Seed Funding

• Yuqing Hu, College of Engineering
• Rui Zhang, College of Engineering

Abstract: Disasters cause significant human, economic, and infrastructural losses, disrupting not only essential systems but also the behavioral and spatial routines that sustain urban life, such as how people move, access resources, and make protective decisions. Understanding these human dynamics is essential for emergency management and urban resilience, as individuals’ responses depend on both what they are doing and where they are when disruptions occur. Existing approaches capture only fragments of this complexity: surveys and interviews reveal motivations but are retrospective and event-specific, while simulation models, though predictive, rely on fixed behavioral rules that cannot represent adaptive reasoning under unprecedented or compounding hazards. To fill these gaps, this project develops AI-Augmented Urban Resilience Agents (AURA), a new framework that integrates reasoning-capable large language models (LLMs) with empirical behavioral data and psychological theory to simulate how people perceive, reason, and adapt during disasters. By combining data-driven and model-based situational realism with theory-guided behavioral coherence, AURA achieves both realism and generalizability, enabling interpretable simulations that extend across diverse disaster types and urban environments, thereby advancing next-generation urban emergency planning, risk communication, and resilience assessment. Specifically, three interconnected tasks will be conducted. Task 1 builds empirically grounded behavioral baselines using American Time Use Survey (ATUS) data to model typical daily activity transitions by demographic group. Demographic-specific Markov models describe probabilistic daily routines such as working, commuting, and shopping. These patterns are incorporated into each LLM agent’s state memory and prompt conditioning to constrain its reasoning within realistic temporal behaviors. A model-based decoding process further ensures that the LLM generated activities remain coherent with the statistical structure of the behavioral model. Task 2 embeds these ATUS-informed agents within real urban environments through activity–location mapping using OpenStreetMap (OSM) and SafeGraph Points of Interest (POI) data. Together with demographic distributions from the American Community Survey (ACS), this spatial embedding provides agents with realistic accessibility constraints, neighborhood heterogeneity, and infrastructure interdependencies that shape feasible decision pathways. Task 3 introduces a Theory-Driven Normal-to-Disaster Transition Module, integrating the Extended Protection Motivation Theory (E-PMT) and the Protective Action Decision Model (PADM) into each agent’s reasoning cycle to generate psychologically consistent adaptations to evolving risks such as heat, power loss, or infrastructure failure, conditioned on both the activity context (Task 1) and spatial context (Task 2). The framework will be validated using the 2025 Philadelphia heatwave–blackout, assessing both behavioral realism and predictive performance through meteorological, outage, mobility, and survey datasets that have been collected or secured through established data-sharing agreement. The interdisciplinary team brings complementary expertise in urban resilience modeling, human mobility analysis, and generative AI system design, ensuring both technical rigor and social relevance. This project aligns with the mission of the Center for Socially Responsible Artificial Intelligence (CSRAI) at Penn State by advancing human-centered, empirically grounded, and accountable AI systems that model how individuals interact with built environment during crises. It also lays the foundation for competitive external funding through programs such as NSF’s Infrastructure Systems and People (ISP) and Cyber-Physical Systems and Connected Communities (CPS), positioning AURA as a scalable, socially responsible framework for AI-enabled urban resilience research.

• Kelley Cotter, College of Information Sciences and Technology
• Andrea Miller, Donald P. Bellisario College of Communications and College of the Liberal Arts
• Sara Liao, Donald P. Bellisario College of Communications and College of the Liberal Arts
• Cora Butcher-Spellman, Donald P. Bellisario College of Communications and College of the Liberal Arts
• Benji Davis, College of Information Sciences and Technology
• Hil Malatino, College of the Liberal Arts
• Lili Dudas, College of Information Sciences and Technology
• Salah Seoudi, University Libraries

Abstract: Trans and gender non-conforming (TGN) individuals currently face increased scrutiny, surveillance, and discrimination within the United States. Amid heightened states of precarity, some trans people and their families are turning to crowdfunding platforms to raise funds for relocation to safer regions in the US and internationally. To escape hostile hypervisibility offline, TGN crowdfunders must seek visibility online, rendering them subject to the dominant gaze and surveillance of platform systems, governance, and audiences. The proposed research will investigate the affordances, vulnerabilities, and ambivalences of TGN crowdfunding campaigns across online platforms, as life-sustaining TGN care practices collide with state and industry uses of artificial intelligence supplementing existing surveillance practices and systems. Deploying a mixed-methods approach of semi-structured interviews and discourse analysis, the project takes up the dilemmas that attend embracing, resisting, and ignoring conditions of hypervisibility through commodification and datafication processes that necessarily accompany the surveillance of TGN crowdfunders. We ask: How do TGN individuals navigate the tensions and ambiguities of in/visibility in relocation-driven crowdfunding campaigns? Addressing this question will contribute novel insight into modalities of survival and liberation for TGN communities and subjects during times of increasing AI-enabled corporate, state, and lateral surveillance enmeshing offline and online environments. The proposed research will culminate in the preparation of an external grant proposal to expand this foundation via external funding opportunities that support interdisciplinary, justice-oriented projects benefitting marginalized and underserved communities.

• Christopher Daryl Cameron, College of the Liberal Arts

Abstract: In this proposal, I seek funding for an interdisciplinary conference on Expanding Empathic AI, to follow up on a successful event I hosted here through the Consortium on Moral Decision-Making in spring 2024. This event will bring together scholars from philosophy, the social sciences (e.g., psychology, political science), and engineering and computer science to discuss descriptive and normative considerations about extending empathy to, and receiving “empathy” from, artificially intelligent interaction partners such as LLM's and robots. The aim will be to use these conversations to refine our understanding of scientific and ethical constructs of empathy in the space of human-AI interactions, and to highlight the value of cross-disciplinary conversation and collaboration. The conference will comprise talks, interdisciplinary panels, early career researcher flash talks, and break-out workgroup discussions meant to encourage new collaborations, and the formation of a “research roadmap” working paper to unite a vision for moving forward in the broader study of empathic AI. Expected outcomes will include the submission of this working paper to an interdisciplinary journal, as well as the formation of a team that will submit for an integrative grant on this topic. The event will be hybrid, in-person and live-streamed online, with recording, production, and dissemination online afterwards. The Empathic AI 2026 event will dovetail with the Rock Ethics Institute's 2025-2026 theme on ethics and human technology, and the College of Liberal Arts' 2025-2026 theme on fragmentations.

• Aron Laszka, College of Information Sciences and Technology
• Yiqi Zhang, College of Engineering

Abstract: Public transportation services provide immense social and economic benefits to communities by enabling community members to access employment, education, and other opportunities. Due to the growing complexity of transportation, public transit agencies in the U.S. increasingly rely on automated tools that augment human decision-making with artificial intelligence (AI) for planning and operational decisions. However, existing AI tools are guided by purely technical objectives, such as maximizing ridership or minimizing vehicle miles traveled. Although logical, these objectives often lead to solutions that are counterintuitive, difficult to manage, and misaligned with the nuanced values of a community, which hinders the adoption of AI and negates its benefits. To address this translational gap, we propose to create the first human-centered AI for the planning and operation of transit services. A crucial element of our innovation is learning the preferences of community members and transit workers through engagement and machine learning, and incorporating these learned preferences into state-of-the-art AI tools. The proposed seed project will demonstrate, with limited scale and scope, that we can align AI for transportation with human preferences, which will serve as a foundation to develop competitive proposals for external funding.

• Cynthia Huang-Pollock, College of the Liberal Arts
• Roger Beaty, College of the Liberal Arts
• Kenneth Huang, College of Information Sciences and Technology

Abstract: Severe temper outbursts, argumentative, and aggressive behaviors in children and adolescents are among the most common reasons for referral to mental health services in children under the age of 14. The most widely adopted evidence-based therapies are parent-based; all include the teaching of relationship-enhancing and behavioral management skills, with “homework” between sessions assigned to increase the rate of skill development. However, homework completion rates are low and live practice with feedback from a therapist is limited by well-known non-trivial barriers such as transportation, scheduling, and childcare for non-target children. This creates a critical implementation gap where the uptake of skills is inefficient, and therapists lack scalable tools to support families between sessions. We therefore propose to develop a natural language system to be used in conjunction with enrollment in parent-based psychotherapy. This platform would act as a secure, scalable practice space for parents, solving the "homework problem" by providing immediate, iterative feedback that is essential for learning complex skills. This proposal aligns with the center's mission to advance interdisciplinary, human-centered AI research for social good, to address a critical need in mental healthcare.

• Michael Clarke, College of Health and Human Development
• Syed Billah, College of Information Sciences and Technology
• Gloria Soto, College of Health and Human Development

Abstract: Despite the availability of speech-generating technologies (SGTs), many non-speaking children with communication disabilities struggle to reach their language learning potential. These barriers stem from the cognitive and physical demands of constructing messages on SGTs and from limited adult proficiency in providing developmentally appropriate spoken language models. Building on our own prior research with non-speaking children using SGTs and established theories of language acquisition, our aim is to address this pervasive issue by developing innovative AI-enhanced SGTs that support non-speaking children and their communication partners in interactions that facilitative child language learning. This pilot project will set the foundations for this work by addressing two important aims to: (1) develop and evaluate child and adult language models that reflect child level of language and conversational context; and (2) design and test of prototype interfaces that unobtrusively deliver language scaffolding prompts to children and their communication partners. This work advances ethical, inclusive AI design and aligns with CSRAI strategic goals to promote equity, inclusion, and human-centered technology.

• Qunhua Li, Eberly College of Science
• Yuehong (Cassandra) Tai, College of the Liberal Arts

Abstract: Transparent, representative, and accountable governance is fundamental to public trust and social justice. Yet analyzing how elected officials communicate with constituents across diverse platforms and contexts poses major methodological challenges. Annotation—a cornerstone of social science research on text and multimedia data—is traditionally slow, costly, and difficult to scale. Large language models (LLMs) offer unprecedented efficiency but introduce new risks: shallow reasoning, misalignment with codebooks, and inconsistent outputs that undermine transparency, reproducibility, and trust in AI-assisted research. This project, led by two PIs with complementary expertise in computational social science and AI for reproducibility, addresses these challenges through two human-centered GenAI frameworks. Aim 1 develops an annotator–inspector dual-agent framework to align LLM reasoning with social science codebooks, producing transparent and reproducible labels of policy domains and stance. Aim 2 introduces a multi-agent, human-in-the-loop system to automate and accelerate codebook refinement, reducing the time and error of traditional adjudication while enabling multimodal annotation. These innovations will be applied to two high-value datasets—both collected and curated by our team: the Digitally Accountable Public Representation (DAPR) Database, containing over 6 million social media posts from 28,000 U.S. officials, and the Local Public Meetings on AI and Governance (LPMAG) dataset, comprising 4,000+ multimodal recordings from 40 states. Applying our frameworks to these data will enable reproducible, large-scale analysis of how public officials communicate, deliberate, and represent their constituents across digital and physical forums, yielding new insights into democratic accountability, participation, and governance. Together, the proposed methods and applications will advance the methodological foundations of responsible, scalable AI in social science, generate open-source tools and datasets for studying democratic communication, and provide hands-on training for graduate and undergraduate researchers—advancing Penn State’s strategic goals in interdisciplinary research and student success.

• Alison Gernand, College of Health and Human Development
• James Wang, College of Information Sciences and Technology

Abstract: A placenta is absolutely vital in a pregnancy and directs the health of both the mother and developing fetus. Standard pathology assessment of placentas, to glean important data for clinical decision making, takes too many resources and too much time to complete for all births. Low-resource settings suffer the most with the lowest access to expert pathology for placental assessment. We are designing a socially responsible artificial intelligence (AI)-based software tool – PlacentaVision – to assess and triage all placentas via digital photograph. Additionally, we plan to test performance metrics by race and ethnicity of the mothers. Our vision is a future where every birth, regardless of hospital setting and resources, will have the placenta assessed and in turn, better clinical care. Our interdisciplinary work will contribute to the Penn State strategic plan themes of enhancing health and empowering through digital innovation and impact socially responsible AI by contributing to social good and equity. The work will be led by faculty in Health and Human Development and Information Sciences and Technology. We will publish our results in top-tier data science/AI journals and submit an NIH R01 renewal application, with these results serving as preliminary data. This seed funding will allow us to move toward creating a socially responsible AI-based tool to assess all placentas, triage those needed to pathology, and ensure equity across racial and ethnic groups to meaningfully contribute to social good.

• Christine Kirchhoff, College of Engineering
• Sarah Rajtmajer, College of Information Sciences and Technology

Abstract: This proposal seeks to develop the Water Adaptation Synthesis Portal (WASP), an AI-driven system to enhance adaptation tracking in California’s water sector. Leveraging small, open language models like Llama and techniques such as few-shot learning, WASP will streamline adaptation tracking by automating evidence extraction with a human-in-the-loop approach. This system aims to address the growing need for timely and scalable adaptation tracking, reducing researcher time while delivering real-time insights into climate adaptation progress. The final output is a fine-tuned model for supporting sub-national adaptation tracking in the water sector as well as a protocol for the use of AI in evidence synthesis.

• Jim Casey, College of the Liberal Arts
• Christopher Dancy, College of Engineering
• P. Gabrielle Foreman, College of the Liberal Arts

Abstract: This CSRAI grant proposes to bring together scholars and students from the Colleges of Liberal Arts, Engineering, Information Sciences and Technology, and Earth and Mineral Sciences in an interdisciplinary, cross-rank, initiative that will advance knowledge of socially responsible AI in Black archives. The proposal includes PSU scholars and students from African American Studies, English, History, Geography, IST, and Mechanical Engineering including Assistant professor Jim Casey, Associate Professor Chris Dancy, Paterno Family Professor, P. Gabrielle Foreman, and PhD candidates Eden Mekonen, Courtney Murray, and Deja Workman. The project will be housed in PSU’s Center for Black Digital Research, which itself is a cross-College (Library/CLA) Center which houses the longest-standing Black digital humanities project in the U.S. as well as Douglass Day, a global annual transcribe-a-thon which has engaged over 33,000 transcribers on four continents in transcribing Black records since 2017, resulting in an ongoing partnership with the Library of Congress. The proof of concept and planning work we will do via this grant will help secure future NEH and NSF funding and will also prepare PSU students to take leadership roles at the intersection of ethical/socially responsible AI and archives, addressing an urgent need identified by our partners at the Library of Congress, Smithsonian, and citizen science organizations.

• Matthew Beckman, Eberly College of Science
• Rebecca Passonneau, College of Engineering
• ChanMin Kim, College of Education
• Dennis Pearl, Eberly College of Science

Abstract: The responsible development of tools that leverage pre-trained large-language models for educational purposes must address the impact of non-normative language use. The proposed study collects a large sample of data from students who use the English language differently from their U.S. peers, and engages faculty from the same cultural and linguistic background, to investigate NLP tool development that prioritizes students’ represented understanding and mutes the influence of superficial details attributable to linguistic diversity such as idiomatic expressions, unconventional grammar and vocabulary, and spelling. The CSRAI project will utilize NLP tools for formative assessment–which include two relational networks for assessment, and two human-in-the-loop methods whose novelty is to machine-learn when to defer the assessment to a human, due to low decision confidence from the automated assessment methods. These tools will be used to evaluate performance for a novel, substantial, data set introducing compelling linguistic diversity to analyze the impact of non-normative language use. The interdisciplinary team from Statistics, Computer Science, and Education is well-positioned for a highly competitive NSF proposal for either the Research on Innovative Technologies for Enhanced Learning (RITEL) or Improving Undergraduate STEM Education (IUSE: EDU) programs.

• Max Crowley, College of Health and Human Development
• Jonathan Wright, College of Health and Human Development
• Alexander Winters, College of Health and Human Development

Abstract: The Penn State University’s Evidence-to-Impact Collaborative seeks a planning grant to integrate AI infrastructure into the research-to-policy pipeline. The planning process will focus on building scalable AI infrastructure that supports researchers in the creation of actionable scientific discovery and policymakers’ use of scientific evidence to make data-driven decisions while mitigating accuracy and misinformation risks. New AI and ML methods require substantial overhead in data architecture as well as human expertise to facilitate lines of questioning, hypothesis testing, and general work of researchers that require AI/ML methods. This planning grant will start building the bridge between researchers and new AI methods that enable cutting edge research.

• Kenneth Huang, College of Information Sciences and Technology
• Dongwon Lee, College of Information Sciences and Technology

Abstract: Logical fallacies are reasoning errors that undermine the validity of arguments. They degrade online discussions, making arguments seem more dubious and contributing to the spread of misinformation. This proposal aims to combat misinformation and foster a healthier online discourse by detecting logical fallacies in online narratives, particularly in social media posts. A key challenge in logical fallacy detection is the low inter-annotator agreement (IAA), as even expert annotators often disagree on judgments. To address this, we will (i) develop an annotation scheme that accommodates multiple interpretations of logical fallacies within a single text instance, (ii) leverage this scheme to achieve a higher IAA for social media misinformation posts, and (iii) apply fallacy detection models to our proposed framework. Ultimately, this work aims to enable early interventions for misinformation by identifying when individuals begin to formulate fallacious arguments, which can lead to disseminating widespread misinformation.

• Stephen Hampton, Penn State Harrisburg
• Anthony Buccitelli, Penn State Harrisburg
• Bernice Hausman, College of Medicine
• Jennifer McCormic, College of Medicine
• Vida Abedi, College of Medicine
• Michael McShane, College of Medicine

Abstract: Artificial Intelligence (AI) has the potential to revolutionize healthcare by enhancing diagnostics, personalized medicine, and operational efficiency. While its benefits may be manifold, AI’s adoption in rural healthcare faces unique challenges. Rural populations, often characterized by limited healthcare access, present distinct barriers to AI adoption, including skepticism, privacy concerns, and algorithm aversion. These challenges may be influenced by socioeconomic status (SES), population density, and exposure to technology. This study investigates the perceptions of AI-mediated healthcare among rural patients and examines factors affecting trust and compliance with AI-generated medical recommendations. Drawing on insights from consumer behavior and services literature, the study explores how source credibility, transparency, and customization impact trust and compliance. Algorithm aversion may also pose a significant obstacle, particularly in healthcare decisions. Using a mixed-methods approach, the research includes semi-structured interviews to describe AI perceptions in rural populations. Key objectives include understanding the role of SES and population density in shaping attitudes toward AI and identifying strategies to build trust through transparent and personalized AI systems. The findings aim to inform public policy and AI governance frameworks, ensuring equitable healthcare access for underserved rural communities.

• Amulya Yadav, College of Information Sciences and Technology
• Rebecca Passonneau, College of Engineering
• Ritu Jayakar, College of the Liberal Arts

This project is jointly funded by the Institute of Computational and Data Sciences.

Abstract: Language is a profound tool for communication and interaction, yet the vast linguistic diversity across the globe poses significant challenges for emerging Natural Language Processing (NLP) technologies. Over the past one year, Large Language Models (LLMs) such as OpenAI ChatGPT and Google Bard have shown remarkable promise in understanding and generating human-like text. However, preliminary research done by the PI suggests that their potential seems skewed towards monolingual English-speaking users, leaving a large portion of the global population—those engaging in code-mixed language usage—at a comparative disadvantage. Code-mixing, the fluid alternation between languages within a conversation or text, is a common linguistic phenomenon, especially in multilingual societies in South Asia (e.g., India) and East Africa (e.g., Kenya). This proposal aims to answer the following research question: How well can state-of-the-art LLMs handle code-mixed data, and how does their performance compare to Natural Language Processing (NLP) methods that are explicitly designed to handle code-mixed text?

• Bernice Hausman, College of Medicine
• Michael Green, College of Medicine
• Priscilla Song, College of Medicine
• Jennifer McCormick, College of Medicine
• Anthony Buccitelli, Penn State Harrisburg

Abstract: We propose to address socially responsible ways of using, building, and deploying AI technology in healthcare by approaching AI from the perspective of the humanities and the arts, emphasizing AI as a broad sociocultural phenomenon. This effort will establish a research network, AiHA (Artificial Intelligence in the Humanities and Arts, pronounced ah ha!), that fosters connections between scholars on all Penn State’s campuses, builds a knowledge base, and develops innovative interpretive approaches to the AI healthcare imaginaries currently in play.

• Qingyun Wu, College of Information Sciences and Technology
• Xiaolong Zhang, College of Information Sciences and Technology
• Ellen Wenting Zou, College of Education

Abstract: Tutoring has been found effective in helping students struggling academically or otherwise, and new technological innovations have been used to address various challenges in tutoring (e.g., matching, scheduling). Recent breakthroughs in Large Language Models (LLMs) offer new ways to support tutoring services. While LLM-based tools like ChatGPT and Bing Chat have emerged as powerful tools in language understanding, coding, reasoning, etc., the potential of LLMs in educational contexts, especially in STEM education, are yet to be fully realized. There is a gap in understanding how these models can be effectively integrated into learning tools to provide meaningful and tailored support to students, helping them navigate the complexities of problem-solving in STEM. As a pilot study, this research proposes MathMentor, an AI-based math tutoring system empowered by large language models. Drawing insights from cognitive theories on problem-solving and theories on STEM education, MathMentor aims to leverage the language and reasoning capabilities of LLMs to provide step-by-step guidance in problem-solving, identify and address common misconceptions (blind points), and offer personalized explanations of mathematical concepts in real-time. Through this, the system intends to enhance the problem-solving skills and conceptual understanding of college students, making mathematics more accessible and less daunting. We will also conduct a thorough evaluation of the system to gain deeper insights into the efficacy of LLMs in aiding students with math concept learning and problem-solving.

• Dongwon Lee, College of Information Sciences and Technology
• ChanMin Kim, College of Education

Abstract: This collaborative project between Penn State (IST & Education) and University of Mississippi (Computer Science) proposes to investigate linguistic discrimination issues of LLMs and their impacts on the auto-assessment of the level of English learners. We hypothesize that automated writing assessments generated by LLMs favor standard English writing while neglecting other English dialects spoken by African American or Hispanic students in the US or by population in other English-speaking countries. Therefore, there is an urgent need to understand linguistic discrimination and limitations of LLMs in assessing student writings in English. With the support from this seed grant, therefore, we propose to design a framework that can systematically demonstrate and detect potential linguistic discrimination in LLM-generated assessments of student writings in English, and obtain preliminary results from experiments that will lead to larger proposals for external funding.

• Priya Kumar, College of Information Sciences and Technology
• Laura Cabrera, College of Engineering, College of Liberal Arts
• Kelley Cotter, College of Information Sciences and Technology

Abstract: Students are using generative AI to brainstorm ideas, better understand concepts, proofread their texts, and, in some cases, to complete assignments. Like others, students express concerns about generative AI providing inaccurate, fabricated, or unoriginal information and how use of generative AI will affect schooling and professional work (Bodnick, 2023; Svlurga & Natanson, 2023; Terry, 2023). At the same time, by “messing around” with generative AI systems, students are living the process of understanding the opportunities and the limits of these technologies (Leaver & Srdarov, 2023). The intellectual and experiential knowledge they are gaining through this process should be incorporated into discussions and actions around the ethics of generative AI in education. Yet current recommendations to address generative AI-driven academic

integrity issues focus on one-way communication, encouraging instructors to learn about the academic integrity concerns, update their academic integrity policies, tell students about the concerns, and monitor student work for violations (Debby et al., 2023; Eke, 2023). In contrast, what could happen if instructors and administrators approach the arrival of generative AI as an opportunity to engage with students on deeper questions about learning (Kumar et al., under review; Warner, 2023)? Instructors across the humanities, social sciences, and STEM disciplines have developed lessons and assignments using generative AI systems that have spurred critical thinking among both students and instructors (Arena, 2023; Laquintano et al., 2023; McMurtrie, 2023b). Building on this momentum, we aim to explore how students are making sense of the ethics of generative AI.

Two research questions guide our pilot study:

1. How are undergraduate students using generative AI systems?
2. What factors influence undergraduate students’ judgment about what constitutes ethical use of generative AI systems?

• Yiqi Zhang, College of Engineering (Department of Industrial and Manufacturing Engineering)
• Aiping Xiong, College of Information Sciences and Technology

Abstract: Artificial intelligence systems in automated driving systems (ADS) are known to be vulnerable to adversarial attacks, making ADS susceptible to safety- and security-critical errors that pose significant road hazards and fatalities. Recent research exploring human drivers' understanding of physical-world attacks indicates a lack of awareness of AI vulnerability, which may pose significant risks to driver safety. This project aims to establish a conceptual framework for driver-AI collaboration in response to adversarial attacks by addressing driver situation awareness. To this end, we will carry out behavioral studies to (1) examine the impact of AI system capabilities and reliability on driver situation awareness, trust, and takeover performance across various adversarial attack scenarios; and (2) investigate central/distributed human-machine interface design to promote driver situation awareness restoration and effective human-AI collaboration. The proposed research will advance socially responsible AI in transportation by deepening our understanding of driver situation awareness of adversarial attacks to enhance AI system resilience. The project's findings could guide Level 3 automated vehicle design optimization, promoting emerging AV technologies and enhancing driver safety. This project's outcomes will be used to obtain external research funding and build a long-term, multidisciplinary research program at Penn State in the area of human-AI collaboration. Additionally, this research will contribute to the Larson Transportation Institute's national leadership in testing automated vehicles.

• John M. Carroll, College of Information Sciences and Technology
• ChanMin Kim, College of Education (Department of Learning and Performance Systems)
• He Zhang, College of Information Sciences and Technology

Abstract: The impact of generative AI technologies, such as large-scale language models (LLMs) and related applications (e.g., ChatGPT), on the education industry, is emerging. Determining the long-term trend of this impact is challenging, and potential risks could have severe consequences. It has become urgent to study how to maximize the benefits of AI tools like ChatGPT while minimizing risks (Zhai, 2022; Perkins, 2023; Dwivedi et al., 2023). This pilot study aims to understand the effects of students' interaction with LLMs-based applications on their summarization and learning abilities.

Summarization is a crucial skill in academic and professional settings and a primary capability of generative AI technology like GPT. It involves condensing large volumes of information into concise and coherent summaries, allowing for better understanding, retention, and communication of essential points. In academia, summarization helps students process and remember key concepts from lectures, textbooks, and research papers, further developing critical thinking skills. Professors and researchers can also use summaries to share their findings succinctly, making their work more accessible and understandable to a broader audience.

Knowledge workers in various industries benefit from summarization as well. Summarization enables them to process and share information quickly and effectively, which is essential in today's fast-paced and information-rich environment. Developing and promoting summarization skills during education is crucial. LLMs' technologies, like GPT, have been designed to perform tasks such as summarization by leveraging their vast knowledge and natural language processing capabilities.

LLMs-based AI systems can process large amounts of text quickly, generating concise summaries by identifying key points and reasoning far faster than humans, increasing productivity and enabling users to focus on higher-level tasks such as analysis, strategy, and innovation.

The intersection of summarization as a key academic and knowledge worker skill with AI technologies like GPT presents opportunities for collaboration and improvement. However, the use of LLMs' applications for summarization could undermine human summarization and create ethical problems for students and teachers. Our proposal aims to understand the relationship between LLMs application and the ability to summarize and learn in educational settings, addressing potential risks and challenges.

The research employs a mixed-methods approach, combining quantitative and qualitative data to comprehensively assess the impact of LLMs-based applications on undergraduate students' learning and summarization abilities. We will conduct several pilot studies, including interviews, questionnaires, focus group discussions, and expert workshops to address our research questions.

The research questions are: RQ1) How does the interaction with LLMs-based applications influence students' summarization and learning ability in comparison to traditional learning methods? RQ2) How do learning behaviors, practices, attitudes, and expectations of students change when interacting with LLMs-based applications compared to traditional learning environments? RQ3) How do students' attitudes and perceptions towards LLMs-based applications impact their learning experience and understanding of course content? RQ4) What are the potential risks and challenges associated with the use of LLMs-based applications in the education sector, and how can they be mitigated? What strategies and interventions can be implemented to ensure generative AI tools like ChatGPT are used ethically and responsibly in the education sector?

The anticipated outcomes include insights into the effectiveness of LLMs-based applications in promoting learning and summarization skills, as well as identifying factors contributing to the success or challenges of using these applications. The research aims to develop risk mitigation strategies to address potential negative impacts on student skills and academic performance while offering recommendations for the ethical and responsible use of generative AI tools in education.

The findings will shed light on student attitudes, learning behaviors, and psychological profiles in relation to summarization skills, promoting LLMs-based applications as an effective tool in Human-Computer Interaction (HCI), education, computer science, and social science. The results will inform policy decisions and improve educational practices, creating a more inclusive and equitable environment. Additionally, the findings will guide future large-scale studies in different contexts, such as across disciplines, stages (e.g., K-12, graduate school, lifelong learning), scenarios (workplace), populations (older adults, people with disabilities), and cultures. Through the research related to this seed proposal, we hope to provide strong support for human-computer interaction and social justice in the context of the rapid development of LLMs. Expected research results may be published in prestigious journals and international conferences such as CHI, Teaching and Teacher Education, CSCW, IEEE Trans, ACII, etc., to extend the impact of the research project further, thereby increasing the likelihood of external funding support.

• Donghyun Rim, College of Engineering (Department of Architectural Engineering)
• Dongwon Lee, College of Information Sciences and Technology

Abstract: While the United Nation’s sustainable goals focus on human health, well-being, and quality education of world population, nationwide closures of elementary and secondary schools due to the novel coronavirus disease (COVID-19) have severed nearly 60 million students from critical educational and health resource. The objective of this proposal is to establish a research and teaching framework for AI-based infection risk analysis for creating clean and healthy school environments. Results from this study will be widely applicable and relevant to the design, operation, and maintenance of densely occupied learning environments as well as social policies to control infectious diseases transmission in school buildings.

• Syed Billah, College of Information Sciences and Technology
• Krista M. Wilkinson, College of Health and Human Development (Department of Communication Sciences and Disorders)
• Dawn Sowers, College of Health and Human Development (Department of Communication Sciences and Disorders)

Background: As noted in the NJC Communication Bill of Rights (Brady et al., 2016, p.1), "All people with a disability of any extent or severity have a basic right to affect, through communication, the conditions of their existence." Access to speech-language therapies that promote optimal communication outcomes is also recognized as a fundamental right by the United Nations' Article 19 of the Convention on the Rights of Persons with Disabilities (2016). Yet many individuals with physical or intellectual disabilities have language limitations that prevent them from using speech as their primary mode of communication. For these individuals, assistive communication technologies (called augmentative and alternative communication, or AAC) offer an important set of supports. The late physicist Stephen Hawking is an example of someone who lost his speaking voice due to an acquired motor neuron disease, Amyotrophic Lateral Sclerosis (ALS), but was able to use synthesized speech from an AAC device. Without these assistive communication technologies, individuals like Dr. Hawking and over 5 million Americans with developmental or acquired communication disabilities could not have access to their fundamental human right to communication.

Problem: AAC technologies are applicable across a broad range of diagnostic conditions and can particularly benefit individuals with motor disabilities and concomitant cortical visual impairment (CVI) who struggle with ocular or cortical processing of visual information. However, current AAC systems are suboptimal for those with CVI and/or motor impairment due to reliance on the visual representation of vocabulary items (letters, words) and the crowded display of symbols that overwhelm visual processing for individuals with CVI. Additionally, access to these systems typically requires finger-pointing or limb movement, which often exceed the motor abilities of individuals with motor disabilities who rely on slower input devices like switches or eye gaze. Repetitive selections in these systems are challenging and fatiguing. For example, individuals with ALS who use eye gaze for typing communicate at a mere 10 to 20 words per minute (wpm), compared to a spoken speech rate of 180 wpm. Consequently, there is an urgent need to develop AAC technologies that minimize the visual and motor burdens for individuals facing challenges in vision and motor behavior.

Specific Aims: The overarching goal of this project is to transform access to AAC technologies through the development of human-centered AI algorithms and intelligent, interactive systems that are capable of sensing and interpreting specific communication messages in various forms of access that do not require advanced vision or repetitive motor movements. We specifically focus on touchscreen gestures and/or natural gestures, and aim to develop personalized, high-performance, and lightweight AI algorithms based on sensor fusion and few-shot learning, suitable for running on edge devices that will enable both AAC users and their care providers to easily create, configure, and tailor communication through a user-friendly, intelligent interface. As a result, the proposed AAC system will shift the burden of access from the users to the algorithms, interfaces, and sensors integrated into AAC devices. This transition would diversify access methods beyond repetitive movements, such as finger-pointing or limb movement, towards rapid, predictive interactions, and include a broader range of message preparation techniques, incorporating alternative touchscreen gestures and natural gestures like facial expressions, eye gaze, or head movements.

• Prasenjit Mitra, College of Information Sciences and Technology
• Bing Pan, College of Health and Human Development (Department of Recreation, Park and Tourism Management)
• Peter Newman, College of Health and Human Development (Department of Recreation, Park and Tourism Management)
• B. Derrick Taff, College of Health and Human Development (Department of Recreation, Park and Tourism Management)
• Shreya Ghosh, College of Information Sciences and Technology

Abstract: In an era defined by climate change and rapid urbanization, the interplay between human activities and wildlife has intensified, frequently instigating conflicts and disturbing the equilibrium of natural ecosystems. The effects of sound and light pollution, which are integral to these interactions, are frequently understated, despite their potential to disrupt wildlife. This project, therefore, proposes the deployment of artificial intelligence (AI) and machine learning (ML) techniques to discern and quantify the consequences of human activities, including movement, sound and light pollution, on wildlife behavior and their habitat utilization. The research is motivated by a commitment to social good and a profound respect for all life forms. By developing AI solutions to monitor and mitigate the impact of human activities on wildlife, we aim to foster a more equitable coexistence between humans and other species, thereby advancing social justice and ethical considerations in technology development. Our work will promote transparency by revealing the often hidden costs of human activities on wildlife and habitats. It will also empower individuals and communities to make more informed decisions about their behavior and its potential environmental consequences, fostering greater social consciousness and accountability. By ensuring the privacy of individual data during the implementation of our AI models, we intend to uphold the principles of data ethics and privacy. Additionally, the project will be designed with inclusivity in mind, considering the impact of sound and light pollution on a broad range of species and habitats. Finally, the project will involve diverse stakeholders, including local communities, policy makers, and conservationists, in the process of developing and implementing AI solutions. By doing so, we aim to ensure that our research not only contributes to scientific understanding, but also meaningfully advances the cause of social good, justice, equity, and sustainability.

Research Questions:

• R1: What patterns can we discern in large-scale human activity data that relate to the frequency and intensity of human-wildlife interactions and conflicts?
• R2: How can the impact of different types of sound and light pollution on wildlife behavior and visitor experiences in protected areas be quantified and predicted?
• R3:What are the optimal strategies for managing soundscapes and lightscapes in protected areas to enhance both wildlife conservation and visitor satisfaction?
• R4: How can adaptive management strategies for protected areas be designed considering the dynamic nature of human activities and wildlife responses and effectively communicated to visitors?

• Jue Wang, Smeal College of Business
• Daniel Mallinson, Penn State Harrisburg
• Eric Best, Penn State Harrisburg
• Pedro Robles, Penn State Lehigh Valley
• Lauren Azevedo, Penn State Harrisburg

Abstract: Digitalization has brought up the importance of fair data practices. Most existing studies have examined data practices at the technological level, which failed to consider the heterogeneity of organizations that implement the practices. This project aims to investigate data practices in organizations and develop stakeholder governance theory to understand how firms design their data practices according to the stakeholders of the data. As these data practices can be fair or unfair according to the stakeholders, governments may design policies to ensure the spread of fair data practices. We further explore how governments design their policies. We will start by examining data practices of about 3,000 AI-based ventures and related government policies from 2002 and 2018.

• Amulya Yadav, College of Information Sciences and Technology
• Samuel Schleipman, World Hope International
• Andy Gatto, Nittany AI Alliance

Abstract: The Philippines is the global epicenter of Online Sexual Abuse and Exploitation of Children (OSAEC) – UNICEF reports that 8 out of 10 children in the Philippines are vulnerable to being victims of online sexual abuse and bullying. Almost ~2.5% of children in the Philippines have had their nude bodies or sexual activities shown on the internet or a cellphone. Unfortunately, the situation has significantly worsened during the Covid-19 pandemic – the number of reported OSAEC cases in the Philippines between March and May 2020 increased by 264% as compared to the same period in 2019. Disturbingly, more than 80% of these OSAEC cases were perpetrated on children by their own parents or relatives. These statistics highlight the critical need to urgently tackle this societal challenge.

Since 2017, World Hope International – Philippines (WHI-PH) has focused on combating OSAEC and protecting survivors. WHI-PH is one of the very few organizations in the world that focuses on the after-care of survivors of OSAEC. In particular, WHI-PH approaches after-care in two ways: (i) training Local Government Units (LGUs), social workers, and shelter managers on the OSAEC specific skills needed to provide targeted care for survivors; and (ii) ensuring the safe and effective reintegration of OSAEC survivors back into a stable family environment, for which WHI-PH either focuses on resolving underlying dysfunctions in the family (e.g., in many cases, destitute poverty and lack of work opportunities causes family members to force their children into sex acts online), or in some cases, WHI-PH may also try to identify financially and emotionally stable relatives of OSAEC survivors who can provide safe and welcoming family environments to survivors.

We aim to tackle the following research question in partnership with World Hope International -Philippines (WHI-PH): Can we build an AI-driven decision support system which can automate the preliminary triage and case assessments of OSAEC survivors?

• Kevin Munger, College of the Liberal Arts
• Sarah Rajtmajer, College of Information Sciences and Technology

Abstract: The production of knowledge by AI systems is fundamentally constrained by the interpretability of that knowledge --- that is, the ability for humans to understand the output of an AI system. There has been significant attention paid to the question of the “interpretability” of machine learning models; our proposed research is related but distinct: how can social scientists best incorporate AI technologies into our workflow of scientific knowledge production?

• Dongwon Lee, College of Information Sciences and Technology
• Amulya Yadav, College of Information Sciences and Technology
• J. Andrew Petersen, Smeal College of Business
• Guangqing Chi, College of Agricultural Sciences

Abstract: This collaborative project between Penn State and two NGOs in Texas—Texas Housers (TH) and Child Poverty Action Lab (CPAL)—proposes a data-driven approach to develop a novel framework to forecast the number of tenants at-risk of formal evictions (i.e., the number of eviction filings) in high resolution (e.g., census tract level, instead of County level) across the USA. Knowing the accurate number of eviction filings in advance helps policymakers and NGO workers make well-informed eviction mitigation plans and direct resources more efficiently. However, at present, the lack of historical eviction filing data in high resolution from courts prevents ML models from accurately forecasting the number of tenants at-risk of formal evictions in high resolution, posing a great intellectual challenge. To address this challenge, we seek to validate our idea of developing deep learning models that leverage publicly available data (e.g., satellite imagery), social media and e-commerce data (e.g., Twitter, AirBnB), historical court filings in low resolution as well as sociological insights (instead of ground truth labels in high resolution). This CSRAI seed grant will use Texas area as a test case to produce sufficient preliminary results so that we can calibrate/extend our solutions toward the entire US in the external grant submissions. Our collaborators at two NGOs will provide relevant datasets and feedback to our solutions as domain experts, and work together for external grant submissions.

• Anna Squicciarini, College of Information Sciences and Technology
• Paul Esker, College of Agricultural Sciences
• Primal Pappachan, College of Information Sciences and Technology
• Mladen Cukak, College of Agricultural Sciences

Abstract: To ensure that farmer data is safely collected and stored on our digital platform, we are proposing this project which would help establish a basis for future collaboration between the Esker and Squicciarini groups in the sphere of data privacy and security in the crop production domain. Through this project, we plan to achieve the following high-level objectives:

1. Preparation: Perform literature research and engage with relevant stakeholders to identify the security and privacy challenges of the agricultural domain
2. Development: Design the framework architecture and develop the necessary privacy and security modules for offering different kinds of data protection.
3. Deployment and Testing: Run simulations to test the framework under various settings and deploy the applications that are part of the framework in the field.

• Christine Kirchhoff, College of Engineering
• Shomir Wilson, College of Information Sciences and Technology

Abstract: Drinking water systems are vulnerable to disruption from extreme events including hurricanes, heavy rain and wind, drought, wildfire, and extreme cold (Nakamura and Brooks 2021; Proctor et al. 2020; Waldrop 2022), and from poor management, delayed maintenance, and age. These disruptions often result in drinking water advisories which require residents to boil their tap water before use, to not consume their tap water, or to not use their tap water (Public Notice Rule 141.202). Disruptions in essential drinking water services can cause ripple effects, with considerable social (Schulman et al. 2004, Pescaroli and Alexander 2015), health (Wade et al. 20, Quist et al. 2022) and economic consequences (Alzahrani and Collins 2022; Reich et al. 2017). Past research on drinking water advisories has predominantly focused on the quality of the notice (Wardlaw and Bruvold 1989; Harding and Anadu 2000; O’Shay et al. 2020), the cause (Hrudey et al. 2003; Hrudey et al. 2006; Kargiannis et al. 2009; Rimhanen-Finne et al., 2010; Robertson et al. 2009; Wallis et al. 2001), and on whether the public is aware and adheres to the recommendations (i.e., to boil water)(Jones-Bitton et al. 2016; Hyllestad et al. 2019; Kargiannis et al. 2009; O’Donnell et al. 2000; Ram et al. 2007; Rundblad et al. 2010; Vedachalam et al. 2016).

To date there has been no investigation that we know of to address the research question, What is the scope and scale of drinking water advisories in the United States (U.S.)? Moreover, anecdotal evidence suggests that drinking water disruptions may disproportionately affect poor and historically marginalized communities (Sneath 2021; Waldrop 2022) but this has not been investigated systematically. We seek to begin to fill this gap using public notices of drinking water advisories in print and broadcast media across the U.S. to characterize the frequency, cause including from extreme events, geographic location, water systems, and characteristics of populations affected by drinking water advisories over time. We will do this by developing a first-of-its-kind dataset of drinking water advisories across the U.S. The computational steps of this research will be (1) using web search and web crawling methods to gather textual artifacts (e.g., announcements, news articles) about drinking water advisories from official sources; (2) using semi-supervised machine learning methods to bin the articles into sets that represent discrete events; and (3) using information extraction and text classification techniques to label each artifact with key metadata, including dates and type of advisories, locations, responsible authority, and causes. This pilot research will lead to an NSF proposal for a more comprehensive collection and analysis of drinking water service disruptions.

• Rebecca J. Passonneau, College of Engineering
• Brian Belland, College of Education

Abstract: Meta-analysis applied to research on classroom interventions informs educational policy from the district to the national levels, and supports teachers and instructors in informed decision making about instructional methods for diverse student groups. However, the high cost in human effort and elapsed time to perform accurate meta-analysis works against the potential benefits at multiple decision points, including which educational studies to fund based on community needs, as well as the speed at which proven methods could transform educational practice. Current manual methods for meta-analysis produce effect size estimates that are quickly outdated, and which often do not reflect research on underrepresented populations. The proposed effort supports a pilot study on the feasibility of natural language processing (NLP) methods applied to meta-analysis, to test the ability to automatically replicate human coding from a recently completed meta-analysis. The PI, whose research area is NLP, and the co-PI, with numerous published meta-analyses, have an existing collaboration that they seek to leverage in the proposed pilot project. Positive feedback on a previous submission to the NSF Directorate of Education and Human Resources indicated that pilot results would greatly increase the likelihood that a resubmission would lead to an award. The pilot study would address how to create the requisite data, given human coding sheets and corresponding full-texts of journal articles, and how to leverage machine learning methods for high accuracy of automated methods. The pilot study variables pertain to education population and other learner characteristics related to diversity. Outcomes would have significant impact on Penn State’s two strategic themes of Transforming Education, and Empowering through Digital Innovation, as well as on theoretical questions in NLP pertaining to semantics and context, and on the explicit and implicit factors that affect human and automated methods for meta-analysis coding variables.

• Qiushi Chen, College of Engineering
• Sunday Azagba, Ross and Carol Nese College of Nursing
• Guodong Liu, College of Medicine
• Paul Griffin, College of Engineering

Abstract: The rapid increase of e-cigarette use in youths and young adults has become a major public health concern in the United States. Although the legislators have endeavored to implement a wide range of e-cigarette regulations, youth e-cigarette use continues to rise, and the most effective e-cigarette policies remain yet unknown. Determining effective e-cigarette control policies can be challenging, because it relies on a clear understanding of the transition process of tobacco product use patterns in youths and young adults, and needs to account for complex interactive effects of multiple control measures entangled with the complex transition process of e-cigarette/cigarette use. To tackle these challenges, we are applying for the CSRAI Collaboration Initiation Funding (CIF) to initiate a multidisciplinary research collaboration, with a long-term goal of developing a data-driven modeling framework to inform the policy making of e-cigarette regulations. In this project, we aim to (1) assemble a research team of faculties with ranging seniority and complementary expertise in health analytics, dynamic modeling, and tobacco and e-cigarette policies across Colleges of Engineering, Nursing, and Medicine, (2) leverage a large national longitudinal survey dataset to examine the dynamic transition patterns of tobacco products use using the latent transition analysis, and (3) develop the e-cigarette policy simulation model prototype to explore possible synergistic effects of various control measures. This collaborative pilot study will allow the team to share domain knowledge and technical expertise needed for e-cigarette regulatory science research, to build a track record of collaboration and joint publication, and to generate key preliminary data demonstrating the feasibility of the proposed analytical framework based on real-world data and its potential translational values, which together will support the proposal of the next-step full-scale research program to attract external funding from NIH or tobacco regulatory agencies.

• Kelley Cotter, College of Information Sciences and Technology
• Priya Kumar, College of Information Sciences and Technology

Abstract: AI-driven systems promise to make our lives easier and make businesses more efficient and profitable. Yet, the quest to realize these promises often obscures concerns about their potential to violate privacy, exacerbate inequities, impinge on human rights and freedoms, and wreak environmental havoc. Moreover, slippage between the possibilities of what such systems could do and what they actually can do help propel cultural and economic investments in AI in spite of these concerns. Visions of technologies as magical, in particular, shift attention away from their potential harmful impacts to the bright future they might bring. The purpose of this project is to explore how discourses about the “magical” qualities of AI-driven systems can obscure—and even legitimize—the exploitative nature of such technologies. We pursue this aim through a multi-phase case study of Disney’s MyMagic+ system, one that mirrors the goals and operational logics of Big Tech platforms like Facebook and Google, but has not faced the same level of scrutiny. The questions guiding this pilot study are:

• How does Disney use the language of magic to describe the MyMagic+ system?
• In what ways does Disney’s use of magical discoursess address, detract, and/or obscure the potential concerns posed by the MyMagic+ system?

• Ting-Hao (Kenneth) Huang, College of Information Sciences and Technology
• Roger Beaty, College of the Liberal Arts

Abstract: Creativity is now the most distinguishing and valuable attribute of the workforce in the global economy. To effectively prepare the workforce, there is a critical need for educational tools that can help to identify what works (or does not) to foster creativity in education. A key barrier to equity in 21st century education is inclusivity of diverse students. The tools teachers currently use to foster creativity were not designed for culturally and linguistically diverse individuals—putting underrepresented minority (URM) students at a disadvantage. Likewise, as AI is increasingly used to support student learning, it is essential that such technology is equitable and not racially unbiased (i.e., algorithmically fair), particularly in high-stakes situations involving student evaluation. Promoting culturally responsive and algorithmically fair assessments—educational tools that measure student potential fairly across diverse cultural contexts—is crucial for equity of opportunity in 21st century education.

Our goal is to empower educators with the technology and knowledge to foster creativity in students from diverse cultural and linguistic backgrounds. Our proposal brings together the complementary expertise of two CSRAI faculty who study creativity from different perspectives: Dr. Ting-Hao ‘Kenneth’ Huang (IST) and Dr. Roger Beaty (Psychology). Seed funding will provide critical pilot data to support a grant application to the NSF (Directorate for Education and Human Resources). Our seed project has two specific aims: 1) to design an AI-Powered tool to support effective scientific writing with URM students; and 2) to develop a natural language processing (NLP) tool to evaluate creative thinking in science education—using best-practices in NLP debiasing to promote algorithmic fairness for URM students.

• Xianbiao (XB) Hu, College of Engineering
• Vikash Gayah, College of Engineering
• Ilgin Guler, College of Engineering
• David Norloff, Donald P. Bellisario College of Communications
• Yiqi Zhang, College of Engineering
• Min Ding, Smeal College of Business
• Sandra Allain, College of Engineering and School of International Affairs
• Bin Li, College of Engineering

Abstract: Autonomous vehicles represent a complex topic that necessitates in-depth analysis across multiple academic disciplines. As some claim that autonomous vehicles represent the moonshot for AI and machine learning, the development and testing of autonomous vehicles, however, go way beyond the technical aspect of algorithm development and training. In this proposal, we propose to initiate the dialogue of socially responsible AI for autonomous vehicles, by working closely with the autonomous vehicle consortium that is currently active at Penn State. Questions to study include, but are not limited to AV decision making in situations of moral dilemmas, AI bias in self-driving, Human bullying AV, and AV adaption to heterogeneous driving scenarios. We will discuss and identify the potential issues of AI in AV, foster cross-disciplinary collaborations, and brainstorm research ideas to pursue external funding opportunities. Faculty from multiple departments will work together to promote, practice and study socially responsible ways of using, building, and deploying AI technology for autonomous vehicles.

• Syed Billah, College of Information Sciences and Technology
• Prasenjit Mitra, College of Information Sciences and Technology
• Marcela Borge, College of Education

Abstract: Access to higher education correlates to the growing income inequality in the US. For example, students pursuing four-year of college in STEM or related fields have a higher income potential than those with lower levels of education or those who pursue in non-STEM fields (e.g., median income: \$77,400 in STEM work vs. \$46,900 non-STEM work per year), according to Pew Research (2019). However, the increasing cost of college education in STEM or related fields in the US is prohibitive to many students in low-income households; thus, making the benefit of higher education inequitable. This proposal aims to make higher education equitable by addressing the wicked problems in Massive Open Online Courses (MOOCs). These online courses have emerged as a low-cost alternative to higher education accessible to students worldwide, including those in low-income households. Drawing on the recent developments in artificial intelligence (AI), education, and human-computer interaction (HCI) research, we first design and develop an interactive, personalized, human-faced, conversation AI agent that will accompany students in MOOCs platforms to enable self-regulated learning practice.

We will simulate a virtual environment with personalized ambiance, where AI agents and humans can collaborate to create a new, desirable learning experience for MOOCs learners. Our work will create a broader societal impact by assisting learners across the US and the world.

• Daryl Cameron, College of the Liberal Arts and Rock Ethics Institute
• Martina Orlandi, Rock Ethics Institute and Schreyer Honors College

Abstract: The project aims at gathering collaborators from psychology, engineering, and philosophy with the objective of leveraging insights within existing conversations to develop a deeper analysis of the ethical implications of moral outsourcing. In particular, the project aims at extending the current debate by exploring how relying on algorithms can blur attributions of blame and how delegating moral labor to algorithms erases a fundamental aspect of what it means to be ethical by outsourcing moral decisions that instead require active reflection.

• Rebecca Napolitano, College of Engineering
• Wesley Reinhart, College of Earth and Mineral Sciences

Abstract: One of the National Academy of Engineering Grand Challenges is restoring and improving urban infrastructure, which currently scores a D+ on the American Society of Civil Engineer’s report card. One prominent barrier to restoring and improving buildings is that the state of materials inside the building envelope is unknown. Radar methods such as Ground-Penetrating Radar (GPR) have been developed to detect embedded elements such as rebar and plumbing. However, these techniques can only provide qualitative information about inhomogeneities such as voids or moisture in the insulation. The proposed work will use a hybrid physics-based simulation and artificial intelligence approach to infer the geometry and material properties inside an existing building envelope based on scattered radar signals. In this seedling project, we plan to explicitly study how the contents of training data sets influence the accuracy of resulting diagnostics and efficacy of resulting interventions in different constructions archetypes. Laboratory experiments will provide a high-quality validation data set to test the simulation protocol and scattering inversion capability of a physics-informed neural network model. Effort will also be directed towards benchmarking the method to provide bounds on various figures of merit related to deployment in the field, making future grant proposals more competitive by measuring the technological and societal impact beyond academic model tuning.

• Hadi Hosseini, College of Information Sciences and Technology

• Dongwon Lee, College of Information Sciences and Technology
• Amulya Yadav, College of Information Sciences and Technology
• Hyung Joon Yoon, College of Education
• Hamoon Ekhtiari, FutureFit AI

Abstract: This collaborative project between Penn State and FutureFit AI1 proposes an interdisciplinary and data- driven approach to building a “GPS for Future of Work” with a focus on mapping career transition pathways designed for both employers and employees severely impacted by automation and AI. By employing methods from computer science, career development, workforce training, and human capital management and using FutureFit AI’s unique dataset on employment demand and talent supply (with coverage across 200+ countries and territories, 30,000 skills, and more than 80,000 job titles), our integrative approach develops a suite of AI solutions to help the employees-at-risk to transition into jobs that have more promising career prospects and to ensure that employers are ready for effective employee career development by innovating HR practices and introducing ethical standard operating procedures. Throughout this seed project, preparing for two NSF grant submissions in 2021 and 2022, we will prioritize key workplace issues in future of work, identify key stakeholders and project partners, streamline research agenda through in-depth discussion in the planning meetings, and perform a preliminary study to obtain a proof-of-concept.

• Amulya Yadav, College of Information Sciences and Technology
• Jonte Taylor, College of Education
• Fei Fang, Carnegie Mellon University

Abstract: Due to their socialization and communication difficulties, young children with autism spectrum disorder (ASD) face many challenges in traditional school environments. Intelligent tutoring systems (ITS) are a promising technology for supplementing instruction for children with ASD. However, previous work in the design of such systems often does not consider crucial cognitive features of students, such as working memory deficits and the zone of proximal development (ZPD), etc. Furthermore, these systems are frequently not tailored for use by children with ASD, potentially causing significant shortcomings in the effectiveness of their learning. Building on top of our prior work, we propose to develop a Reinforcement Learning based ITS for teaching children with ASD by: (i) collecting data from 100 K-12 students suffering from ASD using a tablet-based software application; (ii) using the collected data to learn a parametrized student model, which represents learning behavior of students with ASD; (iii) use this student model with deep RL techniques to design an adaptive tutoring policy for teaching emotion recognition to students with ASD; (iv) piloting our RL based ITS with 100 K-12 students in Pennsylvania to evaluate its effectiveness.

• Daryl Cameron, College of the Liberal Arts and Rock Ethics Institute
• Alan Wagner, College of Engineering and Rock Ethics Institute

Abstract: As the nature of human-technology interactions becomes increasingly complex, there is a strong need to better understand the role of moral emotions, moral intuitions, and moral judgments as people navigate these interactions. In the current seed grant proposal, Drs. Daryl Cameron (Psychology, Rock Ethics Institute) and Alan Wagner (Aerospace Engineering, Rock Ethics Institute) will leverage their joint expertise to explore the conceptual and empirical possibilities for advancing the science of morality at the human-technology intersection.