Taposh Dutta Roy
I created the Autonomy Index, co-authored audits of how frontier language models weigh clinical values, and lead AI strategy across a large hospital network at Kaiser Permanente. My work connects bioethics theory, empirical evidence, and governance that holds AI accountable in practice.
Director of Innovation and AI, Kaiser Permanente · Research affiliate, Harvard Medical School Center for Bioethics · MS in Bioethics, Harvard Medical School (2026) · Author, Intelligent Governance
Published and recognized inarXiv · Tech Policy Press · Harvard Medical School Center for Bioethics · H2O AI 100 (2026) · MIT Hacking Medicine (1st place) · Cureus · medRxiv
Essays and commentary
I argue that the central problem with AI mental health chatbots is legitimacy, not only alignment, and propose function-based classification of these products, post-market monitoring modeled on pharmacovigilance, and independent oversight with authority to delay a launch.
Read the essayOversight of health AI in the US is split across the FDA, CMS, CDC, and NIST, and no single agency owns the gaps between them. I argue that one-time product approval cannot catch harms that surface months or years later in specific patient groups, and propose four steps: Chief AI Officers in each agency, a National Health AI Council of senior officials from those agencies, lifecycle governance with harmonized transparency and bias standards, and public safety audits with coordinated post-market monitoring.
Read the essayThe ideas behind the work
Most responsible AI conversations ask whether a model is compliant. I ask whether it can be trusted, whether it reasons well about ethics, and whether anyone has the legitimacy to decide what "well" means. Each of the five holds only while the other four keep running, so I treat them as one loop rather than a list.
Idea 01 of 05
Asking whether AI shares human values is not enough. We have to ask whether the institutions governing it are legitimate enough to decide which values count, whose risks matter, and who is accountable when harm occurs.
From: Tech Policy Press essayOriginal frameworks and research
With Rebecca Weintraub Brendel, MD, JD
I developed the Autonomy Index, a 13-item instrument for evaluating patient autonomy in clinical ethics, grounded in Appelbaum and Grisso's model of decisional capacity. In an IRB-approved study, 30 trained human reviewers and nine frontier language models rated the same 50 clinical ethics vignettes.
Key finding: Frontier models systematically underweight patient autonomy, rating it 11.0 points lower than human experts on average (95% CI 6.3 to 15.8; d = 0.66). That gap bears directly on informed consent and shared decision-making as these models enter clinical workflows.
Harvard Medical School and Beth Israel Deaconess Medical Center, with Payal Chandak, Isaac Kohane, Gabriel Brat, Rebecca Brendel, and colleagues
A framework for auditing value pluralism in medical AI: 50 clinician-verified dilemmas and an attribution method that recovers how models prioritize autonomy, beneficence, nonmaleficence, and justice directly from their decisions, compared against a physician panel.
Key finding: Frontier models discuss competing values but commit to consistent value preferences. Deploying one model at scale risks replacing clinical pluralism with a deployment monoculture.
I designed BAIA, an agentic AI framework for bioethical clinical decision support. It unifies three established methods of clinical ethics (Beauchamp and Childress's principlism, casuistry, and narrative ethics) in an extensible multi-agent architecture with a four-stage pipeline. Every advisory output includes its reasoning chain, and new ethical frameworks, clinical domains, and AI systems can be added without restructuring.
What it does: Gives clinicians and ethics consultants a transparent AI advisor that supports human moral reasoning rather than replacing it.
Co-author: Rebecca Weintraub Brendel, MD, JD
Book
My practitioner's guide to AI governance, which treats AI accountability and post-quantum readiness as one compliance problem rather than two. Covers AI risk management frameworks, adversarial attacks, industry perspectives in healthcare and finance, and governance for quantum-era systems.
Reviewed by Anil Kumar Pandey, technology leader at Workday.
View the bookFrom principles to practice
Frameworks matter only if they survive deployment. At Kaiser Permanente I lead AI strategy across a large hospital network, where my teams build predictive, prescriptive, and simulation-based products for hospital operations, finance, and population health.
Co-authored "Validation of a Hospital Digital Twin with Machine Learning," which uses machine learning to validate agent-based hospital simulations.
ReadNamed inventor on USPTO patent application 19/247,983 (filed June 2025) for AI-powered hospital digital twins.
1st place, MIT Hacking Medicine (HospiTwin). Named to the 2026 H2O AI 100 list.
What's next
Today's responsible AI tools tell us whether a model is risky or compliant. They do not tell us whether it reasons well about ethics. These are the open problems I am working on.
The problem: Responsible AI ontologies such as AIRO, VAIR, and DPV-AI describe risks, regulations, and documentation. Bioethics ontologies such as ApplE, OBEST, and iOntoBioethics describe ethical principles and concepts. No existing ontology connects the two, so there is no shared, machine-readable language linking an AI system's risk profile to the clinical ethics reasoning it performs.
My approach: Build a formal ontology that bridges AI risk frameworks and clinical ethics reasoning, grounded in the BAIA domain taxonomy (principlism, casuistry, narrative ethics, deontology, virtue ethics) and the Autonomy Index constructs, giving governance committees, EHR vendors, regulators, and model developers a common vocabulary to describe, audit, and compare how AI systems handle ethical decisions.
Why it matters: You cannot govern what you cannot describe consistently. An ontology turns ethical evaluation from one-off studies into reusable, interoperable infrastructure.
The problem: A model that passes an ethics evaluation today can behave differently after its next version update, and most health systems have no way to detect it.
My approach: Turn the Autonomy Index into a standing benchmark that health systems rerun against a human expert baseline whenever a vendor updates a model, with defined roles and clear thresholds for review and pause.
The problem: How much weight physicians give patient autonomy varies across cultures and medical traditions. If language models encode one tradition's weighting, a single global deployment quietly exports it to every patient.
My approach: Extend value audits beyond North American reference panels and study how models should adapt to the values of the patients and communities they serve.
The problem: AI is moving from answering questions to taking actions inside clinical and operational workflows. Accountability frameworks built for static prediction models do not cover agents that plan, act, and hand off to each other.
My approach: Combine agentic ethics architectures (BAIA) with audit-grade logging (MedLog) so every agent decision in a clinical setting is traceable, reviewable, and attributable to an accountable human.
Interactive
Five questions drawn from my research and the decisions health systems face right now. See how you handle policy, model monitoring, a real clinical ethics case, ethical theory, and the trolley problem, then find out your ethical lens.
About 2 minutes · No sign-up · Nothing is stored
Academic service and community
If you are shaping responsible AI governance, evaluating clinical AI, or looking for a speaker, reviewer, or advisor on AI ethics in healthcare, I would be glad to connect.