This preprint presents a structured observational report of anomalous interaction events involving a commercial multimodal AI system. The study documents repeated instances of non-correspondence between user-submitted visual inputs and system-generated outputs within a constrained temporal window (April 6, 2026). Across multiple interaction instances, the system produced internally coherent outputs that did not match the reported input content. Observed behaviors include semantic drift across sessions, safety-layer activation, output generation under null input conditions, and failure of input binding in task-specific operations. A replication test further demonstrates that image-description requests may trigger generative outputs unrelated to the provided input, suggesting inconsistencies in the association between input data and processing operations. The report adopts a strictly observational and descriptive methodology (N=1), without asserting causal mechanisms. Instead, it provides a detailed record intended to support future investigation into multimodal system reliability, input fidelity, and processing transparency in human-AI interaction. This work is released as a preprint for academic and technical scrutiny.
Oberon de Mello Campos de Almeida (Tue,) studied this question.