Large language models miss protein quality in vegan meals

Technical Summary: Large Language Models Miss Protein Quality in Vegan Meals

Problem Statement
General-purpose Large Language Models (LLMs) are increasingly utilized for meal planning, yet their nutritional reliability, particularly regarding protein quality, remains untested. This gap is critical as dietary shifts toward plant-based eating occur; without deliberate combination of plant proteins, such diets often yield lower-quality protein. The stakes are heightened for older adults, whose muscle mass preservation and counteraction of sarcopenia depend on adequate intake of indispensable amino acids (IAAs) at the meal level, not just daily totals. While plant proteins vary in IAA composition and digestibility, current evaluations of LLMs in clinical nutrition are largely qualitative, lacking validation against domain-specific, quantitative benchmarks like the Meal Protein Quality Score (MPQS).

Methodology
The study conducted a first-of-its-kind, meal-level, amino-acid-resolved benchmark of two state-of-the-art LLMs: ChatGPT (OpenAI, GPT-5.4) and Claude (Anthropic, Claude Opus 4.6).

• Generation Protocol: Each model generated 250 single-serving vegan dinner recipes for older adults, totaling 500 recipes. The generation was distributed deterministically across 25 cuisine inspirations and 10 meal formats to vary cultural and dish-format framing while keeping the optimization task constant.
• Constraints: The prompt requested recipes with ≥20 g of protein per serving and "good protein quality" via smart plant protein combinations. It explicitly excluded supplements, fortified products, and meal replacements. The prompt was minimal, avoiding prescriptive instructions on how to achieve quality, to reflect realistic practitioner use.
• Evaluation: Ingredients were mapped to the Dutch Food Composition Database (NEVO) and processed through the Alpha-Tool. This tool calculates the MPQS, which integrates digestibility-adjusted IAA intake, total protein, and personalized requirements for a reference older adult. A score of ≥100 indicates adequate protein quality.
• Statistical Analysis: Comparisons utilized Welch's t-tests for continuous variables and Fisher's exact tests for binary outcomes (meeting quality thresholds). Linear regression assessed the effects of cuisine and meal format on MPQS.

Key Results

• Protein Quantity vs. Quality: All 500 generated recipes met the protein quantity target (≥20 g). However, a significant portion failed the quality threshold (MPQS ≥ 100). ChatGPT met the quality target in 81.2% of cases, whereas Claude achieved this in only 59.6%. Consequently, 18.8% of ChatGPT and 40.4% of Claude recipes met quantity targets but were nutritionally inadequate regarding amino acid composition.
• Strategic Divergence: The models employed distinct strategies. ChatGPT relied heavily on tofu (97.2% of recipes), a near-complete protein source, resulting in higher mean protein (38.5 g) and higher MPQS (118.8). In contrast, Claude utilized a more diverse array of ingredients, including nuts, seeds, and peas, but fell into the "classical complementarity trap" more frequently. Methionine was the dominant limiting amino acid, constraining 40.4% of Claude's recipes compared to 19.6% of ChatGPT's.
• Cultural and Format Variability: Meal quality was strongly dependent on cultural framing. Recipes inspired by Japanese, Indonesian, Korean, and Thai cuisines, as well as curries and grain dishes, achieved the highest MPQS. Conversely, Turkish, Spanish, and Caribbean cuisines, along with soups and stews, yielded the lowest scores. This indicates that culturally framed prompts lead to systematically uneven amino acid adequacy.

Significance and Claims
The paper asserts that while LLMs can reliably generate linguistically plausible and quantity-compliant meal plans, they fail silently on domain-specific tasks requiring quantitative scientific constraints. The study highlights that total protein is an incomplete indicator of adequacy, especially for older adults.

The authors claim this work demonstrates a "quality-versus-robustness" trade-off: ChatGPT's high adequacy relied on a narrow ingredient strategy (tofu) that may lack resilience to real-world constraints like allergies or sourcing, while Claude's diversity often resulted in nutritional inadequacy. The central conclusion is that plausibility must not be mistaken for nutritional adequacy. The paper proposes that for AI meal planning to be dependable, generation must be grounded in curated nutritional infrastructure, such as retrieval-augmented generation coupled with automated, meal-level quality checks like the MPQS.

Limitations
The authors note that the study evaluated single dinners rather than whole-day completeness, did not assess consumer acceptability, cost, or feasibility, and relied on a single food composition database. The findings represent a snapshot of two specific models under a minimal prompt.