AI Meal Plans for Teens Undercount Calories by Nearly 700 Per Day

Frontiers

A 15-year-old trying to lose weight asks an AI chatbot for a meal plan. The chatbot responds with something that looks reasonable - three meals, two snacks, specific foods and portions. But compared to what a registered dietitian would recommend for the same teenager, the AI plan contains nearly 700 fewer calories per day. That deficit is roughly the size of an entire meal.

This is not a hypothetical scenario. A new study published in Frontiers in Nutrition tested exactly this, and the results raise pointed questions about relying on AI tools for nutritional guidance during adolescence - a period when getting enough calories and the right balance of nutrients is not optional.

Five chatbots, four fictional teenagers

Researchers at Istanbul Atlas University used free versions of five AI models - ChatGPT 4, Gemini 2.5 Pro, Bing Chat-5GPT, Claude 4.1, and Perplexity - to generate three-day meal plans. Each plan covered three meals and two snacks per day. The prompts specified age, height, and weight for four fictional 15-year-olds: a boy and girl in the overweight percentile, and a boy and girl in the obese percentile.

The AI-generated plans were then compared against meal plans created by a dietitian specializing in adolescent health, following established nutritional guidelines.

The calorie gap and the macronutrient distortion

On average, the AI models calculated energy requirements almost 700 calories lower than the dietitian did. But the problem was not simply that the plans were too restrictive. The macronutrient balance was also substantially distorted.

AI plans recommended about 20 grams more protein than the dietitian's plans, pushing protein to 21-24% of total energy intake. Lipid (fat) intake was also elevated, making up 41-45% of energy. Carbohydrates, meanwhile, were dramatically reduced - an average difference of about 115 grams, leaving carbs at only 32-36% of energy intake.

For context, the U.S.-based National Academies of Sciences, Engineering, and Medicine recommend that for adolescents, 45-50% of calories should come from carbohydrates, 30-35% from fats, and 15-20% from proteins. The AI plans essentially inverted the carbohydrate-to-fat ratio.

Plausibility over precision

Why do the models get it wrong? Lead author Ayse Betul Bilen, an assistant professor at Istanbul Atlas University, offered a straightforward explanation: AI models are primarily trained to generate responses that appear plausible and user-friendly rather than clinically precise. The models may rely on generalized or popular diet patterns - think high-protein, low-carb trends - rather than integrating age-specific nutritional requirements from established guidelines.

This is a fundamental design characteristic, not a bug that will be fixed with the next update. Large language models optimize for coherent, convincing output. They are not nutritional databases, and they do not consult clinical guidelines the way a trained dietitian does. The resulting meal plans look professional enough to inspire confidence, which may be precisely the problem.

What is at stake for growing bodies

Adolescence is a period of rapid physical growth, bone development, and cognitive maturation. Insufficient calorie intake during these years can impair growth, weaken bone density, and affect brain development. Macronutrient imbalances carry their own risks: excessive protein and fat with insufficient carbohydrates can alter metabolic pathways and may contribute to disordered eating patterns.

Bilen pointed out that following such unbalanced or overly restrictive meal plans during the teenage years may negatively affect growth, metabolic health, and eating behaviors.

The concern is amplified by the context in which these tools are used. Teenagers dealing with weight issues often lack access to professional dietary counseling. They may turn to free AI chatbots precisely because they cannot afford or reach a dietitian. The tools are most likely to be used by the people who can least afford to get bad advice.

A study with its own limitations

Several caveats are important. The researchers used free versions of the AI models, which may perform differently than paid or professional-tier versions. The study was conducted at a single point in time, and AI models are updated frequently - performance may have shifted since the analysis was completed.

The comparison was against a single dietitian's recommendations. While those recommendations followed established guidelines, different dietitians might produce somewhat different plans for the same teenagers, introducing variability that the study design does not capture.

The study also did not test whether anyone actually followed the AI-generated plans or experienced health consequences. It is a comparison of outputs, not outcomes. Real-world harm from AI meal plans remains plausible but unquantified.

And four fictional patient profiles, while useful for a controlled comparison, cannot represent the full diversity of adolescent body types, metabolic conditions, and cultural dietary patterns.

Tools, not replacements

The researchers were careful to note that AI models should function as complementary aids in nutrition education, not replacements for professional dietary counseling - particularly for vulnerable populations. That framing is sensible but may be optimistic. When a free, instant, confident-sounding chatbot is available and a trained dietitian is not, the distinction between tool and replacement dissolves quickly.

For now, the clearest takeaway is specific: if a teenager asks an AI chatbot for a weight-loss meal plan, the result is likely to recommend too few calories, too little carbohydrate, and too much fat compared to what a professional would advise. That gap - nearly 700 calories a day - is not subtle. It is the difference between a restrictive but adequate diet and one that could meaningfully impair growth.