All participants gave written informed consent. Ethical approval was not required as the study had low risk to participants.

The study uses a case study approach to achieve the research objectives stated in the earlier section. It provides rich and detailed data that can be used to gain a deep understanding of a particular case. It allows for the exploration of complex phenomena that cannot be easily studied through other research methods. ¹⁴ Although there are limitations to the case study method it is one of the most useful tools in the exploratory study of abstract and evolving phenomena. The type of case study method utilized in this study is Illustrative case studies. ¹⁵ The approach used in this study is borrowed from functional testing and quality Assurance practices in software development. Functional testing involves creating test cases based on the functional requirements of the software application. These test cases are designed to evaluate whether the software performs as expected. Functional testing is typically performed using black box testing techniques, which means that the tester does not have access to the source code of the software application. In this case, ChatGPT acts as a black box for the researchers involved in this study.

To evaluate the performance of ChatGPT in medical nutrition therapy a well-defined Study Protocol was used. The steps followed in the study as follows:

Step 1: Creation of questions (scenarios) of varied complexity by public health professionals. The questions were selected by the licensed medial nutrition therapist working in UAE. The selected scenario was simple diet consultation to patient with comorbid conditions.

Step 2: The response of ChatGPT was taken and recorded for further analysis.

Step 3: The responses from Step 2 were evaluated by a registered nutritionist for accuracy.

Step 4: Based on the accuracy the potential of harm score for the response was created.

Step 5: Data was summarized and analyzed by the expert group used in Step 1.

The expert group for deciding complexity contained five public health professionals working in the United Arab Emirates. The experts were selected from Gulf Medical University, UAE and method of selection was nonrandom purposive sampling. The inclusion criteria for the expert were master’s degree and clinical experience greater than five years. The researchers involved in this study approached 7 healthcare professionals out of which five agreed to be part of the expert group. The researchers wanted to recruit five to nine experts as a number greater than that if difficult to handle and a number less than that may result in bias.

For accuracy, one registered nutritionist’s response was taken for step 3. The nutritionist gave a score on a ordinal scale of one to ten where one being least and ten being most accurate.

To ascertain the potential of harm in Step 4 all five experts discussed earlier worked together.

The method utilized for reaching consensus was the Delphi method depicted in Figure 1. ¹⁶ Using the steps mentioned above and data provided in the support material the reproducibility of the research can be established. Again, a scale of one to ten was used to ascertain the potential to harm where one being least and ten being highest.

The Delphi method is a structured communication technique originally developed as a systematic, interactive forecasting method that relies on a panel of experts. The experts answered questionnaires in three rounds. After each round, a researcher VM provides an anonymous summary of the from the previous round as well as the reasons they provided for their judgments. Thus, the experts are encouraged to revise their earlier answers considering the replies of other members of their panel. It was observed that the range of the answers decreased, and the group converged. Finally, the process is stopped after a predefined stop and the median scores of the final rounds determined the results ( Figure 1).

The conceptual definitions of the terms used in the study are as follows:

Clinical Accuracy: “A clinical accuracy is a qualitative approach that describes the clinical outcome of basing a treatment decision on the result of a measurement method being evaluated”. ¹⁷

Complexity of the Clinical Problem: “Clinical complexity is a protean term encompassing multiple levels and domains. Illustratively, a prominent concern in health care involves a multiplicity of disorders and conditions experienced by a person along with their cross-sectional and longitudinal contexts”. ¹⁸

Potential for Harm: “Harm means an injury to the rights, safety or welfare of a research participant that may include physical, psychological, social, financial or economic factors”. ¹⁹

The question is simple, with age, gender, and a weight loss goal provided. The statement emphasizes the importance of sustainable weight loss and the potential risks of rapid weight loss. The provided diet chart is low carb, high fiber/protein, suitable for the given condition. There is a negligible risk for a user following this diet unless they have comorbid conditions. The statement that a diet chart need not contain caloric information is not true, as it serves as a guideline for achieving a caloric deficit to aid in weight loss. In terms of the evaluation criteria, the statement receives a complexity score of 2, an accuracy score of 8, and a potential risk score of 1. This result suggests that in case of lower complexity the accuracy of the information is higher and potential risk is lower.

The question is slightly more complex than the previous one, with the addition of BMI information. For a person with a BMI of 34, a calorie-deficit diet is required for weight loss. The diet, however, does not specify the amount of oil to be consumed, which can significantly increase the calorie count. The diet is not specific, and portions are assumed, which may result in a diet of around 1400-1500 calories, which may not be enough to achieve the target weight loss. A layperson following this guide may not achieve their weight loss target as the diet provided is not guided. In terms of the evaluation criteria, the statement receives a complexity score of 3, an accuracy score of 7, and a potential risk score of 3. This result again supports the finding of the case 1 as increase in the complexity score reduces the accuracy while increasing the potential to harm.

The complexity of the question increases with the addition of the condition of PCOS. The diet provided is like the previous question with the addition of extra guidelines for PCOS, which is general information. However, the diet provided is not specific to the condition, and a user following it may not achieve their weight loss target, but they are not at potential risk for harm. In terms of the evaluation criteria, the statement receives a complexity score of 4, an accuracy score of 6, and a potential risk score of 4. The result of this case is also concurring the hypothesis complexity to question asked results in less accuracy and higher risk to harm.

The question is complex due to the mention of diabetes, which requires consideration of many factors before preparing a diet chart. A simple statement of diabetes does not provide enough information, and the patient should be asked about the type of diabetes, medications, and recent blood reports. Calories, BMI, and current physical activity are critical considerations for a diabetic diet. The patient is at risk of developing hypoglycemia if they are on insulin and have a low BMI or high activity levels. A dietitian would consider all these factors while preparing a plan for a diabetic patient. In terms of the evaluation criteria, the statement receives a complexity score of 4, an accuracy score of 5, and a potential risk score of 6. The complexity score of four for an older patient has less accuracy and high potential of harm. Does age contribute to potential to harm? This question needs to be further tested empirically.

The complexity of the question increases with the addition of chronic kidney disease (CKD), which requires consideration of several factors while preparing a diet chart, such as the stage of CKD and the current level of potassium and sodium in the blood. However, the statement receives a low accuracy score of 4 as the diet generated does not mention limiting the sodium intake to at least 1.5 g/day, which is essential for CKD patients. Additionally, the diet contains high sources of protein, 75-80 g, which is much higher than what is recommended for a CKD patient and not calculated as per patient weight and CKD level. As a result, a layperson following this diet may be at a high potential for risk, indicating a high potential for risk score of 8. In terms of the evaluation criteria, the statement receives a complexity score of 5, an accuracy score of 4, and a potential risk score of 8. Increasing complexity of the query from 4 to 5 increases the potential of risk from 6 to 8. This makes us conclude that with increasing complexity the increase in potential harm increases exponentially after a point. This phenomenon needs to be further tested empirically.

The complexity of the question increases with the addition of hypertension as a comorbidity. However, the diet chart provided is the same as the previous question, which does not pose much risk for diabetes and hypertension but poses all the risks previously mentioned for CKD. Therefore, the statement receives a low accuracy score of 4. Additionally, patients need to be educated about sugar and salt sources, and general guidelines are not enough. Measurements should be incorporated into the diet plan itself to avoid potential risks. As a result, the statement receives a high potential for risk score of 8. In terms of the evaluation criteria, the statement receives a complexity score of 6, an accuracy score of 4, and a potential risk score of 8. The finding of this study doesn’t concurs the finding of the case 5 that increase in potential harm with increase in exponential as increasing complexity potential risk remains same.

The complexity of the question increases with the addition of gluten sensitivity. As a result, a proper dietitian is required to prepare a diet plan that takes into consideration the patient’s multiple comorbidities and dietary restrictions. However, the statement is lacking in accuracy as it does not provide any specific information on how to prepare a diet plan for a person with these conditions. Therefore, the accuracy score is low at 4. Additionally, without a specific diet plan, there is potential for risk for the patient with so many comorbidities and dietary restrictions. As a result, the statement receives a high potential for risk score of 8. In terms of the evaluation criteria, the statement receives a complexity score of 7, an accuracy score of 4, and a potential risk score of 8. The findings from this case also concurs the finding from the earlier cases. The increase in complexity is inversely proportional to accuracy while directly proportional to the potential risk.

This question is extraordinarily complex with multiple parameters given, including the condition of hypothyroidism. The ideal diet for this patient is a low-carb, high-protein, anti-inflammatory diet, with the need to avoid goitrogenic foods like soy products. However, the accuracy of the given information is relatively low, and there is still a potential risk for the patient if not properly guided by a qualified dietitian. Overall scores are as follows: Complexity - 8, Accuracy - 3, Potential for Risk - 6. This study examines the case of highest complexity and hence minimum accuracy. The risk score for this study was expected to be highest but that is not the case. This finding does not concur the findings from the earlier seven cases.

The summary of the analysis of eight cases is listed in Table 1.

As depicted in Figure 2, the complexity of the scenario increases risk potential also increases. That suggests that ChatGPT should be avoided for complex medical conditions/scenarios. Researchers believe accuracy does not change much with an increase in complexity and needs to be further evaluated empirically.

The findings of the study are supported by the researcher Johnson et al. (2023). They observed that ChatGPT can produce accurate information to diverse medical queries as judged by academic physician specialists although with important limitations. Further research and model development are needed to correct inaccuracies and for validation. ²⁰ Another group of researchers found that ChatGPT provides medical information of comparable quality to available static internet information. ²¹ Another recent study suggests cautious approach against use of the ChatGPT in clinical practice. They lament that it doesn’t provide references for the information hence is not reliable for clinical use. Thus, the findings of this study also suggest the cautious use of ChatGPT in medical nutrition therapy as irresponsible use has potential harm for the user. The study assessing the accuracy and potential risks of using nutrition therapy information provided by ChatGPT, evaluated by nutritionists and a group of experts, has several limitations that must be considered when interpreting its results. ChatGPT’s responses are based on the information available up to its last training data, which might not include the latest research or updated guidelines in nutrition therapy. This time-lag in information can introduce a bias towards outdated practices or missing new evidence-based approaches. The accuracy and risk assessments made by the nutritionists and experts are subjective and can vary based on their individual experiences, knowledge, and biases. This variability can introduce both direction and magnitude biases in the evaluation process. The experts and nutritionists might have preconceived notions about the reliability of AI-generated information, which could influence their assessment of ChatGPT’s responses, either positively or negatively. The range and type of nutrition therapy questions asked may not comprehensively cover the vast field of nutrition. Thus, the study’s findings might not be generalizable across all areas of nutrition therapy.

The study used a small sample size which could affect the accuracy of the results. Another limitation is the dynamic nature of technology. Since technology is constantly evolving and improving, the results of the study may need to be reevaluated after a few days or weeks to account for any changes or updates. Additionally, the study’s reliance on only one nutritionist to assess accuracy introduces the possibility of bias and human errors.