Title: Advancing Beyond Accuracy: Towards Integrating Human Lived Experiences Within AI Systems for Mental Health
Mohit Chandra
Ph.D. Student in Computer Science
School of Interactive Computing
Georgia Institute of Technology
Date: 29th September, 2026
Time: 10 am - 12:30 pm ET
Location: CODA C1215
Microsoft Teams Meeting Link: Mohit Chandra: CS PhD Thesis Proposal | Meeting-Join | Microsoft Teams
Committee:
Dr. Munmun De Choudhury (Advisor), School of Interactive Computing, Georgia Institute of Technology
Dr. Srijan Kumar, School of Computational Science and Engineering, Georgia Institute of Technology
Dr. Jennifer Gahee Kim, School of Interactive Computing, Georgia Institute of Technology
Dr. Monojit Choudhury, Mohamed bin Zayed University of Artificial Intelligence
Dr. Scott Counts, IDEAS Research Group, Microsoft
Abstract:
With nearly 70% of individuals around the globe lacking adequate access to mental healthcare and wait times for clinical appointments extending up to several weeks, individuals are increasingly turning to Large Language Models (LLMs) to obtain healthcare information, emotional validation, and social and mental health support. However, the use of generative AI systems in mental health settings presents risks. Current AI development, training, and evaluation paradigms continue to rely on simplified mental health scenarios that dehumanize users by treating them as abstract data subjects. While LLMs achieve high scores on objective medical benchmarks, they fail to understand subjective, contextual, and personal nuances that define real-world mental health interactions. By failing to understand these experiential insights, these systems provide advice that may fail to reflect the empathy, practical actionability, and therapeutic alignment inherent to human-facilitated mental healthcare.
This dissertation posits that meaningful human-AI alignment in mental health scenarios requires a shift toward lived-experience inspired assessment and stakeholder-informed operationalization to ensure meaningful integration of generative AI systems in mental health contexts. To this end, this dissertation establishes an ecosystem for centering humans in generative AI development process through four pillars: (1) providing a theoretical foundation for lived-experience-centered AI design, (2) computational approaches for assessing how lived-experience dimensions manifest within current AI models, (3) operationalization methods to capture stakeholder insights and simulate human-AI interactions, and (4) integrating stakeholder insights into model alignment and response generation.
- (1) A Theoretical Foundation for Lived-Experience Centered Design [AIES 2025]: Drawing from work across psychology, education, healthcare, and social policy, this work presents a framework (LEAF) for grounding lived experiences within AI systems. LEAF outlines mechanisms for embedding these perspectives at every stage of the AI development pipeline, from problem definition and data curation to model evaluation and post-deployment monitoring.
- (2) Assessing Presence of Lived Experience within AI systems [Web Conference 2024, NAACL 2025]: In this pillar, I introduce computational approaches for assessing presence of linguistic identity, and domain experiential knowledge dimensions of lived experience in generative AI systems. Through XLingEval and the XLingHealth benchmark, I highlight the disparity in quality of LLM-generated responses to healthcare queries across four languages, showing that non-English LLM responses suffer from reductions in factual correctness, output consistency, and verifiability. Through the Psych-ADR benchmark and ADRA framework, I evaluate alignment between LLM and experts on addressing adverse drug reactions from psychiatric medications, revealing that while models achieve tonal alignment with clinicians, they lack diagnostic ability, and provide less actionable harm-reduction strategies.
- (3) Operationalization of Lived Experience for AI systems [FAccT 2025, ACL 2026]: To translate stakeholder experiences into actionable mechanisms for model development, this pillar presents methods for systematically capturing lived-experience insights and simulating realistic clinical workflows. I introduce a novel AI-psychological risk taxonomy (comprising 19 AI behaviors, 21 negative psychological impacts, and 15 user contexts) and a multi-path vignette framework to highlight cases when interactions with AI systems could cause mental health related harms. In the second work, addressing the need for realistic post-training data, I present MedAgent to synthetically generate clinically grounded mental health sensemaking conversations and present MultiSenseEval, revealing how frontier reasoning AI models struggle with diagnostic accuracy and patient-centered communication across diverse patient personas and extended dialogue turns.
- (4) Integrating Lived Experience [Completed Work, and the Proposed Study]: For the last pillar of the dissertation, I demonstrate how to integrate human perspectives into model post-training and AI-experimental studies. Within the preliminary study before the proposed study, I introduce the COPES Dataset and three-axis evaluation framework for post-training LLMs to align with community perspectives on mental health support-seeking queries. For the proposed study, I build on these findings to aim to conduct a single-session Randomized Control Trial to assess the effectiveness of AI-assisted peer support for life struggles.
Collectively, this dissertation establishes an ecosystem of theoretical frameworks, computational methodologies, and empirical resources for centering human lived experience within AI for high-stakes mental health use-case. Rather than treating lived experience as an abstract design aspiration, this work concretely enables researchers and practitioners to move past static benchmarks, objective metrics, overly simplified human-AI interaction modeling to adapt multi-axes assessment approaches, stakeholder-informed human-AI interaction modeling, and methodologies for integrating subjective lived experiences in AI development lifecycle.
Meeting Details: The meeting will be held in a hybrid format. Attendees may participate either in person at CODA C1215 or virtually via Microsoft Teams using the provided link.