7 preprints
This paper develops a granular taxonomy of epistemic failure in large language models, distinguishing between confident error, synthetic coherence, and context-sensitive reliability collapse. I argue that benchmark performance cannot substitute for process-level justification and propose an audit architecture grounded in process reliabilism and virtue epistemology.
By analyzing how symbolic structures are compressed during representation learning, this preprint examines the gap between linguistic fluency and semantic grounding. I show why lexical competence in model outputs can mask referential fragility and propose criteria for distinguishing symbolic simulation from meaningful reference.
This paper argues that dignity-preserving design requires more than fairness metrics. Drawing on Kantian ethics and capabilities theory, I outline institutional and interface-level constraints that preserve contestability, recognition, and agency in automated welfare, labor, and healthcare decisions.
This preprint critiques benchmark-centric evaluation paradigms by demonstrating how similar scores can conceal divergent epistemic profiles. I distinguish performative accuracy from knowledge-relevant reliability and introduce a framework for stress-testing models under epistemically novel conditions.
This paper connects inference economics to epistemic quality. I show how latency and cost optimization decisions can systematically redistribute model reliability across user groups, creating hidden normative asymmetries. The analysis proposes governance principles for ethically constrained serving policies.
Focusing on recommendation, moderation, and ranking infrastructures, this preprint examines how responsibility diffuses across distributed technical and institutional actors. I propose a layered accountability model for tracing normative responsibility when social harms emerge from interacting machine systems.
Rather than treating agency as binary, this paper defends a graded account based on functional autonomy, representational plasticity, and normative exposure. I argue that this framework clarifies public confusion about AI agency while avoiding both anthropomorphism and reductive instrumentalism.