Popular AI chatbots generate unsafe diet plans for teenagers

Many teenagers trying to manage their weight are turning to artificial intelligence chatbots to generate personalized diet plans. However, a new study published in Frontiers in Nutrition reveals that these computer-generated menus often fall short on basic calories and fail to provide the proper balance of essential nutrients. Following these flawed digital recommendations could put young people at risk of nutritional deficiencies during a period of rapid physical growth.

Childhood and adolescent obesity rates are rising rapidly across the globe. This condition carries immediate and long-term health risks, including a higher chance of developing type 2 diabetes, high blood pressure, and sleep apnea. Health experts typically recommend that young people manage their weight through a combination of increased physical activity, behavioral changes, and specialized nutrition therapy.

Dietitians are the ideal professionals to guide this process because they design sustainable meal patterns tailored to a growing body. A trained dietitian takes into account a teenager’s developmental needs, family habits, and the school food environment. Access to professional nutritional counseling, however, is often limited by high costs and heavy clinical workloads.

Because of these barriers, a notable portion of teenagers now rely on the internet for health advice. Artificial intelligence tools, which generate human-like text in response to user prompts, have become especially popular resources for dietary information. Prior research has explored how well these text-generating programs handle general nutrition questions or adult meal planning.

Few studies have tested how safely these tools can craft diets for teenagers. Adolescence is a time of immense physical and cognitive maturation, meaning young people have specific nutritional requirements that differ from those of adults. Ayşe Betül Bilen, an assistant professor of nutrition and dietetics at Istanbul Atlas University in Turkey, led a research team to investigate this issue.

Bilen and her colleagues wanted to know if popular chatbots could produce safe and balanced meal plans for young people without professional oversight. The research team noted that adolescents often type simple requests into search engines, rather than providing technical medical details. To see how the algorithms would respond to a typical teenager, the researchers decided to compare the software outputs against a professional standard.

The researchers set up a test using five widely available artificial intelligence models. The programs tested included OpenAI’s ChatGPT, Google’s Gemini, Microsoft’s Bing Chat, Anthropic’s Claude, and Perplexity. The team used the free versions of these tools because a typical teenager would be most likely to access the options without a paywall.

To simulate the varied needs of real adolescents, the team created four hypothetical profiles of fifteen-year-olds. These profiles represented a boy and a girl classified as overweight, along with a boy and a girl classified as obese. The researchers determined these categories using body mass index percentiles, a standard measurement that compares a child’s weight and height against growth charts for their specific age and sex.

For the purposes of this study, the eighty-fifth to ninety-fourth percentile was defined as overweight, while the ninety-fifth percentile and above was defined as obese. For each profile, the researchers prompted the five chatbots to generate a three-day weight loss nutrition plan. The instructions asked for three main meals and two snacks per day, using portion sizes in grams or milliliters.

The prompts specifically requested foods that are easy to find in Turkey, where the study was based. The researchers did not include target calorie counts or specific dietary guidelines in their prompts. They wanted to mimic the natural language a teenager might use when asking a computer for help.

Each software program was run in two separate sessions for the four profiles, resulting in sixty total days of computer-generated meals. Next, the researchers needed a benchmark to evaluate these digital menus. A pediatric dietitian created a one-day reference plan for each of the four adolescent profiles.

These human-made plans strictly followed international guidelines for energy balance and nutrient distribution. The research team used a specialized nutrition database to analyze the contents of both the computer-generated and human-made diet plans. They looked at the total calories provided, as well as the balance of macronutrients.

Macronutrients are the main building blocks of our diet, categorized as carbohydrates, proteins, and fats. The analysis revealed a massive calorie deficit in the computer-generated diets. On average, the algorithms underestimated the adolescents’ daily energy needs by nearly 700 calories when compared to the dietitian’s plan.

A deficit of this size is equivalent to skipping an entire meal, which the researchers noted could harm metabolic health and eating behaviors over time. “We show that diet plans generated by AI models tend to substantially underestimate total energy and key nutrient intake when compared to guideline-based plans prepared by a dietitian,” Bilen said in a press release.

The computer programs also struggled to provide the correct balance of macronutrients. According to established health guidelines, teenagers should get about forty-five to fifty percent of their daily calories from carbohydrates. In the algorithms’ meal plans, carbohydrates made up only thirty-two to thirty-six percent of the total energy.

This severe drop in carbohydrates often leads to a drop in dietary fiber, which is necessary for healthy digestion and maintaining a balanced gut microbiome. To make up for the lack of carbohydrates, the artificial intelligence models heavily relied on proteins and fats. The digital plans recommended that proteins provide up to twenty-four percent of daily energy, which exceeds the standard recommendation of fifteen to twenty percent.

A diet excessively high in protein can force the kidneys to work harder than normal and increase the amount of calcium lost through urine. Fats made up roughly forty-one to forty-five percent of the computer-generated calories, far above the ideal target of thirty to thirty-five percent. “AI-generated diet plans consistently deviated from the recommended macronutrient balance, which is particularly problematic for adolescents,” Bilen pointed out.

Bilen suggested that the software might be pulling information from trendy internet diets. Many popular weight-loss routines favor high-protein and low-carbohydrate meals. Because artificial intelligence programs learn by scanning vast amounts of online text, they often repeat these popular concepts rather than prioritizing age-appropriate medical guidelines.

“AI models are primarily trained to generate responses that appear plausible and user-friendly rather than clinically precise,” Bilen said. “Our findings suggest they may rely on generalized or popular diet patterns instead of fully integrating age-specific nutritional requirements.”

The researchers also checked the computer menus for micronutrients, which are the vitamins and minerals our bodies need in small amounts to function properly. The algorithms showed high inconsistency in this area. Some programs provided far too much vitamin D and folate, while others overestimated calcium and phosphorus.

Trace minerals like iron, zinc, copper, and manganese also varied wildly depending on which program the researchers tested. No single computer model consistently matched the human dietitian across all nutritional categories. The research team noted that teenagers are already at risk for vitamin and mineral deficiencies, and relying on an algorithm could worsen these issues.

The researchers acknowledged a few limitations to their work. Because the study relied on hypothetical profiles, it did not track real teenagers eating the computer-generated meals. This means the researchers could not observe actual metabolic changes or shifts in eating behaviors in human subjects.

The study did not incorporate the prior dietary habits of any participants, as that would require making assumptions about their daily routines. The prompts were also written in Turkish, which might limit how well the results apply to other languages or cultural contexts. Artificial intelligence software updates rapidly, so the specific errors found in these versions might change as the programs are refined over time.

The researchers noted that future studies should track real-life dietary habits alongside more advanced versions of the software. Health professionals advise that algorithm-based tools should not replace traditional nutritional counseling. While the technology might serve as a basic educational aid, it currently lacks the precision needed to manage adolescent health safely.

Teenagers and their parents should consult qualified medical experts before starting any restrictive weight-loss programs. “Adolescence is a critical period for physical growth, bone development, and cognitive maturation,” Bilen concluded. “Lower energy and carbohydrate intake, combined with increased protein and fat ratios, may pose risks during the adolescent growth period.”

The study, “Artificial intelligence diet plans underestimate nutrient intake compared to dietitians in adolescents,” was authored by Ayşe Betül Bilen, Gülen Ecem Kalkan, and Hülya Yılmaz Önal.