Introduction: Large visual-language models (VLMs) frequently exhibit "hallucinations," yet these errors are often treated as undifferentiated stochastic failures rather than specific cognitive deficits. In clinical neuroscience, pathological visual persistence, known as palinopsia, occurs when a patient continues to see a visual form after it has been removed from the environment. This study introduces a novel clinical-neuroscience testing platform designed to turn AI hallucinations into measurable cognitive phenotypes. By applying psychophysical rigor to VLM evaluations, we aim to detect and subtype visual persistence in AI agents compared to human baselines.
Methods: We developed Snellen-based engines for standardized visual acuity control and the Persistent Stimulus-Response Generation (PSRG), allowing structured A-B-C trial sequence consisting of stimulus-on, blank camera-on, and blank camera-off phases. Gemini Live Models were primed with tumbling "E" stimuli and subsequently presented with null trials (0 px stimuli or camera-off states). Persistence was operationalized as any contentful claim of an "E" orientation during these null trials.
Results: The data revealed that VLMs are highly susceptible to instructional and perceptual palinopsia. Following successful priming, models showed high persistence rates, often reporting directional orientations for over 20 consecutive blank trials. Statistically, these hallucination-like states manifested through significant reaction time instability and "motor block" phenomena, characterized by fragmented vocalizations and processing delays. Mechanistic subtyping successfully differentiated between context-dependent persistence during live camera feeds and procedural confabulations that persisted even when the system signaled a camera-off state.
Conclusions: This research demonstrates that AI hallucinations can be rigorously quantified using extinction metrics and persistence curves derived from clinical psychophysics. The manifestation of motor blocks and directional fixation suggests that these failures are tied to internal processing congestion or failures to flush visual buffers. This framework provides a critical new benchmark for AI alignment and safety, offering a pathway to diagnose and mitigate specific cognitive failure in multimodal AI.
