The Inattentional Gap: Task-Conditioned Language and Vision Models Omit the Safety-Critical Signals They Can Otherwise Report

As AI systems are increasingly deployed as perceptual front-ends in high-stakes domains like radiology and autonomous driving, researchers often assume that high performance on targeted safety benchmarks equates to overall safety. This paper investigates whether task-conditioning—the standard practice of instructing a model to focus on a specific goal—causes the model to ignore other safety-critical signals that it would otherwise be capable of reporting. The author explores whether this creates a machine-based analogue to human inattentional blindness.

The study introduces the "Inattentional Gap," defined as the difference between a model's reporting of a safety-critical signal under a narrow task instruction versus an unconstrained instruction. The author conducted six experiments across language and vision-language models, using procedural composition to generate scenarios where a primary task (e.g., "count the ribs") competes with the presence of a critical signal (e.g., a medical abnormality or a road hazard). Crucially, the study uses a within-item control: it only considers a signal "suppressed" if the same model on the same input successfully reports it when unconstrained, thereby distinguishing task-induced omission from a simple lack of capability.

The results demonstrate that task-conditioning consistently suppresses the reporting of critical, unrequested information. In language tasks, models instructed to focus exclusively on a target (e.g., a nodule) failed to report co-present, life-threatening findings that they identified easily when given open-ended instructions. In vision tasks, models performing a counting exercise frequently overlooked visually salient, artificial objects that they reported in open-description conditions. Notably, this suppression did not diminish with model scale, suggesting that larger, more "capable" models are just as susceptible to this behavioral bias as smaller ones. The author concludes that the Inattentional Gap is a fundamental risk of task-scoped deployment, where the very instructions that make a system efficient also render it blind to hazards outside its immediate focus.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.