Failure Modes and Trade-offs in Low-Risk AI Systems An Investigative Analysis of Cognitive Access, Non-Clinical AI, and Governance Boundaries

Low-risk artificial intelligence systems are increasingly deployed in public-sector, educational, and mental health–adjacent contexts under the assumption that the absence of clinical decision-making implies limited harm. However, emerging practice suggests that risk may arise not from system outputs, but from interaction design, access modality, and long-term use dynamics. This paper provides an investigative analysis of failure modes and non-resolvable trade-offs in low-risk and non-clinical AI systems, with a focus on language-based interaction, cognitive access, and indirect exclusion. Building on existing governance frameworks, including SUPREME-1, OMNIA-1, and EROS-1, the analysis stress-tests key assumptions underlying low-risk classification. Four recurrent failure modes are identified: channel-induced cognitive exclusion, over-reliance under non-clinical framing, responsibility diffusion, and temporal accumulation of risk. Rather than proposing new regulatory categories, the paper highlights mechanisms through which harm may remain invisible to static, output-focused assessments. The analysis supports a shift toward interaction-based and temporally aware governance approaches, emphasizing the need for monitoring, cognitive access evaluation, and recognition of trade-offs that cannot be fully optimized. The paper aims to inform regulators, auditors, and system designers concerned with proportional and evidence-informed AI governance.