Blinded professional ratings of artificial intelligence-generated and dietitian-designed meal plans

Artificial intelligence-assisted systems increasingly produce personalized meal plans, but their value depends on whether clinicians judge those plans credible and usable. We evaluated complete plans from a constrained artificial intelligence-assisted pipeline under blinded professional review.

In a cross-sectional exploratory web study, 54 eligible anonymous evaluator records from Brazilian dietitians contributed 130 evaluations of five matched plan pairs, with one pipeline-generated and one dietitian-designed plan per persona. Evaluators rated nine items across three domains and were offered an optional pairwise plan preference task. The overall score required all nine items and was available for 124 evaluations.

Pipeline-generated plans received a mean rating of 3.67 (standard deviation, 0.57) (n = 59) compared with 3.11 (standard deviation, 0.73) for comparator plans (n = 65) (d = 0.86). Among 38 evaluator records with complete scores for both origins, 31 favored pipeline-generated plans, 6 favored dietitian-designed plans, and 1 had equal mean scores for both origins (paired t[37] = 5.32, P < 0.001). Pipeline-generated plans received higher ratings in all five persona pairs. This pattern was descriptive because evaluator records overlapped across pairs. Domain differences were +0.76 for nutritional adequacy, +0.69 for technical quality, and +0.21 for practical applicability. In the optional task, 32 of 42 eligible pairwise choices selected the pipeline-generated plan as more appropriate.

Under blinded presentation, practicing dietitians gave plans from the constrained pipeline a mean rating above the neutral anchor of the 5-point scale. These findings support prospective evaluation of a structured first draft workflow, with required clinician edits and patient-relevant outcomes as endpoints.