New research reveals that popular AI tools may be giving adolescents incomplete and imbalanced diet advice, raising important questions about whether these technologies are ready to guide growing bodies without expert oversight.
Study: Artificial intelligence diet plans underestimate nutrient intake compared to dietitians in adolescents. Image credit: ilona.shorokhova/Shutterstock.com
Artificial intelligence (AI) is increasingly used for dietary planning among adolescents, but a new study suggests it may fall short of expectations. The investigation, published in Frontiers in Nutrition, found that AI-backed recommendations may consistently underestimate the nutritional intake required for adolescents.
Rising adolescent obesity drives demand for accessible diet advice
Adolescent overweight and obesity rates are rapidly increasing globally, affecting about 390 million adolescents in 2022. In fact, multiple regions now report this as the major form of malnutrition. Excessive body weight is associated with multiple adverse health outcomes, including type 2 diabetes, abnormalities in blood cholesterol, high blood pressure, and sleep apnea. These youngsters are also more likely to be obese as adults and have a lower quality of life.
Adolescents are also prone to body image worries and desires to lose weight, including potentially dangerous methods like vomiting after meals or overusing laxatives.
Dietary modification is key to improving child health in this area. Dietitians are health professionals who design and supervise individualized nutrition plans in accordance with established guidelines. However, their services are not always accessible, and their heavy workload may prevent adolescents from receiving the dietary advice and follow-up they need.
AI-based tools, such as chatbots, are being used to overcome these limitations, but only a handful of studies have evaluated their role in adolescent nutrition. Similarly, large language models (LLMs) like ChatGPT can provide useful information to support nutritional planning, but with important limitations.
Existing research indicates they may not meet safety standards or international nutritional recommendations, especially under real-world conditions. AI tools are also unlikely to provide the same level of tailored patient services that dietitians do. However, most of this evidence is based on adult studies or on clinical cases.
The current study sought to directly compare AI-generated diets with individualized dietitian-prepared diets for overweight or obese adolescents. The areas of comparison were energy and nutrient content, safety, and feasibility. The comparison could show whether AI chatbots can substitute for dietitians in nutritional planning for this patient category or be used as aids under dietitian supervision.
Researchers compare five AI tools with dietitian plans
The researchers used five AI models (ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT, and Perplexity) to generate 60 diet plans over two sessions. Three-day diet plans were created by each model in response to prompts using four standardized adolescent profiles: a boy overweight or obese, and a girl overweight or obese.
These were compared with a reference one-day dietary plan prepared by a dietitian for each profile. This followed the nutritional recommendations with energy distribution as follows: 45–50 % from carbohydrates, 30–35 % from lipid and 15–20 % from protein.
The researchers then analyzed the energy and macronutrient content of each plan.
AI diets underestimate energy and key nutrients
The results revealed a consistent and potentially concerning pattern. AI models included less energy and macronutrients than dietitians did in their plan. The energy shortfall was 695 kcal, while protein was 20 g short, fats decreased by 16 g, and carbohydrate by 115 g. The potential energy gap may have important clinical implications, especially given adolescents' high energy demands.
The authors suggest that, given this typical oversupply of fat and lower carbohydrate content, LLMs may rely more on popular diets like the ketogenic diet than on scientific guidelines, which explains the low-carbohydrate, high-fat approach. This could upset growth, metabolism, and cognitive development in this crucial developmental window. The long-term safety of such recommendations is thus unproven.
The five models recommended protein content up to 23.7 %, and fat content up to 44.5 %. Both were above the recommended levels for adolescents. In contrast, carbohydrates accounted for at most 36.3 % of the diet, which was lower than the recommended level.
Dietitian plans contained 44 %-46 % carbohydrate depending on the profile. Protein percentage varied between 18 % and 20 %, and fat between 36 % and 37 %. Overall, these plans aligned with the national recommendations.
The authors point out that “This pattern illustrates a systematic shift across all AI models to lower CHO, higher protein, and higher lipid meal structures, indicating that the macronutrient balance, not just the amount of gram-based nutrients, is significantly disrupted in AI-generated plans.”
Micronutrient composition varied significantly across AI-generated diets, with notable variability between models and compared to the dietitian reference. This could contribute to micronutrient inadequacies in adolescents, indicating that these plans may not yet be suitable for clinical use without professional supervision. No model adhered closely to the dietitian reference diet across all nutrients.
The authors note that this is the first time different LLMs have been compared for nutritional needs in adolescents, with a detailed evaluation of macro- and multiple micronutrients, as well as macronutrients. As earlier research suggests, this may indicate AI’s lack of technical expertise in this area. This may hinder the accurate estimation of energy and macronutrient composition in an AI-generated personalized dietary plan.
Strengths and limitations
The study has several strengths. It evaluated five different AI models, enhancing the robustness and comparative power of the analysis. By generating three-day diet plans, the researchers were able to assess consistent patterns rather than isolated anomalies, strengthening the reliability of the findings. The use of dietitian-designed plans grounded in international dietary guidelines provided a credible and clinically relevant reference standard. In addition, the comprehensive assessment of macro- and micronutrients enabled a detailed, multidimensional evaluation of dietary quality.
Despite these strengths, the study also has limitations. Its findings may apply only to the specific AI models tested, which are continuously evolving, and some potentially relevant information may have been absent from the standardized adolescent profiles, limiting personalization. The statistical approach, including the use of averaged multi-day outputs, may affect the independence of results and variability estimates. Furthermore, the study relied on simulated scenarios rather than real-world adolescent behaviors, which may limit ecological validity. Finally, the use of standardized prompts in a single language could restrict the generalizability of the findings to other populations and settings.
The risks of unsupervised AI nutrition advice
“AI models have exhibited clinically significant deviations in diet plans for adolescents at both macro and micro levels.” They consistently recommended diets with lower energy and carbohydrate content than the dietitian-designed diet.
Until these gaps are addressed, the authors caution that AI-generated diet plans should not replace professional dietary guidance for adolescents.
Download your PDF copy by clicking here.
Study finds most sustainable diet indices overlook industrial food processing