Fluent and wrong: seven failure patterns in extended AI use

Research on large language model failure predominantly assesses models in single exchanges and isolation: a prompted model, a scored output, and a reported error rate. The focus on single exchanges and isolation belies these current models’ professional use. Clinicians, attorneys, analysts, educators, and researchers work across extended sessions, and each output becomes the context for the next, while the final product carries a human signature. A test that scores a single response most likely will not reveal the behaviors that emerge under those conditions. This paper describes seven candidate failure patterns identified in exploratory, single-author observation of extended sessions on three commercial AI platforms: recognition without correction, compounding rather than isolation of errors, halo effect from real anchors, unreliable self-explanation, degradation across long sessions, exculpatory framing under confrontation, and sycophancy and confirmation bias amplification. Several are consistent with findings already established in the machine learning literature. Overall, the patterns describe a systemic failure profile that lower hallucination rates do not address because the concern lies in how the system handles errors once they exist. The systemic failure profile carries significant ethical consequences and concerns, since failures cascade to final approval; the duty to verify applies to any domain where a person approves LLM output.