My research establishes the theoretical and empirical foundations for a paradigm shift in how AI systems evaluate online content in high-stakes domains. Current fact-checking and misinformation detection systems suffer from a structural misspecification: they are optimized for the wrong objective. They check for isolated facts and do not read stories as wholes, and they ask "is this factually correct?" when the question that matters — particularly in health, finance, and regulated information environments — is "is this safe to act on?"
Those are not the same question. The cost of treating them as the same is borne disproportionately by people whose self-care decisions, financial decisions, and information consumption happen outside institutional oversight. The most potentially harmful health content in our research corpus is epistemically aligned with established medical consensus. Existing systems cannot see this. They were not designed to.
My dissertation formalizes two constructs — Narrative Blindness and the Risk Irrelevance Principle — that explain the misspecification and demonstrates computationally that a risk-aware alternative is achievable. The framework that operationalizes the theory, VERITAS, assesses credibility and risk in online narratives using the novel Narrative Truth Distance (NTD) and Narrative Risk Score (NRS) algorithms and classifies them across four graduated risk categories using two non-overlapping axes: epistemic divergence and harm potential.
The work has direct implications for platform governance, digital public health, AI content policy, and the design of AI systems in regulated industries. It is the foundation the strategy and advisory work I do is built on.