Turkish researchers have confirmed that AI tools can create educational nutrition plans or support diet planning. However, these cannot yet replace professional dieticians, underscoring the need for human oversight to ensure personalized care and safety.
The study is especially relevant as AI is not just being used privately. The US government has recently integrated former Trump administration adviser Elon Musk’s Grok into guidance on the country’s new dietary guidelines. FDA deputy commissioner Kyle Diamantas says the tool aims to “provide parents and consumers with clear and concise answers.” However, several reports question its trustworthiness and usefulness in advising people and creating healthy diet plans.
The study researchers tested ChatGPT-4o, DeepSeek V3, and Grok-3 by prompting them to design menus for 40 simulated patients recovering from sleeve gastrectomy (post-SG), a common weight-loss surgery.
Dieticians reviewed the AI-generated menus to test whether they met professional guidelines, based on a checklist. The team aimed to learn whether AI models could develop diet plans that matched professional recommendations during different stages of recovery, such as liquids, puréed foods, and solids.
The analysis revealed that ChatGPT was the most consistent in aligning with the guidelines, especially for early recovery stages.
DeepSeek was found to include more vitamins and minerals in recommendations, and Grok often suggested too little energy and protein.
The Nutrients study points out that all three missed important details, such as expert advice on vitamin B1, iron, and multivitamins, which are essential to prevent complications after bariatric surgery.
Researchers tested how three AI systems handled post-surgery diet planning after sleeve gastrectomy.To learn more, Nutrition Insight speaks with study authors Dr. Emre Batuhan Kenger and Dr. Tugce Ozlu Karahan, who specialize in AI applications in health sciences at Istanbul Bilgi University, and Aylin Bolat Yilmaz, Ph.D. student at Acibadem University and a clinical dietitian.
Grok consistently underdelivered on energy and protein across all phases. What does this tell us about its approach?
Kenger: There is no universally defined energy target in current guidelines for the postoperative bariatric period; instead, this process is largely guided by the clinical assessment of an experienced bariatric dietitian, the patient’s individual tolerance, and close follow-up.
In our study, Grok generated noticeably lower energy and protein values compared with the other models; however, this finding is limited to the specific model versions and testing conditions used. Large language models are continuously evolving, and greater integration of guideline-based frameworks may modify these differences over time. Therefore, our results should be interpreted not as definitive conclusions but rather as a snapshot of current model behavior.
All three AIs mostly skipped thiamine, iron, and multivitamin advice. Why might they overlook important nutrients in post-SG care?
Karahan: It is noteworthy that large language models tend to generate general nutrition advice rather than guideline-based clinical protocols. Since micronutrients such as thiamine and iron are commonly included in standard multivitamin preparations, models may address them under a general “multivitamin use” heading rather than highlighting them individually.
However, given the risk of postoperative complications following bariatric surgery, individualized and specific recommendations — particularly for critical micronutrients like thiamine and iron — are clinically essential. For this reason, postoperative nutrition management should be based not only on general supplementation but also on patient-specific biochemical monitoring and targeted micronutrient strategies.
ChatGPT did best on liquid-phase guidance, but none handled complications well. Could better prompts fix this, or will human oversight always be needed?
Yilmaz: In this study, model outputs were evaluated using standardized single inputs; in real clinical practice, however, AI systems are typically used within interactive workflows guided by specialists and refined through continuous feedback. With such structured and optimized communication, more refined results — particularly regarding protein targets and phase-specific dietary guidance — can be achieved.
Clinical experts emphasize that AI tools can support — but not replace — dietitians in postoperative care.Nevertheless, postoperative complications after bariatric surgery vary considerably between individuals and cannot be managed through generalized nutrition advice alone. Clinical follow-up requires personalized calculations based on patient tolerance, metabolic status, and biochemical parameters. Therefore, AI should be considered a supportive tool, while postoperative nutrition planning must be individualized and conducted under the supervision of a dietitian experienced in bariatric care.
Models evolve fast — what updates might change these rankings, and how should dietitians adapt?
Kenger: AI-generated outputs are directly influenced by technical factors such as model updates, sampling parameters — including “temperature,” which affects response variability — and hallucination tendencies. These variables may lead to different nutritional recommendations even when identical clinical inputs are provided. In the future, more structured integration of guideline-based datasets and the development of clinical validation layers may help reduce such inconsistencies and improve the reliability of outputs.
During this transition, the role of dietitians becomes increasingly important. AI-derived information must be interpreted alongside clinical experience and patient-specific parameters. Professional judgment informed by patient tolerance, biochemical findings, and individual risk profiles enables the safe and effective use of digital tools. Dietitians will be able to adapt to this technological transformation in a sustainable and meaningful way to the extent that they integrate AI outputs with their own clinical reasoning.
