My research examines how organizations can govern and use artificial intelligence when the signals available to decision makers do not fully reveal the meaning, motivation, or risk that matters. I am particularly interested in situations where language evolves to evade detection, users adapt or circumvent technological controls, and human supervisors must make consequential decisions about autonomous systems. Across these settings, I study how governance can become more accurate, auditable, and proportionate to the risks involved.
A complementary stream of my research focuses on healthcare information systems. In this work, I develop interpretable and deployable AI methods that support clinical judgment and operational decisions while preserving transparency for physicians and managers. Across both streams, I combine large language models, natural language processing, agentic pipelines, deep learning, and behavioral and causal experiments.
Governing Beneath the Surface: Meaning, Motive, and Risk in AI-Mediated Systems
My dissertation develops a unified account of AI governance for settings in which surface-level observations are insufficient for responsible intervention. The three essays examine different forms of divergence between what a governance system can readily observe and the latent state it must understand: the meaning encoded in language, the motive behind circumventing a control, and the actual level of human control over an AI agent. Together, they develop technical and behavioral approaches for governance that is context-sensitive, auditable, and aligned with the consequences of error.
Essay 1 — Meaning: My job-market paper develops a mechanism-grounded method for detecting evolving coded language, reconstructing its concealed meaning, and distinguishing benign from malicious intent in context. Rather than matching known terms, it identifies the encoding mechanism by which meaning is hidden and then proposes and verifies candidate decodings, so it flags coded expressions it has never seen. Every decision carries an explicit evidence trace, and accumulated coded-language knowledge is retained as governed, revisable memory, improving both the accuracy and the auditability of AI-enabled content moderation.
Essay 2 — Motive: This study reconceives attempts to circumvent generative-AI guardrails as forms of technology appropriation. It distinguishes malicious pursuit from legitimate workarounds and examines how user motives and work context shape engagement, the packaging of reusable methods, and their diffusion.
Essay 3 — Risk: This behavioral study examines when human oversight of autonomous AI agents produces substantive control and when approval becomes largely symbolic. It compares pre-action approval, post-action trace review, and risk-based hybrid oversight to identify which architecture best supports prevention, detection, and mitigation.
Published Work
Firoozfar, H., Abolhasani, M. S., Mousavi, R., & Hu, P. J.-H. (2026). “Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection.” EMNLP 2026 Main Conference. arXiv: 2606.27314.
Firoozfar, H., Xu, D., Hu, P. J.-H., & Hwang, S.-Y. “CirrhosisCare: A Decision Support System Enabled by Complication-Aware Deep Learning to Predict One-Year Mortality among Cirrhotic Patients.” Decision Support Systems, accepted.
Heidary Dahooie, J., Vanaki, A. S., Firoozfar, H., Zavadskas, E. K., & Čereška, A. (2020). “An Extension of the Failure Mode and Effect Analysis with Hesitant Fuzzy Sets to Assess Occupational Hazards in the Construction Industry.” International Journal of Environmental Research and Public Health, 17, 1442.
Heidary Dahooie, J., Zavadskas, E. K., Firoozfar, H., Vanaki, A. S., Mohammadi, N., & Brauers, W. K. M. (2019). “An Improved Fuzzy MULTIMOORA Approach for Multi-Criteria Decision Making Based on an Objective Weighting Method and Its Application to Technological Forecasting Method Selection.” Engineering Applications of Artificial Intelligence, 79, 114–128.
Manuscripts in Preparation
Firoozfar, H., Hu, P. J.-H., Mousavi, R., & Abolhasani, M. S. “MIND: Mechanism-Grounded Detection of Indirect Linguistic Encoding for AI-Enabled Content Moderation.” (Job-market paper)
Firoozfar, H., Abolhasani, M. S., & Hu, P. J.-H. “Interpretable Privileged Learning for Hospital Capacity Decisions: A Concept-Relational Distillation Approach to Trauma Length-of-Stay Prediction.” Under review at Information Systems Research.
Firoozfar, H., Hu, P. J.-H., Mousavi, R., & Abolhasani, M. S. “When AI Guardrails Become Workarounds: Technology Appropriation and Intent-Aware Governance.”
Work in Progress
Firoozfar, H., Hu, P. J.-H., & coauthors. “Symbolic versus Substantive Oversight: How Oversight Architecture Shapes Human Control of Agentic AI.”
Abolhasani, M. S., Firoozfar, H., Hu, P. J.-H., & coauthors. “Five Registers of Blame: Ethical Verdict and Fallacy Auditing for Blame Adjudication in Online Harm Narratives.”
Conference Presentations
MIND: Mechanism-Grounded Detection of Indirect Linguistic Encoding for AI-Enabled Content Moderation
— CIST 2026
— INFORMS Workshop on Data Science 2026.
When AI Guardrails Become Workarounds: Technology Appropriation and Intent-Aware Governance
— CIST 2026.
Detecting Emerging Coded Language for Platform Governance: A Taxonomy-Guided LLM Approach
— INFORMS Annual Meeting 2026.
CirrhosisCare: A decision support system enabled by complication-aware deep learning to predict one-year mortality among cirrhotic patients
— CIST 2025.
A knowledge distillation method to enhance interpretability in predicting HCC among cirrhotic patients for clinical decision support and patient management
— INFORMS Workshop on Data Science 2024 (Best Student Paper Award Finalist).