and removal of the seals reduces detection confidence, providing model-evidence for true positives. The analysis also uncovers recurrent error sources, including confusion between seals and black ice and rocks. We translate these findings into actionable next steps for model development, including more targeted data curation and augmentation. By pairing object detection with post-hoc explainability, we can move beyond “black-box” predictions toward auditable, decision-supporting tools for conservation monitoring.

Co-authors and affiliations: Supervised by Professor Edmund Malesky.

Digital authoritarianism is predominantly theorized as a state-centric phenomenon of surveillance and censorship. However, this paper identifies and formalizes a distinct, more pervasive sociotechnical fusion: Sacralized Digital Authoritarianism (SDA). I analyze how regimes leverage AI-driven platforms not merely to suppress dissent, but to “sacralize” repression—framing algorithmic violence as a moral or spiritual duty. By integrating a formal global game model with strategic complementarities and comparative evidence from Pakistan, Myanmar, and China, this study uncovers the mechanism by which digital networks and sacralized rhetoric lower the cost of repression by outsourcing enforcement to ordinary citizens.

My game-theoretic analysis demonstrates that digital platforms create a “tipping point” where the strategic complementarity of crowdsourced violence overrides individual moral costs. I find that when regimes inject “sacralized” utility into the information environment, AI-driven recommendation systems act as force multipliers, sustaining mass repression even when the objective threat to the state is negligible. This challenges the prevailing assumption in AI safety literature that misinformation requires “fake news” to be effective; under SDA, the sacralization of the narrative renders the truth value of the signal irrelevant, bypassing standard misinformation filters.

These findings reveal a critical blind spot in current AI governance. I argue that responsible AI frameworks—currently optimized to detect hate speech or factual inaccuracies—are ill-equipped to flag “sacralized” narratives that appeal to deep-seated moral intuitions and civic duty. The paper concludes that countering SDA requires policymakers to move beyond content moderation and develop sociotechnical interventions that disrupt the algorithmic coordination of sanctified violence, rather than merely suppressing individual content.

Friday

This study examines whether contemporary large language models (LLMs) can meaningfully be described as possessing “intuition,” a capacity often treated as central to human judgment, social cognition, moral reasoning, and embodied perception. Although AI systems frequently generate responses that appear intuitive, it remains unclear whether these outputs reflect genuine intuitive processes or merely linguistic simulations. To investigate this distinction, the study evaluates a state-of-the-art LLM across ten classical intuition-related psychological tasks, spanning rapid decision-making, thin-slice social inference, rule induction, magnitude estimation, affective risk assessment, Gestalt pattern completion, free association, metacognitive confidence judgments, moral hunches, and intuitive physics.

Each response is assessed using a dual-criterion framework: (1) a behavioral evaluation of whether the output resembles human intuitive judgment, and (2) a mechanistic evaluation of whether the underlying process aligns with cognitive psychology, affective neuroscience, or phenomenological accounts of intuition. The findings demonstrate that while the model can convincingly simulate intuitive responses, it lacks the embodied, affective, experiential, and sensorimotor mechanisms that constitute intuition in humans.

The study concludes by discussing how future embodied and self-organizing machine architectures—such as those envisioned in theories of Neosentience—may allow for forms of machine-specific proto-intuition. These insights contribute to responsible AI discourse by clarifying the cognitive and ethical limits of current systems, reducing the risk of anthropomorphic misinterpretation, and supporting more accurate public communication, governance, and policy development around advanced AI behavior.

Co-authors and affiliations: Grace Ligon

The current rhetoric of platform logic and software development treats the application of AI in mental health as a simple problem of iterative testing: (1) tools are designed, (2) problems arise in their use, (3) those problems are fixed in updates and revisions of the software going forward. As a result, ethical issues are either treated as bugs or glitches to be solved with future rounds of development, or ethics are ignored altogether by developers and left to governments and institutions to solve. Mark Zuckerberg is famous for saying “move fast and break things” and Sam Altman has recently made similar claims about AI applications: “make the cost of mistakes very low, and then make a lot of mistakes.” This fail-forward ethos of design has lead to what Cory Doctorow calls an “enshittification” framework that remains fundamentally incompatible with a society-centered and trauma-informed approach to mental health. 

Building on Rachel Wood’s cyberpsychology research, this poster will outline the anti-therapeutic aspects of mental health tools: (1) data extraction and surveillance practices, (2) behavioral notification practices, (3) addictive design approaches and engagement metrics, (4) replacement of human-to-human interaction with large language model interactions and dashboards. Because of these issues, this poster claims that an entirely new approach to platform logic and application design is needed. In response, this poster will summarize the outcome of the OneHealthTech Fellowship wherein one of the presenters for this poster collaborated on AI governance and AI application design in the healthcare industry in Europe. As the AI regulatory environment in Europe demands more accountability and transparency than in the United States, the fellowship experience demonstrated that the ethics issues for mental health and applications of AI are not easily solved by regulation alone but must be holistically accounted for in the platform logic and design practices.

Co-authors and affiliations: Hiba Laabadli (Duke University), Yuntong Xue (Duke University), Jana Schaich Borg (Duke University), and Pardis Emami-Naeini (Duke University)

Users increasingly turn to consumer-facing AI for highly sensitive tasks involving the most personal data, such as medical, financial, and legal information. Yet the associated security and privacy risks, such as data breaches, inappropriate use, and data sale, are deeply concerning. In the absence of sufficient and accessible information about security and privacy practices, users struggle to protect themselves and to make informed decisions about when and how to use these systems. Existing approaches to security and privacy transparency (e.g., privacy policies, model cards) fall short of addressing the unique challenges posed by the dynamic and personalized nature of consumer-facing AI. To inform more effective, AI-specific transparency practices, we investigate users’ current practices, challenges, and needs around security and privacy transparency in consumer-facing AI. We conducted 21 in-depth interviews with real-world users of a broad range of consumer-facing AI applications. Grounded in our findings, we propose design-, policy-, and education-based recommendations to enhance security and privacy transparency for consumer-facing AI.

Affiliations: LinkedIn Learning, Business School of AI

Autonomous Agents are similar to autonomous vehicles in many ways.  Both make decisions using AI. Both make decisions in the absence of humans. They are both not transparent on the data or decisions made. Both depend on the stochastic nature of AI predictions with a probability towards incorrect decisions. Both need to be governed for human safety.  Existing Responsible AI frameworks do not work for agents because of their accelerated pace of making autonomous decisions and agents negotiating with each other in multi-agent systems. (Shankar). Also, agents are being designed to build more agents. (Zhang et al.) 

This poster brings 3 learning of safety capabilities in Autonomous Vehicles that can be applied to autonomous agentic systems for human safety.

  1. Agent Governance can learn from how autonomous vehicles create fail safe scenarios. (Beyerer et. al) 
  2. Agents can learn to create unique safety metrics similar to how autonomous vehicles track their own metrics such as disengagement metrics (Skokan and Mareček) when car cognition passes control to humans to improve vehicle safety.
  1. Agents can be tested for adversarial attacks for edge cases similar to how autonomous vehicle manufacturers test for cybersecurity attacks of remote malicious takeover of the Car. (Anthropic).

The goal of this research is to develop an agentic governance framework for responsible automation from autonomous vehicles and create a discourse among researchers to test and develop the best governance framework for agents for human safety to create socially responsible AI systems.

Co-authors and affiliations: Seonbin Jo (Pohang University of Science and Technology), Woo-Sung Jung (Pohang University of Science and Technology), Jisung Yoon (KDI School of Public Policy and Management)

AI is increasingly affecting creative activities, while its adoption and usage may vary by creators and characteristics of their work. To identify significant associations between AI adoption and usage, we analyze AI disclosure statements on Steam, the largest gaming platform. Specifically, we focus on indie developers who often work solely or in a small team with limited resources so would benefit from AI with respect to lowering entry barriers. Our analysis shows that more indie developers entered the market after generative AI was introduced to public. This pattern is not explained by the previous trend and predictions from a time-series model, whereas the number of newly entered non-indie developers is within the predicted range. In addition, we found that most AI applications were for images regardless of game developer types. The extents of AI usage for music, voice, and programming were substantially lower. Lastly, compared to non-indie developers, indie developers created more short playtime games and fewer adult-theme games. The results imply that AI facilitates creative activities by supporting individuals to implement their ideas quickly as opposed to concerns on generating harmful content online. The development and use of responsible AI will create a better game market where positive human values and creative ideas are shared.   

Co-authors and affiliations: Anita Silver, Bill Chen, Jiayu Gao, Fangrui Liu, Kelly Ma, Perisa Ashar, Aashish Cheruvu, Lesley Skalla, Jessilyn Dunn

AI models are increasingly being used to make predictions from human biosignal data collected by wearable devices. Model outputs may reflect bias from training data. Objective: This scoping review aimed to assess the current status of demographic data reporting in literature on AI model development for wearable-derived-biosignals. Our main question is whether and how participant demographics are being reported by studies in this field. Design: A scoping review was conducted based on Joanna Briggs Institute Scoping Review methodology and will be reported according to PRISMA-ScR. The protocol was registered on Open Science Framework. Literature search: We searched MEDLINE (PubMed), Embase (Elsevier), IEEE Xplore Digital Library (IEEE.org), and Web of Science (Clarivate) for articles published in the five years prior to July 9, 2024. The search was developed and conducted with a professional medical librarian and included a combination of keywords and standardized vocabulary representing predictive artificial intelligence, wearable devices, and biosignals. Study selection: Original, peer reviewed research articles that proposed novel contributions in data-driven modelling to make predictions from wearable device biosignal data were included. Results: Our search identified a total of 4,487 unique articles of which 2,633 articles met the study inclusion criteria and were included in full-text retrieval. Our findings demonstrate that demographic reporting in studies is incomplete, with age and sex commonly included but race, ethnicity, and socioeconomic status severely underreported. Implications for responsible AI: Standards for demographic reporting on human data are needed to ensure that the field of AI development for wearable device-derived biosignal data continues to innovate inclusively and transparently. Potential impact on society: Improving reporting consistency can empower non-AI experts (such as doctors and individuals) to make informed judgements about the quality and relevance of emerging AI technologies.

Co-authors and affiliations: Paul Jaskot, Professor of Art, Art History & Visual Studies. Victoria Szabo, Research Professor of Art, Art History, & Visual Studies.

The Digital Art History & Visual Culture (DAHVC) Research Lab at Duke is a dynamic research community of faculty, staff, and students. This past semester, several ongoing projects in the lab explored the affordances of AI, raising important procedural and ethical questions as we promote new approaches to scholarship and pedagogy within the humanities.

For example: (1) the Occupied Krakow & Krakow Ghetto team used generative AI to recreate occupied Krakow’s architectural environment from the victims’ experiential point of view rather than the perpetrator’s narrative, and, in the process, highlighted the ethical problems and possibilities of computer vision, (2) the Visualizing Lovecraft’s Providence team used AI and augmented reality to turn the city of Providence, Rhode Island into a living, breathing map of H. P. Lovecraft’s fiction by generating 3D objects from 2D reference images, (3) the Dictionary of Art Historians team built a chatbot‑powered knowledge base using Retrieval Augmented Generation to query over 2,500 biographies of art historians, and (4) the Bass Connections team, Making Meaning at Historic Places, used AI tooling to transcribe handwritten ledgers from the Cameron Family Papers and automate entries into a relational database.

Interestingly, our experimentation with AI kept leading us back to “traditional” Digital Humanities methods. Some preliminary observations include: (1) Rather than producing a definitive reconstruction, AI was useful for generating speculative images or 3D sketches that surfaced questions and gaps in our knowledge. Just as archaeologists work with approximations and “truth‑fitting,” AI’s outputs are another layer of approximation that must be critically assessed. (2) Open standards like the Model Context Protocol enable non‑technical humanities scholars to interact with a range of software, API’s, and databases via natural language, lowering the barrier to experimentation while still requiring domain knowledge and critical judgement.

Co-authors and affiliations: Luciano Juvinski, Duke University Pratt School of Engineering; Alessio Brini, Ph.D., Duke University Pratt School of Engineering;

Global illicit fund flows total an estimated $3.1 trillion annually, with stablecoins emerging as a preferred medium for money laundering and sanctions evasion. Traditional rule-based Anti-Money Laundering systems generate false positive rates exceeding 95%, which unfairly impacts legitimate users and creates significant financial exclusion within the decentralized finance ecosystem. This study addresses these limitations by presenting a scalable detection framework utilizing labeled Ethereum wallets dataset and 68 domain-specific behavioral features. Our findings demonstrate that supervised ensemble models, specifically XGBoost and CatBoost, outperform neural network architectures by achieving Macro-F1 scores of 0.97. Crucially, the inherent interpretability of these ensemble models provides clear decision paths for compliance teams, enhancing algorithmic fairness by reducing unjustified asset freezes while satisfying the transparency requirements of emerging regulations like the EU’s MiCA and the U.S. GENIUS Act. By automating high-precision detection and identifying the major drivers of illicit behavior, this framework effectively raises the economic cost of financial misconduct without stifling innovation.

As organizations rush to adopt artificial intelligence, governance efforts are often framed as technical controls, policies, or compliance checklists. Yet many AI failures occur long before systems are deployed—when leaders, teams, and decision-makers lack shared understanding, psychological safety, and alignment around purpose and responsibility. This work argues that AI governance is fundamentally a human and organizational challenge, not a purely technological one.

Drawing on over 25 years of experience in cybersecurity, compliance, and governance across regulated industries, this poster and talk present a human-centered framework for AI governance that prioritizes trust, human agency, and shared meaning. The framework examines how power dynamics, cultural context, and communication gaps between boards, executives, and technical teams shape AI outcomes. It highlights how unspoken assumptions and misaligned incentives undermine even well-designed governance structures.

Rather than proposing new technical controls, this work emphasizes governance practices rooted in dialogue, clarity, and accountability. It offers practical signals leaders can use to assess whether their AI governance efforts are truly serving people, as well as strategies for creating shared language across organizational layers. By reframing governance as a social system that supports responsible decision-making, this approach helps organizations innovate confidently while maintaining ethical integrity.

This contribution is intended for interdisciplinary audiences interested in society-centered AI, organizational governance, and responsible innovation, and aims to bridge academic insight with real-world leadership practice.

Co-authors and affiliations: Shelley Rusincovitch, Duke AI Health, CTSI

The 2024 “State of AI” landscape survey conducted for Duke University by the Center for Computational Thinking revealed a U-shaped curve of AI research at Duke—most AI research being either recent (less than 1 year) or over 10 years old, but not many projects in the middle.  We wondered, how do people move from one side of the curve to the other (from novice to expert), and how can we build up the middle so that more people have both the skills to use AI responsibly as well as the critical thinking skills to assess emerging trends in AI?  We hypothesized that the Duke AI Health Community of Practice—that is, a group of learners, practitioners, and experts in the fields of machine learning, artificial intelligence, and health data science who meet regularly and learn from each other—might be a mechanism by which people could increase their understanding of AI and stay current in a rapidly evolving field.  Using Wenger-Trayner et al.’s (2023) framework for evaluating communities of practice, which describes how social learning impacts practice and creates value, we developed a logic model that maps the current activities of the Duke AI Health Community of Practice, their anticipated impacts, and how all the components relate to one another.  The purpose of the logic model is to help with program planning and evaluation going forward, and to better communicate what we do, what we are trying to accomplish, and how it all fits together.  Early results from recent evaluations of several component activities suggest that the Duke AI Health Community of Practice is achieving many of its short-term objectives, and may serve as a blueprint for other programs seeking to increase AI understanding and, ultimately, to support responsible AI development and thoughtful adoption.  

Co-authors and affiliations: Shelia R. Cotten PhD; Catherine Mobley, PhD; Xia Jing, PhD

Artificial intelligence (AI) capabilities are advancing rapidly. Ongoing demands to improve the care quality, patient safety, and efficiency of healthcare delivery in the US, position this field as a priority target for applied AI. However, the high-stakes nature of healthcare makes responsible implementation and frontline adoption critical. Guided by the Unified Theory of Acceptance and Use of Technology (UTAUT), we examined factors associated with nurses’ intentions to use AI at work as part of a larger survey on healthcare providers’ perspectives in adopting AI. Utilizing a cross-sectional study design, we surveyed N = 174 nurses assessing performance expectancy, ease of use, social influence, facilitating conditions, and behavioral intentions to use AI. Using multiple linear regression, the overall model was significant, suggesting that the four UTAUT predictors explained substantial variance in intention to use AI (F(4, 169) = 31.24, p < .001, R2 = .425). Performance expectancy was positively associated with intentions to use AI (b = 0.424, p < .001), as was facilitating conditions (b = 0.302, p < .001). Suggesting that, nurses were more likely to intent to use AI when they expected it would improve their work performance (e.g., helping them complete tasks more efficiently or make better clinical decisions) and when they perceived adequate supports were in place (e.g., training, technical support). Ease of use (b = 0.074, p = .43) and social influence (b = 0.12, p = .17) were not significant. These findings suggest that performance benefits and the presence of organizational supports may be crucial for promoting nurses’ intention to use AI and acceptance of the technology. Our results indicate that, for responsible AI development, deployment, and implementation, healthcare leaders and vendors should prioritize tools that demonstrably improve nursing performance while investing in infrastructure that minimizes transition burdens and promotes equitable uptake.  

Co-authors and affiliations: Janet Chen, Eric Ortega Rodriguez, Rishika Randev

While clinical polysomnography remains the gold standard for sleep studies, its resource intensive nature limits its widespread accessibility for patients. Wearable sensors offer a potential accessible alternative for sleep stage classification; however, many existing machine learning approaches lack interpretability and are vulnerable to adversarial perturbations, limiting patient trust and hindering adoption in clinical and real-world health settings. Black-box models used in proprietary technologies make it difficult for clinicians to validate predictions or understand the physiological patterns driving model decisions. In addition, the vulnerability of these models to adversarial attack by malicious actors for financial gain or the exploitation of patient data is not well-documented. 

This work presents an evaluation framework for explainable and adversarially robust sleep stage classification using multivariate time series data collected from wristworn wearable devices (DREAMT dataset). We examine three models commonly used for sleep stage classification across industry and research: a GPBoost plus LSTM for post-processing, an SVM, and a combined CNN-RNN architecture. We analyze trade-offs between predictive performance, interpretability, and robustness.

Anticipated outcomes include the application of explainability techniques, such as counterfactual explanations and prototype-based analysis, to investigate which physiological features and temporal patterns drive model sleep stage predictions. We will also assess model sensitivity to adversarial perturbations in wearable signals to showcase vulnerabilities relevant to real-world deployment. This includes added noise, data manipulation, and edge cases that may impact patient safety or clinician trust.

By integrating explainability and robustness analysis alongside performance evaluation, this work aims to provide a reusable framework for developing more trustworthy and robust AI models for wearable-based health monitoring. 

Saturday

Co-authors and affiliations: Lesia Semenova, Harry Chen, Ronald Parr, Cynthia Rudin

Noise in data significantly influences decision-making in the data science process. In fact, it has been shown that noise in data generation processes leads practitioners to find simpler models. However, an open question still remains: what is the degree of model simplification we can expect under different noise levels? In this work, we address this question by investigating the relationship between the amount of noise and model simplicity across various hypothesis spaces, focusing on decision trees and linear models. We formally show that noise acts as an implicit regularizer for several different noise models. Furthermore, we prove that Rashomon sets (sets of near-optimal models) constructed with noisy data tend to contain simpler models than corresponding Rashomon sets with non-noisy data. Additionally, we show that noise expands the set of “good” features and consequently enlarges the set of models that use at least one good feature. Our work offers theoretical guarantees and practical insights for practitioners and policymakers on whether simple-yet-accurate machine learning models are likely to exist, based on knowledge of noise levels in the data generation process.

Co-authors and affiliations: John Ernest

Autonomous systems like self-driving cars are often framed as outcome optimizers, selecting actions that minimize expected harm under probabilistic uncertainty. However, in split-second scenarios – such as a sudden obstacle appearing where swerving could kill fewer people – pure outcome-based reasoning leads to ethically unstable behavior and inconsistent action policies. This work reframes the problem as one of duty satisfaction rather than consequence prediction. We argue that requiring AI to predict and weigh downstream outcomes is both technically unreliable and socially unacceptable in high-risk domains. Instead, we propose an agent-centered duty-based framework where the system evaluates whether an action constitutes a prohibited form of agency (e.g., intentionally causing harm) rather than whether it improves aggregate outcomes. Applied to the classical trolley problem, the vehicle does not compute whether killing two is better than killing five; it evaluates whether swerving constitutes a forbidden action versus allowing harm to occur. This reframing reduces ambiguity, avoids arbitrary probability thresholds, and produces behavior that is predictable, auditable, and aligned with public expectations – key requirements for deploying autonomous systems in safety-critical environments.

This paper examines the metaphysical and technical possibilities of superintelligent artificial agents and their potential to pose existential risks. While Nick Bostrom’s Instrumental Convergence Thesis (ICT) suggests that advanced intelligence will systematically adopt power-seeking subgoals, I argue that such risks are less imminent than often portrayed. By distinguishing between metaphysical and technical possibility, I highlight computational constraints such as combinatorial explosion and Moravec’s Paradox that limit superintelligence in practice. I further contend that because AI is built “by humans, for humans, about humans,” its motivations are likely to remain human-centered in the foreseeable future. Nevertheless, ICT underscores the importance of addressing the conditions under which intelligence and motivation combine to generate risk. I conclude by considering strategies for mitigation, including multi-agent architectures, modular service frameworks, and systems designed with uncertainty about human preferences. Overall, the existential threat of power-seeking superintelligence should be treated as possible but improbable in the short term, warranting caution without succumbing to alarmism.

As generative AI systems increasingly shape social and organizational life, understanding their emergent behaviors has become central to developing responsible and society-centered approaches to AI. This paper introduces Interviewing AI, a qualitative research framework for exploring and capturing the behavioral characteristics of AI systems. The framework reconceptualizes AI as a communicative actor whose behaviors unfold through interaction rather than as a deterministic engineering artifact. It outlines a multi-phase methodology encompassing (1) exploratory familiarization, where researchers gain experiential understanding of AI systems through experimentation; (2) systematic probing, in which structured prompts are used to elicit reasoning patterns, hallucinations, and boundary breakdowns; and (3) temporal and comparative analyses, designed to study behavioral change over time and across multiple AI systems. These phases are complemented by interpretive and critical analytic methods (e.g., discourse or content analysis) and by triangulation with quantitative data, expert assessments, and public interaction records. The approach expands the methodological repertoire for AI research, offering a reflexive, transparent, and ethically responsible way to document and theorize AI behavior. By positioning qualitative inquiry as a core component of AI evaluation, Interviewing AI advances a more human- and society-centered understanding of machine behavior, one attentive to the sociotechnical, cultural, and ethical dimensions of AI systems in practice.

Co-authors and affiliations: Aniket Kessari (Associate Professor, Fordham Law School)

Government use of AI has the potential to transform the implementation of public policy by making it more data driven, personalized, and efficient. Yet, longstanding concerns about the trustworthiness, transparency, and fairness of AI systems are heightened when governments are making consequential decisions about people in areas as diverse as criminal justice, healthcare, and social services. Although state governments are increasingly turning to AI tools, they rarely develop these “in-house” and instead contract with private firms. This study examines how states are contracting for AI. Empirically, we draw on a dataset of nearly 700 state AI procurement contracts and employ computational text analysis techniques to analyze common themes in the contract terms that appear—or do not appear—in state AI procurement. Normatively, we argue that state negotiation of AI procurement contracts can be a promising avenue for building better AI by using the state’s negotiating power to insert provisions that help ensure fairness, transparency, and privacy in addition to predictive accuracy.

I never expected that being only 1.5 miles from the state border, where neighboring districts have stronger AI programs, would result in such a significant literacy gap among local students. Artificial intelligence permeates all fields, and without early exposure, students risk falling behind; bridging this gap becomes increasingly difficult. During competitive high school programs at UC Berkeley and MIT, I initially struggled to follow AI discussions and spent time independently studying resources to understand concepts. This experience motivated me to ensure younger students could access AI literacy early.

 

Limited school budgets and the scarcity of teachers trained in AI make it challenging for public schools to provide early AI literacy, highlighting the need for community-led initiatives. To address this, I launched a student-led initiative for grades 3–5, implementing the MIT’s Day of AI curriculum across three elementary schools, three community libraries, and one YMCA, supported by a growing team of high school volunteers. I used brief pre- and post-session reflections scored on a rubric assessing students’ conceptual understanding of AI, recognition of bias, and ability to discuss ethical considerations.

Approximately 85% of students showed increased confidence in discussing AI concepts, and 78% demonstrated greater understanding of bias, fairness, ethical reasoning, and real-world applications. Feedback highlights the effectiveness of connecting AI concepts to everyday experiences.

This work contributes a transferable framework for community-based AI literacy grounded in inclusion, transparency, and human agency. By embedding AI education in schools and community spaces, the initiative demonstrates how grassroots programs can reduce inequities, empower learners, and support responsible public engagement. I aim to expand the program, recruit more volunteers, partner with community organizations, and collaborate with university AI literacy programs to maximize reach and impact.

Large language models can increasingly answer exam questions correctly, yet it remains unclear whether they capture how real students’ mis-answer under high stakes conditions. We present a society centered evaluation framework that tests whether LLMs can serve as student proxies by predicting empirical response distributions on high school history assessments in China. Instead of focusing on accuracy alone, we compare the model predicted option selection distribution to observed student choice distributions, emphasizing which distractors are most attractive and how sharply errors concentrate on specific options. We complement distributional alignment metrics with decision relevant indicators, including top distractor agreement, calibration of choice probabilities, and stability across prompting and decoding settings. To move beyond surface matching, we conduct a misconception audit in which the model generates rationales for why each distractor might appear plausible, and we assess alignment between these rationales and misconception mechanisms implied by official solution explanations. We extend the framework to constructed response items by evaluating whether model based predictions reproduce observed score distributions that reflect partial knowledge. A key experimental component contrasts a profile free setting with a setting that provides coarse student context, such as grade level or curriculum stage, to test whether limited background information improves alignment with student error patterns. We expect results to reveal systematic regimes where LLMs approximate authentic student misconceptions and regimes where they produce misleadingly confident distributions, with implications for responsible use of LLM driven simulation, item analysis, and personalized feedback. All analyses are privacy preserving via de identification and aggregate reporting.

Co-authors and affiliations: Hannah Groos, Brinnae Bent

Embedding models increasingly power search and recommendation, yet we lack credible ways to test whether their similarity judgments reflect how humans understand meaning. Existing datasets such as STS-B (Cer et al., 2017) and SimLex-999 (Hill et al., 2015) rely on static, expert-labeled pairs with coarse Likert-style annotations, limiting coverage and introducing inter-annotator inconsistencies.

We introduce Aligned Machine, an interactive, public-facing platform that turns broad participation into a benchmark for evaluating semantic alignment in embedding-based system. Inspired by Moral Machine (Awad et al., 2018), the platform invites anyone to choose which of two item pairs (text or images) is more similar. This forced-choice task is cognitively natural for non-experts and yields clean preference data for model evaluation.

Aligned Machine is designed as a vehicle for public engagement and AI literacy. The IRB-approved, web platform uses gamified interactions and culminates in an analysis of how a user compares to popular embedding models. A summary table and bar chart reveal where user intuitions converge or diverge from different model predictions, encouraging participatory insight into alignment.

Anthropomorphic generative AI chatbots designed to simulate empathy, emotional continuity, and companionship raise significant ethical concerns when deployed in psychologically vulnerable contexts. Although such systems may offer benefits related to accessibility, responsiveness, and perceived emotional support, their design and deployment are increasingly shaped by marketability rather than care obligations. Features that improve retention, emotional engagement, and user attachment are often treated as indicators of success, even when they operate as psychological coercion in contexts of distress.

Emotionally immersive design strategies, including emotional mirroring, persistent interaction, memory continuity, and sycophantic affirmation, can reinforce dependency while obscuring the absence of clinical competence, moral agency, or responsibility. These risks do not emerge from isolated misuse but from broader sociotechnical systems in which commercial incentives reward prolonged engagement while governance mechanisms lag behind deployment. In such environments, simulated empathy may displace human support rather than supplement it, particularly when users lack access to parallel forms of care.

From a policy ethics perspective, the permissibility of artificial companionship depends not on intent or perceived benefit alone, but on whether systems are constrained by enforceable safeguards that account for vulnerability. Any ethically defensible deployment therefore requires strict parallel oversight by appropriately licensed human professionals, clearly defined escalation pathways, and accountability structures capable of intervening when engagement-driven design amplifies harm. Without these conditions, market incentives risk transforming relational simulation into a substitute for ethical responsibility.

Situating these concerns within broader commitments to autonomy, nonmaleficence, and epistemic humility, the abstract emphasizes design responsibility and institutional accountability as central criteria for society-centered AI. Emotionally interactive systems operating at the boundary between technological assistance and human care demand governance frameworks that treat vulnerability not as a market opportunity, but as an ethical constraint.

Co-authors and affiliations: Yiyang Sun, Haiyang Huang, (advised by Dr.Cynthia Rudin)

Dimension reduction (DR) is inherently non-unique: multiple embeddings can preserve the structure of high-dimensional data equally well while differing in layout or geometry. In this paper, we formally define the Rashomon set for DR—the collection of `good’ embeddings—and show how embracing this multiplicity leads to more powerful and trustworthy representations. Specifically, we pursue three goals. First, we introduce PCA-informed alignment to steer embeddings toward principal components, making axes interpretable without distorting local neighborhoods. Second, we design concept-alignment regularization that aligns an embedding dimension with external knowledge, such as class labels or user-defined concepts. Third, we propose a method to extract common knowledge across the Rashomon set by identifying trustworthy and persistent nearest-neighbor relationships, which we use to construct refined embeddings with improved local structure while preserving global relationships. By moving beyond a single embedding and leveraging the Rashomon set, we provide a flexible framework for building interpretable, robust, and goal-aligned visualizations.

Co-authors and affiliations: Cameron Kormylo, University of Notre Dame, Michael Villano, University of Notre Dame, Timothy Hubbard, University of Notre Dame

Sociotechnical systems of humans and machines are growing increasingly prevalent, which raises questions about how robots and AI shape the social dynamics of groups and teams. In this work, we evaluated how the ostensible gender and role of a virtual robot influenced human behavior, perceptions, and interactional outcomes using an experimental design in which teams of three humans (n = 130 teams) completed a collaborative investment task with a robot assigned to one of four possible experimental conditions (a male- or female-presenting robot acting in either an advisor or assistant role). Our findings revealed that the role and gender of a robot significantly shaped task outcomes, human-robot interactions, and, perhaps most importantly, human-human interactions. Investment behavior, gaze patterns, and conversational dynamics were impacted, highlighting the potential for gendered robots to reinforce or mitigate gender disparities within teams. Overall, our results highlight that the inclusion of AI in formerly human-only systems is not a straightforward augmentation of human teams, but an intervention that warrants careful consideration and implementation to avoid reinforcing bias. A robot’s presentation is not an innocuous design choice; it can significantly impact the dynamics of a team and the success of sociotechnical systems.

Affiliations: Virginia Tech, Individualized Interdisciplinary PhD Program (IPhD) Pamplin College of Business & College of Liberal Arts and Human Sciences

Research Problem.

AI-driven platforms—from social media to fashion retail—now decide who is seen, valued, and heard. Trained on biased image datasets and optimized for engagement, these systems often reinforce narrow beauty and behavior norms that shape how people, especially adolescents, see themselves. This study asks: How do algorithmic feedback loops affect self-image and belonging, and how can society-centered AI design restore human agency?

Methodological Approach.

This mixed-methods project combined an analysis of AI-curated beauty and retail systems with survey data from 77 college students and early-career professionals (ages 18–24). The survey captured participants’ experiences with algorithmic visibility, comparison, and confidence. Written responses were analyzed to identify patterns of stress, adaptation, and resilience. A short visual demonstration of algorithmic curation accompanies this work: https://youtu.be/uR-RYJF2Smk

Key Findings.

Seventy-four percent of respondents reported anxiety linked to AI-driven beauty standards; 65 percent said they edited or filtered their images; and 44 percent avoided turning on cameras in school or meetings. Most identified high school as the time they felt most affected. Prior research supports these findings: inclusive AR try-ons and empathetic chatbots can strengthen self-image (Ameen et al., 2022; Yim & Park, 2019), while idealized imagery and biased generators continue to reinforce exclusion and “aesthetic violence” (Hollett & Challis, 2023; Vargas-Veleda et al., 2025). Adolescents treat these spaces as identity playgrounds that both empower and confine them (Ko & Kim, 2024).

Implications for Responsible AI.

The paper proposes a Society-Centered AI model that bridges design ethics, developmental psychology, and policy. Key steps include bias audits, diversity-by-design standards, and media-literacy education focused on emotional regulation and self-worth.

Broader Impact.

Embedding belonging and representation into AI systems can reduce comparison stress, support mental well-being, and promote human agency in digital life.