researchGrant
Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
Sign Up For Latest News
Does AI have the potential to redress disparities and improve health outcomes for communities historically marginalized and underserved? Yes. Does it also carry the risk of perpetuating discriminatory bias and worsening disparities often impacting already vulnerable populations? Yes. So how do we facilitate the former while preventing the latter? Our work takes a community-centered, multi-stakeholder approach to explore how diverse populations perceive the integration of AI into health care in an attempt to further enhance our understanding and capacity to ensure AI is optimized in a way that advances health equity. We want to understand how people from different ethnic and intersectional backgrounds perceive AI-enabled health care; what the barriers and enablers are to access and what the potential opportunities are for AI to improve population-specific health challenges. By engaging civil society, policymakers, care providers, researchers, and industry, we seek to facilitate collective action in building capability for stakeholders and communities to work together in identifying and developing tailored policy and practice that advances AI’s human impact for all and not just those most privileged in society.
Name | Role | School | Department |
Main PI | School of Medicine | Epidemiology and Population Health | |
Co-PI | School of Medicine | Primary Care and Population Health | |
Research Scholar | School of Medicine | Clinical Excellence Research Center (CERC) |
People with Autism Spectrum Disorder (ASD) face unique social communication challenges, particularly around nuanced aspects of communication that impact their ability to make friends, have romantic relationships, and secure employment. In fact, the ASD demographic faces the highest unemployment rate among any group with disabilities. While some face-to-face interventions with specialized clinicians target social conversation skills, there is a huge gap: The rise in ASD diagnoses, now 1 in 36 (CDC), far outpaces the availability of these highly-skilled clinicians. As a result, many with ASD have no access to traditional interventions. We introduce Noora, an AI-driven platform featuring conversation modules and LLM-based feedback. It is designed to help those with ASD improve their social communication through specialized exercises in challenging areas. We are starting an approved IRB randomized clinical trial to test Noora’s impact on verbal individuals with ASD. We aim to extend essential evidence-based intervention services to those lacking traditional clinical support and to bolster existing interventions with clinicians with this cost- and time-efficient addition.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Medicine | Psychiatry & Behavioral Sciences |
Our research aims to enable the next generation of multimodal situational apps by combining principles from Human-Computer Interaction (HCI), machine learning, and software engineering. The project introduces a new suite of developer libraries and tools, simplifying the creation of multimodal situation apps. These apps can understand user inputs based on physical surroundings, user profiles, and application data, allowing for more natural and intuitive interactions. To achieve this goal, our libraries and tools will assist developers in handling user actions and accommodating diverse user needs. By democratizing application development and promoting multimodal interaction, this research has the potential to bridge the gap between idea and action for users and enhance inclusivity, accessibility, and diversity in the app market.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science |
A prosecutor’s decision to charge an individual with a crime can have severe consequences for that individual, their family, and their community. Unfortunately, despite these high stakes, there is evidence that charging decisions are racially biased, with Black, Hispanic, and other marginalized race and ethnic groups bearing the costs of this discrimination. To mitigate bias in these charging decisions, we designed an algorithm that automatically redacts race-related information from crime reports, allowing prosecutors to make race-blind charging decisions. Legislators in California, prompted by our research, have mandated that prosecutors across the state implement a way to make all charging decisions race-blind by 2025. We will remove barriers to adoption so that our race-blind charging algorithm can scale to meet the state's mandate, and interest from prosecutors across the country. We plan to: (1) develop better scientific and technical solutions to ease adoption by new offices; (2) conduct research to study the impacts of blind charging, including its potential to increase perceptions of procedural justice; and (3) use this experience to inform the policy around its wider use.
Name | Role | School | Department |
Main PI | School of Law | School of Law |
Brain-controlled mobile robots have the potential to restore mobility and perform assistive tasks for those with sensorimotor impairments. Recently, intracortical brain-computer interfaces (iBCIs), in which neural signals are decoded from arrays of microelectrodes implanted into the motor cortex, have shown great promise over a variety of tasks, including computer cursor control, text messaging and communication via attempted handwriting. In this project, we propose to leverage advancements in iBCI as well as autonomous systems to fly a drone safely, as a demonstration of safe iBCI control of a dynamic physical system in real-time.
Name | Role | School | Department |
Main PI | School of Engineering | Aeronautics and Astronautics |
Clinical reasoning, a key skill for doctors to diagnose illnesses, is traditionally taught using clinical cases in formats like paper-based, online, and panel exams. These methods offer key information and questions for students but don't fully replicate the real-world scenario where doctors directly interact with patients, gather data, and make decisions. The most realistic training method available now is clinical simulation with actors portraying patients, but it's expensive and hard to standardize for large-scale use. To address scalability and affordability issues, researchers introduced the concept of virtual patients. Nonetheless, these apps lack AI capabilities and are limited to their initial design contexts, restricting their utility in other beneficial settings. In this project, our team will develop Clinical Mind AI, a novel application designed to harness generative AI for training medical students in clinical reasoning, specifically focusing on analyzing patients' medical histories. This app will enable medical educators to input illness scripts, which the application will then convert into simulated patient scenarios. This process will allow educators to provide clinical cases in a more interactive and realistic manner, closely mirroring real-world physician-patient interactions. Additionally, Clinical Mind AI will give students customized feedback aimed at improving their problem-solving practices and decision-making skills, enhancing the learning experience. We plan to conduct tests with medical students and expert physicians to evaluate the effectiveness of Clinical Mind AI, comparing it to traditional in-person clinical simulations. If our approach proves successful, it has the potential to revolutionize medical education globally, offering a more affordable and adaptable training solution.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education | |
Co-PI | School of Medicine | Pediatrics - Infectious Diseases |
We propose to build ConsistentControlNet, a novel architecture that can maintain frame-to-frame consistency (no flickering) for video, while using a pre-trained text-to-image diffusion model combined with ControlNet as the core component of the architecture. We will consider two approaches; (1) a training-free method that uses optical flow to enforce consistency in the denoising step of the diffusion process between corresponding pixels from frame to frame, and (2) an end-to-end method that combines the ControlNet architecture with a text-to-video diffusion model. In addition to generating flicker-free video we will apply the approach to generating high-quality textures for 3D assets.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science |
The overall goal of this project is to develop advanced algorithms that enable a team of agents to solve tasks such as assembling furniture, changing bed sheets, setting up a tent, or transporting bulky items. By leveraging collaborative manipulation techniques, a team of agents (either a multi-robot or human-robot team) can work together to perform these tasks with ease, offering several advantages over individual agents. Notably, a team of agents can achieve faster task completion times, improved robustness through data fusion, information sharing, and redundancy, and greater reliability, flexibility, scalability, and versatility. Prior work on collaborative robotic manipulation has shown impressive results on low-level sensorimotor control, coordination of multiple agents, sequencing of subtasks as well as on robustness to some uncertainty in the object parameters and geometry. However, high-level coordination of collaboration remains a challenge, with many approaches relying on hand-designed task plans, except in simple cases with fixed low-level actions. Despite the importance of high-level coordination, there is a lack of research on automating these aspects, especially in the area of multi-robot collaborative manipulation and when deployed in real world scenarios with noisy perception. This lack of solutions significantly hinders the scaling up of collaborative multi-robot or human-robot manipulation teams to a diverse set of tasks in real world scenarios. Our aim is to enable effective communication and collaboration among robots and human teammates. Using a large language model (LLM), we aim to facilitate task decomposition into subgoals, efficient task assignment to subteams or individual agents, and high-level team coordination. By leveraging LLMs, we eliminate the need for engineers to hand-design specific task domains, enabling robots to dynamically adapt and coordinate in complex environments.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Electrical Engineering |
Humans bear the primary responsibility for the ecosystem destruction. Yet we are also active agents that can make a significant change right now. In this process, what cultural ideas and practices can guide people to collaboratively participate in the natural environments? We describe two approaches to interacting with nature: dualistic thinking and ecological thinking. We propose that fostering the second approach through the use of human-centered AI has the potential to increase pro-environmental behaviors. To explore this important direction, we leverage cultural psychological sciences and AI to construct cultural agents for bridging different cultural perspectives on human-ecology interactions. Our hope is to eventually build an interdisciplinary field of inquiry to advance environmental justice and strengthen the public's capacity to coordinate on environmental actions.
Name | Role | School | Department |
Main PI | School of Humanities and Sciences | Psychology | |
Co-PI | School of Sustainability | Earth System Science |
Instruction-following language models (LMs) have powered recent and dramatic advances in natural language processing and artificial intelligence. These systems have shown a remarkable ability to generate humanlike text, perform classification tasks, and have open-ended conversations with users. If harnessed to benefit society at large, these systems could bring about tremendous benefits, accelerating scientific discovery and empowering human creativity. However, major questions remain about who these LMs serve. Ensuring these models are aligned with the communities they serve requires that the most critical aspect of these models – the data – be created and controlled by the community that these models serve. In this proposal, we articulate an alternative vision for LLMs, where the alignment of language models is done in a more democratic fashion through distributed data collection across many communities.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science |
Foundation models in AI have the potential to transform many areas of science. This project considers specialized foundation models for analyzing worker careers and labor market outcomes, models that are designed to perform well at specific downstream tasks. We propose training a model using large scale but non-representative data, trained to predict career transitions and other outcomes as a function of history. The model takes into account a predefined taxonomy over worker jobs as well as textual descriptions of the jobs and the skills associated with them. Our method produces low-dimensional representations of workers’ career history while avoiding various categories of bias. We particularly focus on biases that might arise when predicting differences across demographic groups in career transitions or in outcomes such as wages. We will develop and evaluate methods for fine-tuning the foundation model using representative survey data. The methods are designed to avoid “omitted variable bias' ' that arises when the representations do not capture elements of history that are important for future outcomes, and that are also systematically different across demographic groups. We will evaluate the extent to which fine-tuning to avoid omitted variable bias along multiple demographic dimensions hurts or helps performance in any given dimension. When omitted variable bias is avoided in a given dimension, the method produces representations that can be used to reliably decompose the sources of differences in, e.g. wages, across groups of workers, and can also be used to analyze changes in group differences over the course of worker careers. We further propose several downstream analyses, including analyzing gender gaps in the nature of text posted on resumes; we will further develop tools for workers to do career scenario planning, considering various constraints on how the planning tool accounts for systematic demographic differences in worker transitions.
Name | Role | School | Department |
Main PI | Graduate School of Business | Graduate School of Business |
Although large-scale patient health data like biobanks and EHRs have the potential to unlock novel clinical insights, their usage is currently limited due to privacy concerns and dataset bias that often mis- or underrepresents certain groups. To address these challenges, we propose an approach that generates synthetic patient data in a manner that simultaneously ensures privacy and representational fairness. Our proposed generative method will be well-suited for EHR and biobank data which is high-dimensional, sparse, and contains multiple variable types. Furthermore, we will develop a method that is both private, such as by learning a differentially private generator, while also optimizing for fairness across minority subgroups. We will evaluate our proposed approach by measuring the synthetic dataset's resemblance to the real dataset, utility on clinical classification tasks, privacy metrics, and fairness with respect to model mode collapse and group fairness. By evaluating UK Biobank and All of Us, two large-scale biobank data, we will ensure our approach generates synthetic data that is privacy-compliant and fairly represents a patient population, thereby democratizing access to informative health datasets and leading to novel medical insights.
Name | Role | School | Department |
Main PI | School of Engineering | Bioengineering | |
Co-PI | School of Engineering | Computer Science |
Globally and locally, governments and community partners often are committed to addressing community determinants of health (e.g., pollution, traffic safety, food access, healthcare access); yet many do not have readily available ways to systematically engage community members in identifying problems and forming locally relevant solutions. This research aims to amplify the positive effects of a tested citizen science community engagement method developed by the Stanford School of Medicine’s Health Equity Action Research and Technology Solutions (HEARTS) Lab, called Our Voice. Our goal is to enrich the information that community members gather and share using cutting-edge artificial intelligence (AI) tools that maintain the participatory nature of this work. We plan to augment the Our Voice method’s community ideation phase by leveraging text-to-image chat AI tools (e.g., Midjourney, DALL-E) to aid in generating quick visualizations of the local problems and potential solutions offered up by citizen scientists. Then, leveraging AI chat tools (e.g., ChatGPT) as an additional “pseudo-participant” within their community meetings, we seek to stimulate deeper and more provocative examination and debate among community members. Through these activities the project will address the following questions: Can equipping citizen scientists with generative AI tools democratize the ideas and solutions considered in the change-making process within their local communities? And will doing so ultimately lead to increases in health equity? We will systematically explore how such tools can help community members themselves, as well as the local decision makers who they seek to enlighten, better “imagine” and address the positive changes that need to be made in their local communities.
Name | Role | School | Department |
Main PI | School of Medicine | Epidemiology and Population Health |
Visualizations are essential for building data intuition but often do not cater to the abilities of those who are blind or have low vision (BLV). Many BLV students lag behind in skills to efficiently and accurately interpret graphical information. While haptic and audio are non-visual methods for representing data, differences in human perception across these modalities pose unique challenges for data exploration. For example, touch and audio convey information in a part-to-whole manner, in contrast to the whole-to-part nature of visual perception. This requires users to progressively form holistic representations through the sequential intake of perceptual information. Developing effective data exploration skills is non-trivial and requires significant training.
Currently, little is known about the decision-making process involved in data inference and exploration, especially in the realm of accessible visualizations. Our proposed work will develop a computational model to understand how BLV experts approach the data exploration process based on the perceptual cues that they gain. We then aim to investigate how such a model integrated into an intelligent tutoring system might provide targeted, context-specific, real-time interventions. These interventions aim to assist learners in developing more effective data exploration strategies when engaging in visualization tasks.
Name | Role | School | Department |
Main PI | School of Engineering | Mechanical Engineering | |
Co-PI | School of Humanities and Sciences | Music |
Algorithms promise the transformation of businesses by integrating greater and broader data within organizations (e.g., Iansiti and Lakhani, 2020) and by enabling reformed workflows to take advantage of this data (e.g., Agrawal et al., 2018). In March 2019, 86% of U.S. workers were employed by organizations founded ten or more years ago (Sadeghi, 2022) - organizations that, ostensibly, pre-date modern algorithms. To realize the full promise of algorithms, these organizations must reform existing workflows, rather than start from workflows native to algorithms. To smooth this transition both for workers and organizations, it is important to understand if, when, and how organizational structures and processes change as organizations adopt algorithms. The purpose of this study is to increase our knowledge of how organizations change as these organizations adopt algorithms. Most research to date focuses on individual-level impacts of algorithms. Academics, however, have theorized several effects of algorithm development and use on organizational structures and processes. To expand our understanding of if, when, and how organizational structures and processes change when organizations adopt algorithms, we propose participant observation of the development and deployment of an algorithm which has significant and disparate data dependencies and which has the potential to create new organizational interdependencies as a result of leveraging that data. Such an algorithm may change workflows at both the individual and organizational levels and afford real-time observation of changes resulting from algorithm development and adoption.
Name | Role | School | Department |
Main PI | School of Engineering | Management Science and Engineering | |
Co-PI | School of Engineering | Computer Science |
Algorithms promise the transformation of businesses by integrating greater and broader data within organizations (e.g., Iansiti and Lakhani, 2020) and by enabling reformed workflows to take advantage of this data (e.g., Agrawal et al., 2018). In March 2019, 86% of U.S. workers were employed by organizations founded ten or more years ago (Sadeghi, 2022) - organizations that, ostensibly, pre-date modern algorithms. To realize the full promise of algorithms, these organizations must reform existing workflows, rather than start from workflows native to algorithms. To smooth this transition both for workers and organizations, it is important to understand if, when, and how organizational structures and processes change as organizations adopt algorithms. The purpose of this study is to increase our knowledge of how organizations change as these organizations adopt algorithms. Most research to date focuses on individual-level impacts of algorithms. Academics, however, have theorized several effects of algorithm development and use on organizational structures and processes. To expand our understanding of if, when, and how organizational structures and processes change when organizations adopt algorithms, we propose participant observation of the development and deployment of an algorithm which has significant and disparate data dependencies and which has the potential to create new organizational interdependencies as a result of leveraging that data. Such an algorithm may change workflows at both the individual and organizational levels and afford real-time observation of changes resulting from algorithm development and adoption.
Name | Role | School | Department |
Main PI | School of Engineering | Management Science and Engineering | |
Co-PI | School of Engineering | Computer Science |
This project proposes an AI-driven system for designing ergonomic, functional, and creative objects. The central goal is to integrate generative AI and physics simulation to create objects that conform to individual user profiles (e.g., body shape and movement patterns) and specifications (e.g., style and manufacturing constraints). Integral to this process is a physical human simulation module, designed to evaluate each object’s ergonomic viability and functional integrity with respect to a particular user. Our approach seeks to generate designs that not only resonate with individual aesthetic preferences but also adhere to ergonomic and functional constraints. Ultimately, this research aims to transform the way objects are designed, making customized product design more accessible.
Twenty-eight million people are trapped in forced labor worldwide, and the supply chains for the construction and agriculture sectors are among those where the practice is most prevalent. This project seeks to assemble unified datasets that combine structured and unstructured data stemming from diverse sources including supply chain databases, governmental surveys, social media, or remote sensing tools, and use these to develop more robust and more equitable predictive models for assessing the risk of forced labor across distinct geographic and industry contexts.
Name | Role | School | Department |
Main PI | Graduate School of Business | Graduate School of Business | |
Co-PI | School of Engineering | Civil and Environmental Engineering | |
Co-PI | School of Sustainability | Center for Ocean Solutions |
Given the paucity of training data in healthcare settings, there is a need to both learn from multiple modalities as well as improve data efficiency to make clinical foundation models (FMs) useful. Therefore, we propose to integrate unstructured clinical notes and medical images, alongside with structured medical data in a FM to enhance clinical risk-stratification tasks and to address the critical challenge of limited data. We introduce a novel, similarity-guided mechanism to find clinically similar patients among those with analogous diagnoses. This mechanism is designed to significantly boost data efficiency by overcoming the key limitation of existing methods that ignore the subtle clinical similarities among the patients, which results in low sample efficiency during pre-training. By implementing such localized pre-training modifications, we can extract more information from limited medical data. We aim to evaluate the resulting model by collaborating with clinicians and Stanford Medicine to validate the model’s usefulness to improve patient outcomes.
Name | Role | School | Department |
Main PI | School of Medicine | Emergency Medicine | |
Co-PI | School of Medicine | Biomedical Informatics |
This study explores the degree to which the responses of generative models in applied interactive settings converge, and critically analyzes the implications of such convergence in case studies of educational and academic applications. For example, we measure the epistemic and linguistic diversity of synthetic output from early studies on the use of generative models in peer review. We situate the results within empirical and theoretical literature about the importance of epistemic variation between human reviewers to the robust and resilient construction of scientific knowledge. We hope our interdisciplinary collaboration will inspire more critical studies on generative tools in knowledge, education, and information ecosystems.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education |
Television (TV) and media hold tremendous power in society; representing people and telling stories in ways that reflect and shape our culture. The last decade has seen significant investments in diversity from major networks and advertisers seeking to have both content and content producers better reflect the communities and experiences that make up the United States (e.g., Procter & Gamble, 2022). Is this increased investment translating to real onscreen racial representation or to improved racial attitudes among Americans? As a team of computational linguists and social psychologists, we propose to pair Natural Language Processing (NLP) methods with social psychological experiments to measure novel features of racial representations in TV scripts and test how these features affect viewers’ racial attitudes. Specifically, we will specify linguistic metrics for the Presence, Positivity, and Positionality—what we call the 3Ps of representation—of Black and White characters in TV scripts. In tandem with these NLP efforts, we will experimentally test whether the same 3P features can improve viewers’ perceptions of Black characters and their broader racial attitudes. Finally, we will develop an AI script-analysis tool that equips content creators and network executives with knowledge to better understand and improve racial representation on TV.
Name | Role | School | Department |
Main PI | School of Humanities and Sciences | Psychology | |
Co-PI | School of Humanities and Sciences | Linguistics | |
Co-PI | School of Humanities and Sciences | Psychology | |
Co-PI | School of Humanities and Sciences | Psychology |
Motivational drives guide decision making. Animals regulate behavior differently depending on their drive states. Recent progress in neuroscience has begun to dissect how animals encode drives, such as thirst, and how they are translated into behavior — at the neuronal level. Artificially intelligent (AI) agents also must act in response to motivations that vary over time. Rapid development and widespread impact of AI systems has highlighted the urgency of understanding precisely how agent behavior is guided by motivational drives. What algorithms are implemented within an agent's neural networks that translate drive states and high-dimensional context into actions? Inspired by neuroscience, we will investigate AI systems at the circuit level and seek a mechanistic understanding of how AI agents represent drives at the neural level, and how these representations impact behavior.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education | |
Co-PI | School of Engineering | Bioengineering |
The number of older adults 60+ is expected to double—and 80+ triple—by 2050, creating urgency for innovative solutions to improve care practices. Ambient intelligence has the potential to improve the work conditions of personal health aides, workers that are essential for supporting older adults in need of assistance. It can collect important health information for a patient’s care, automatically document care activities, and suggest health and wellbeing activities tailored to a specific patient and their environment. This support can augment the capabilities of personal health aides, potentially leading to career growth, better pay, and better treatment by patients and agencies. We will work with personal health aides to design a system that will improve their work conditions while enhancing the care they provide. We will build the system to be more privacy-preserving than cameras by using a variety of sensing mechanisms and combining their information via a multi-modal inference algorithm. We will also utilize voice assistants to ask relevant medical questions, provide learning opportunities, and deliver a customized care plan in an intuitive manner. We aim to use ambient intelligence to reimagine care work with personal health aides and older adults while actively considering the ethical implications of the system.
Name | Role | School | Department |
Main PI | School of Engineering | Civil and Environmental Engineering | |
Co-PI | School of Medicine | Psychiatry and Behavioral Sciences | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Medicine | Med/Primary Care and Population Health | |
Co-PI | School of Engineering | Civil and Environmental Engineering |
Generalized learning of concepts in children during development is often dependent on semantic, contextual, encultured, and scaffolding processes that enable transfer learning and development of internal cognitive representations. However, a large number of children suffer from a variety of learning deficits including developmental dyscalculia or mathematical learning difficulties (MLD). These have been shown to be widely heterogeneous, and learning related representational deficits may arise from multiple domain-general or domain-specific mechanisms. Standard training and educational practices may demonstrate a neurotypicality-bias – they are often geared towards providing effective scaffolding mechanisms to optimally support the development of robust internal representations for transfer learning in neurotypical children, but may be relatively ineffective in children with MLD.
Along these lines, we propose debiasing training materials and methods for imparting mathematical training to optimally support scaffolding mechanisms for generalizable and transfer learning in non-typical populations who may otherwise demonstrate learning deficits. Typicality biases – for example, the lack of adequate use of virtual or physical manipulatives – place the burden of slower learning on children with learning deficits, rather than on teaching practices. Our specific goal is thus to assist children with MLD develop robust latent representations related to numbers and mathematical concepts that support better generalizability and transfer learning.
This objective will be pursued via the design of personalized AI networks – biologically inspired neural networks that have been structurally or functionally fine-tuned to represent distinct individualized profiles of learning deficits. Such networks will serve as a simulation environment to design and evaluate the efficacy of diverse and non-typical training techniques that can improve learning in diverse groups of children. This innovative approach will facilitate creation of optimal cognitive tutoring strategies in children with MLD. It is our hope that successful outcomes will lead to solutions that may reduce the learning gap between neurotypical children and children with MLD.
Name | Role | School | Department |
Main PI | Stanford School of Medicine | Psychiatry & Behavioral Sciences |
A collaboration among faculty, PhD students, and postdocs at Stanford, Oxford, and Imperial College London, aimed at merging expertise in the history and theory of democracy, practical ethics and law, and intelligent systems and self-organization. Together, we will develop new work at the intersection of AI ethics and “democracy by design.” Our project is intended to augment existing work on AI ethics that centers on questions of rights and welfare. Among the distinctive features of our proposed collaboration are (1) resituating AI ethics in the context of democratic politics, (2) refocusing the debate on human flourishing, (3) formalizing (1) and (2) in socio-technical systems. We anticipate producing a series of scholarly papers, a model university course, and an intensive AI policy course for thought leaders. Our collaborative work will be advanced through a series of workshops, to be held at each of our institutions, with a culminating conference. Our goal in reframing the debate on AI ethics around democracy and flourishing is to offer scholars, students, and policy-makers better resources for evaluating the possible benefits and harms of new and emerging AI technologies.
Name | Role | School | Department |
Main PI | School of Humanities and Sciences | Political Science |
Name | Role | School | Department |
Main PI | School of Engineering | Bioengineering |
As generative AI tools rapidly transform the media landscape, filmmakers are grappling with how this technology will shape their creative practice and livelihoods. Most of the current conversation focuses on generative AI’s influence on the future of film production, overlooking the potential for filmmakers to play an active role in influencing AI development through stories. Stories—and storytellers—have immense power to shape public discourse. Popular film and fiction narratives (e.g., The Terminator) disseminate social and technological visions that influence policy and impact research funding. Through cinematic prototypes, audiences imagine how technologies like AI might become part of our society and embedded in our everyday lives. The current landscape of AI narratives tend to utopian and dystopian extremes, but at their best, films can be used to challenge traditional mythic narratives, highlight faults in a system, and introduce new concepts into public discourse. Our one-and-a-half day spring convening will bring together an intimate group of AI and HCI researchers, science fiction screenwriters, film directors, and producers to craft new depictions of AI, weaving in recent advancements and pressing philosophical/ethical concerns. Unlike traditional conferences, Stories for the Future centers speculative design activities and collaborative storytelling. Our hope is to create a space for curiosity, disagreement, creativity, and generative discourse that will foster long-term relationships and collaborations between researchers and filmmakers.
Name | Role | School | Department |
Main PI | School of Humanities and Sciences | English |
Over a quarter of the world’s population is physically inactive, failing to meet global recommendations for weekly physical activity. Mobile health applications present a promising avenue for low-cost, scalable physical activity promotion, but often suffer from small effect sizes and low adherence rates, particularly in comparison to human health coaching. We propose an LLM-based conversational agent for personalized health behavior change to encourage physical activity. Our proposed system architecture extends a purely conversational LM’s abilities by exposing APIs for querying personal data and for important tasks like goal setting and fact checking. Unlike existing approaches, this architecture allows the system to integrate various sources of context, including semantic information captured in natural language interaction (e.g., high-level goals and motivation, life circumstances and constraints) and quantitative data about the user’s physiology and behaviors (e.g., biosignals from a wearable). As a first step, we will conduct needfinding and usability studies with various stakeholders while simultaneously iterating on a functional implementation of our system. Once the system design is complete, we plan to evaluate the conversational agent as a technology probe in a 4 week pilot study with approximately 20 participants.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science |
Differentiation, a crucial process of adjusting instruction to cater to individual learners, bolsters both academic and non-academic outcomes. However, 70% of educators face challenges addressing student diversity, primarily due to time constraints and lack of appropriate tools. Research with teachers and district administrators suggest that, while many districts provide high-quality curriculum materials, these resources are typically one-size-fits-all and assume grade-level proficiency. Teachers then have to spend considerable time customizing and scaffolding these materials to suit the varied needs of their students. In partnership with a network of school districts, this project is focused on the potential for AI-powered tools to support middle-school math teachers to adapt and tailor their curriculum for the specific needs of their students.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education | |
Co-PI | Graduate School of Education | Graduate School of Education |
Human-caused global change challenges ecosystem conservation and restoration efforts. Climate change, overexploitation of resources, biological invasions, and land-use changes bring additional uncertainty to the efficacy of management actions. To address this challenge, ecosystem management has focused on ecosystem resilience, the ability of the ecosystem to withstand disturbance while preserving its function. Measuring and tracking ecosystem resilience has focused on expensive empirical measurements and scope-limited theoretical models. This project will develop an AI-based approach to measure resilience that integrates insights from theoretical models with the realism and specificity of monitoring data. Our vision is for these AI tools to advance knowledge and applications of ecosystem resilience to inform management, conservation and restoration with a smaller cost.
Name | Role | School | Department |
Main PI | School of Sustainability | Center for Ocean Solutions | |
Co-PI | School of Sustainability | Earth System Science |
Nearly 100,000 individuals in the US are diagnosed with melanoma each year. Sunburns and ultraviolet radiation exposure have been identified as risk factors for melanoma and this risk factor is increasing as a result of the changing climate. As the climate warms, the ozone layer that protects us from ultraviolet radiation (UVR) is being depleted leading to more UVR reaching the earth's surface. As a consequence, people need to take more protective actions to prevent skin damage. Tailored educational messages have been shown to be more efficient and effective at improving sun protective behaviors than generic materials. Indeed, immersive technologies including Augmented Reality and Virtual Reality, have the power to enhance behavioral modification and habit formation. We will use Generative AI, a form of machine learning that is able to produce text, video, images, and other types of content, as an innovative method to create tailored, immersive educational interventions that meet the needs of college students from diverse racial and ethnic backgrounds to promote a better understanding of sun safety practices. Students will be able to create a personalized avatar representing their skin tone and facial features, then create specific scenarios to foreshadow skin aging based on the sun protection behaviors chosen and skin tone. We will train Generative AI to demonstrate the features of UVR-induced photodamage on the aged avatar face in the future. Our long-term goal is to use Generative AI for a positive human impact to prevent melanoma.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education | |
Co-PI | School of Humanities and Sciences | Communication | |
Co-PI | School of Medicine | Dermatology | |
Co-PI | School of Medicine | Dermatology | |
Co-PI | School of Medicine | Dermatology |
The ability to write is a critical skill that supports the transition from K-12 schooling to post-secondary employment, education, and independent living. Societal expectations for students with intellectual and developmental disabilities (IDD; formerly known as mental retardation; including students who are on the autism spectrum and students with Down syndrome) continue to increase. Thus, it is essential that we understand how to support learners with IDD by using cutting edge technologies like artificial intelligence (AI). Write Path - AI will allow us to design and evaluate the use of an AI tool to support and enhance content area writing outcomes for students with IDDwho are in secondary settings. Our proposed study aims to integrate evidence-based interventions with AI technology to create an AI-based tool. The tool will provide structured support to help students with IDD improve their content area writing. It will enable them to effectively identify the main idea or gist in content area readings and craft well-structured written responses to purpose questions. Our project is guided by these research questions: (1) Do AI tools enhance content area writing outcomes for students with IDD in middle and high school? and (2) What are the perceptions of students with IDD, their family members, and educators regarding the use of AI tools and their ability to enhance writing outcomes? Enhancing writing outcomes with AI has the potential to substantially improve post-secondary outcomes and to integrate individuals with IDD more fully into our society.
Name | Role | School | Department |
Main PI | Graduate School of Education | Graduate School of Education |