AI Chatbots Cut 695 Calories Daily From Teen Diet Plans

Facing a teenager who wants to lose weight with the help of a chatbot, parents now have a peer-reviewed number to point to. A March 2026 study in Frontiers in Nutrition compared 60 three-day diet plans produced by ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT, and Perplexity against reference plans drafted by a registered dietitian for four 15-year-old profiles. The AI plans came in short on calories by an average of 695 kcal a day, roughly the energy of a missed meal, and shifted the protein and fat shares above the recommended range while compressing carbohydrates well below it.

No chatbot closed the gap on energy. None of the five held a consistent macronutrient profile within the adolescents’ reference range, and the gaps showed up on every nutrient the researchers measured.

Where the Five Chatbots Came Up Short on Calories

The team used the free version of each AI tool, the variants a curious teen is most likely to open in a browser tab. Each model was fed the same prompts covering age, height, weight, and a weight-loss goal, and asked to draft three days of meals: breakfast, lunch, dinner, and two snacks a day. The four target profiles were 15-year-olds in two body types, a boy and a girl in the overweight percentile and a boy and a girl in the obese percentile.

Across 60 plans the AI models undercalculated total energy by an average of 695 kcal per day. They also undercounted 19.9 g of protein, 15.8 g of fat, and 114.6 g of carbohydrates against the dietitian’s targets. The calorie gap is large enough to have clinical consequences in a still-growing body, the paper notes.

The bias ran in one direction across every model. That uniformity matters. A teen who cycles through two chatbots for a second opinion is unlikely to find one that contradicts the first.

• 5 chatbots tested: ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT, Perplexity
• 60 three-day diet plans evaluated
• 4 standardized 15-year-old profiles
• Mean calorie underestimation: 695 kcal per day
• Paper published March 12, 2026 in Frontiers in Nutrition

How the Macronutrient Mix Tilted Off Target

Calories were not the only line the chatbots overshot. The protein share of each AI-generated plan ran between 21.5% and 23.7% of total energy, above the 15% to 20% window the U.S. National Academies of Sciences, Engineering, and Medicine recommends for adolescents. Fat ran between 41.5% and 44.5%, above the 30% to 35% range. Carbohydrates, the energy source adolescents most need for growth, sat between 32.4% and 36.3%, below the recommended 45% to 50%.

The averages mask wide spread between models. Micronutrient content moved significantly from one chatbot to the next, and no model showed consistent proximity to the dietitian across every nutrient tested. 19.9 g of protein per day, the average protein gap, looks small on a plate. Over months of dieting it is not.

Why a Plausible Answer Is Not a Safe One

The study frames the failure as an incentives problem. The chatbots are tuned to produce text that sounds confident and reads as helpful, not to match the precise nutrient targets a registered dietitian would apply to a 15-year-old in a growth spurt.

AI models are primarily trained to generate responses that appear plausible and user-friendly rather than clinically precise.

That quote is from Dr Ayşe Betül Bilen, an assistant professor at the Faculty of Health Sciences at Istanbul Atlas University and the paper’s lead author. She and her co-authors point to training data of variable quality: international guidelines from reputable scientific societies are freely available online, but the models also absorb popular content with thinner scientific support, including strongly low-carbohydrate patterns. Add a tendency to orient answers toward user expectations, and the resulting plan reads right on the screen while missing what a growing body actually needs.

What a Vertically Trained Health Model Looks Like

The paper’s authors and outside commentators point to one structural fix: train health chatbots on a narrow, vetted corpus rather than the open web. The most cited prototype in that direction has been the World Health Organization’s S.A.R.A.H., the Smart AI Resource Assistant for Health, which the WHO describes on its S.A.R.A.H. campaign page as a research prototype built to explore how generative AI can improve public health outreach.

S.A.R.A.H. ran around the clock and operated in eight languages. It answered questions on healthy habits, stress reduction, quitting tobacco, and nutrition, and was built on validated WHO materials, not the open internet. The aim was to put public-health content ahead of the diet-blog and forum content that drove the gaps in the Turkish study.

The WHO Director-General, Dr Tedros Adhanom Ghebreyesus, called S.A.R.A.H. a glimpse of how AI could expand access to health information in a more interactive way. The program is no longer in active public deployment, and WHO has folded its lessons into wider research on safe AI deployments. The S.A.R.A.H. approach is vertical integration: one objective, one vetted source set, and one validation pipeline aimed at a defined outcome.

Where AI Coaching Has Held Up

The story is not all negative. Among adults with prediabetes and overweight or obesity, an automated AI-led lifestyle program was noninferior to a human coach-led Diabetes Prevention Program in a 12-month pragmatic randomized trial published in JAMA. Among 368 participants, 31.7% in the AI arm and 31.9% in the human arm met the composite outcome of at least 5% weight loss, hemoglobin A1c targets below 6.5%, or at least 150 minutes of weekly physical activity, and the difference met the prespecified 15% noninferiority margin.

The DPP trial used a reinforcement-learning algorithm, not a large language model, and was scoped to a single objective: hit the standard DPP targets. The diet content was standardized, the metric predefined, and the patient pool an adult population the algorithm had been calibrated against. A separate 2023 systematic review and meta-analysis of broader chatbot interventions also reported measurable effects: 735 steps a day more on average, about one extra serving of fruits and vegetables, and roughly 45 minutes more sleep per night.

The pattern is consistent across both studies. AI health tools work best when narrowly built, validated against a defined outcome, and used with a defined population. They underperform when asked to substitute for clinical judgment in an unfamiliar demographic, the exact case the Istanbul paper makes for adolescents. The randomized DPP trial comparing AI and human coaches is a useful counterweight, and it is a different kind of tool from a chatbot.

The Margin for Error at 15

The Turkish study is part of a wider pattern in which LLM-based health advice improves when the population, the metric, and the training data are tightly defined, and slips when any of those pieces are loose. Teenagers land on the loose end of all three.

Adolescent overweight and obesity is a fast-growing public-health problem. The WHO put the number of overweight 5-to-19-year-olds at roughly 390 million in 2022, including about 160 million who met the threshold for obesity. In a Bulgarian cross-sectional survey of 315 high-school students aged 15 to 19, 31.4% said they got nutrition information from the internet, including ChatGPT, against 10.5% who had consulted a family doctor. For that population the chatbot is the default channel.

The paper suggests parents and clinicians treat AI diet plans for adolescents with caution, and as a complement at most. The exact macronutrient figures matter because of who is on the receiving end. A 15-year-old body does not absorb a standard adult intake, and the chatbots in this study did not adjust for that.

Bilen’s group is extending the work to other age brackets and clinical scenarios, and the DPP researchers are following their adult cohort for longer outcomes. Each next paper is likely to tighten the gap. For now, the rule the March study surfaces already applies: a chatbot optimized for plausibility gives a teen a confident answer, and that answer can be wrong by hundreds of calories a day.

Frequently Asked Questions

What did the Frontiers in Nutrition study measure?

The study compared 60 three-day diet plans generated by five AI chatbots against reference plans built by a registered dietitian for four standardized 15-year-old profiles. Each plan was analyzed for energy, macronutrients, and selected micronutrients using BeBiS software, with single-sample t-tests, Bland-Altman agreement analysis, and heatmap visualization for micronutrient variability.

Which AI chatbots were tested, and which versions?

The researchers used ChatGPT-4o from OpenAI, Gemini 2.5 Pro from Google DeepMind, Claude 4.1 from Anthropic, Bing Chat-5GPT from Microsoft, and Perplexity. All five were used in their free versions, the variants most accessible to a teenage user.

How do the AI macros differ from the official adolescent ranges?

Across the 60 plans, protein ran 21.5% to 23.7% of total energy against a target window of 15% to 20%. Fat ran 41.5% to 44.5% against 30% to 35%. Carbohydrates ran 32.4% to 36.3% against 45% to 50%. The reference windows come from the U.S. National Academies of Sciences, Engineering, and Medicine.

Are AI-generated diet plans unsafe for teenagers?

The study described the deviations as clinically significant because all five models underestimated total energy and skewed the macronutrient ratio in ways that diverge from adolescent guidelines. Whether a specific plan causes short-term harm depends on what else the teen eats and for how long the plan is followed. The research did not track health outcomes in real teen users; it compared the nutritional content of the plans themselves.

What should a parent or teenager do instead?

Treat any chatbot-generated plan as a starting point for questions, not a prescription. Lead author Dr Ayşe Betül Bilen recommends using AI as a complement to professional counseling rather than a substitute. Where possible, verify a plan against published guidelines such as the WHO Adolescent Nutritional Guidelines, or have it reviewed by a registered dietitian familiar with adolescent growth.

Disclaimer: This article is for general informational purposes and does not constitute medical advice. AI-generated diet plans carry documented risks for adolescents and other groups with specific nutritional needs. Consult a qualified healthcare professional before making changes to diet, exercise, or medication. Figures cited are accurate as of the publication date of the underlying studies.