When the Algorithm Becomes the Dietitian: Large Language Model Hallucinations, Algorithmic Food Environments, and the Erosion of Nutritional Autonomy

Abstract

Large language models (LLMs) can answer dietary questions, but their use as substitutes for registered dietitians remains insufficiently quantified. Evaluations have identified inaccurate nutrient targets, incomplete clinical advice, and insufficient personalization, alongside useful performance on some general nutrition tasks. Social media and food-delivery interfaces may also influence dietary choices through food cues, social endorsement, and choice architecture. This conceptual perspective presents ADRIFT (Algorithmic Dietary Reconfiguration through Informational, Framing, and Technological mechanisms) to examine how these influences might interact. Drawing on a targeted, nonsystematic synthesis of nutrition, behavioral, neuroscience, and AI literature, ADRIFT distinguishes three axes: informational errors and unsupported advice; framing through content and interface design; and adaptive personalization and cognitive delegation. Direct evidence supports selected component mechanisms, but their convergence into a feedback loop that reduces nutritional autonomy remains hypothetical. We propose a six-category taxonomy covering hallucinations and related nutrition errors, define nutritional autonomy erosion as declining capacity for informed, self-directed dietary decisions, and outline eight testable hypotheses. The framework concerns unsupervised use and does not predict that language models will replace dietetic professionals. Longitudinal validation, platform experiments, and clinical safety evaluations are needed to determine whether, for whom, and under which conditions the proposed harms occur.