Research Projects
Publications: Google Scholar · Data access: Available upon request
My research develops methods and evaluation frameworks to improve the safety, security, and ethical reliability of artificial intelligence systems, with a particular focus on large language models. I study how AI systems behave in sensitive or high-risk deployment settings and design approaches to systematically assess and mitigate associated risks. This work supports the responsible deployment of AI in complex domains, including healthcare, and provides evidence to inform policymakers and non-technical stakeholders working to reduce algorithmic harm.
Stress Testing and Failure Analysis of Large Language Models
This line of research focuses on developing methods for stress testing large language models to systematically identify failure modes, including jailbreak vulnerabilities, unsafe outputs, and robustness breakdowns under distribution shift. The work emphasizes evaluation frameworks and adversarial testing approaches for characterizing model behavior in high-risk or deployment-relevant settings.
Responsible AI
Developed through close collaboration with interdisciplinary teams at NU’s Responsible AI Practice and industry partners, this work spans domains such as gaming, healthcare, and finance. Spanning domains such as gaming, healthcare, and finance, the work translates technical evaluation, monitoring, and guardrail methods into applied settings. Research outputs are informed by real-world deployment constraints and emphasize issues that emerge over the system lifecycle.
AI Systems and Suicide
This work examines how AI systems interact with suicide-related content, focusing on system behavior, responses, safeguards, and evaluation. Ongoing work includes (i) understanding how suicide is expressed through language and emotional expression, and (ii) identifying the medical and sociodemographic contexts that drive suicide risk. These findings are used as evidence to inform policy and deployment decisions involving AI systems.
Building an Actionable Framework to Responsible AI Integration in Clinical and Public Health Contexts
Building on the AI Ethics Box, this work adapts bioethical principles to healthcare settings and links them to concrete technical methods for evaluation, governance, and post-deployment monitoring. The framework addresses the gap between high-level ethical guidance and operational decision-making in clinical and public health contexts. It is refined through interdisciplinary collaboration and stakeholder co-design, supporting practical use by clinicians, health system leaders, and developers in real-world AI deployment.
RAI4MH: Responsible AI for Mental Health Initiative
RAI4MH is an international partnership developing evidence-based guidance for the responsible use of AI in mental health contexts. As the U.S. lead, I contribute to coordinating interdisciplinary collaboration across computer science, mental health, ethics, and policy. Key outputs include white papers that synthesize international expert consensus and contribute directly to policy discussions, including a POSTnote for the UK Parliament on responsible AI in mental health.
The Hidden Weight of Words: Using Natural Language Processing to Uncover Weight Stigma in Health Insurance Coverage Policies
This work examines how weight stigma is embedded in health insurance coverage policies for obesity care and how such language shapes access to evidence-based treatments. Using natural language processing methods, the study analyzes large-scale insurance policy documents to identify, characterize, and quantify stigmatizing language across payer types. The resulting evidence supports policy reform, coverage standardization, and the development of stigma-free obesity treatment benefits, with findings disseminated through open-access publications and scientific venues.
Using Natural Language Processing to Detect Bias in Emergency Department Clinical Notes
This line of research focuses on developing methods for stress testing large language models to systematically identify failure modes, including jailbreak vulnerabilities, unsafe outputs, and robustness breakdowns under distribution shift. The work emphasizes evaluation frameworks and adversarial testing approaches for characterizing model behavior in high-risk or deployment-relevant settings.
COMPASS: Co-produced Mental Health Participation and Safety Strategy for AI
COMPASS develops a co-produced AI literacy and safety toolkit for people living with mental health conditions, their families, and clinicians. Grounded in lived-experience participation, the work identifies priorities and concerns around AI use in mental health care and translates them into accessible guidance. The toolkit supports critical engagement with AI tools, helping users articulate expectations, navigate ethical issues, and participate meaningfully in decisions about digital and clinical technologies.
SHAPE-AI: Smart Health Analytics and Predictive Engagement with AI
This work studies how AI learning companions intersect with student mental health, cognition, and academic outcomes. It integrates multimodal data including biometric signals (sleep, activity, physiology), academic performance, and patterns of AI use to model wellbeing and learning trajectories. The study examines AI use as a digital phenotype of academic engagement and supports evidence-based guidance for intervention design and responsible AI deployment in higher education.
Home 