Search

Materials AI models invariably produce predictions for every input, even when operating under high uncertainty, distributional shifts, or conditions where the potential costs of error far outweigh any informational benefit. This reflexive prediction habit represents a critical gap in the field, as models rarely—if ever—choose to abstain despite the high-stakes nature of materials discovery, where erroneous outputs can trigger wasteful synthesis campaigns, compromise safety assessments for novel compounds, or mislead decisions involving rare-event phenomena such as phase instabilities under extreme conditions. Justified abstention is defined here as the deliberate, epistemically grounded decision by a model to withhold any prediction when the expected utility of outputting a value falls below the utility of remaining silent, thereby prioritizing scientific integrity over forced coverage. This paper articulates a novel theory of justified abstention built on three core principles—competence boundary, risk threshold, and resource consideration—alongside five explicit operational criteria that together provide a principled framework for when abstention becomes not only permissible but obligatory in materials contexts. Four distinct types of abstention are delineated (input-based, prediction-based, risk-based, and resource-based), each with clear triggers and materials-specific illustrations that underscore their necessity. The implications extend to transformed design pipelines, where abstention mechanisms foster greater trustworthiness, enable more efficient allocation of experimental resources, and shift materials AI from indiscriminate oracles to responsible scientific partners capable of signaling their own epistemic limits. By embedding justified abstention as a core design feature rather than an afterthought, the framework addresses a longstanding oversight in the literature. It offers a pathway toward more reliable, ethically defensible AI systems for materials science.