Do AI Chatbots Give Nutrition Advice Aligned with the Dietary Guidelines for the Brazilian Population? An Exploratory Analysis Informed by the Commercial Determinants of Health

1. Introduction

The rapid expansion of digital technologies has transformed how individuals access health information, including guidance on diet and nutrition. Artificial intelligence (AI)-based conversational systems (hereafter referred to as AI chatbots) can be used by the population to obtain information on health behaviors, including food choices and dietary patterns [1,2]. Exposure to health information can influence knowledge, perceptions of risk, and behavioral intentions related to diet [3]. As a result, AI chatbots may represent an increasingly important channel through which individuals encounter nutrition guidance; at the same time, limited evidence exists regarding the quality, reliability, or evidentiary basis of the nutrition information they provide.

Nutrition and health information available to the population is produced by multiple actors, including public health institutions, researchers, media outlets, social media, and private-sector organizations. In the context of food and nutrition, these messages can include advertising, nutrition and health claims, promotions, and other forms of food-related communication that do not always or necessarily prioritize public health [4]. These dynamics have been described within the framework of the commercial determinants of health, which examines systems, practices, and pathways through which commercial actors drive health and equity [5]. Within this evolving information environment, emerging digital technologies may further reshape how nutrition and health information is generated, disseminated, and interpreted by the population.

Transnational food and beverage companies continue to influence public understandings of diet and health through diverse communication strategies across digital platforms [6]. At the same time, AI tools are becoming more widely used to generate and disseminate health information on a scale. The growing presence of AI-generated nutrition and health content introduces new questions regarding the accuracy, framing, sources, and potential commercial influences embedded in information communicated to the public [7].

AI tools and AI chatbots are often based on Large Language Models (LLMs), which are artificial intelligence systems trained on extensive text datasets to generate human-like responses to user prompts. These systems synthesize information from a wide range of digital sources, including scientific publications, health websites, and other publicly available online materials. Although AI chatbots may facilitate access to health information, the characteristics and quality of the content they generate remain uncertain. In particular, the lack of transparency regarding the origin of specific statements and the difficulty in reconstructing algorithmic decision-making processes have led experts to describe these systems as “black boxes”[8] .Recent studies evaluating AI-generated health information suggest that chatbot responses can vary in their accuracy, completeness, and alignment with evidence-based recommendations [1,2,9]. However, their responses to nutrition-related questions have rarely been evaluated against national dietary guidelines and, to our knowledge, never in relation to the commercial determinants of health.

In the field of nutrition and health, recommendations provided to the public are expected to align with evidence-based guidance developed by public health institutions and health organizations. In many countries, national dietary guidelines play a central role in translating scientific evidence into recommendations that support healthy eating practices [10]. In Brazil, the Dietary Guidelines for the Brazilian Population (DGBP) is centered on a “golden rule” that prioritizes the consumption of fresh and minimally processed foods while discouraging the intake of ultra-processed products [11,12]. DGBP also recognizes access to reliable information as essential for supporting healthy food choices [11]. However, it remains unclear whether dietary recommendations generated by AI chatbots align with established public health guidance, such as national dietary guidelines, or whether they are more strongly influenced by the broader digital information environment on which these systems are trained, which may include references with conflicts of interest, such as recommendations focused on nutrients that align with food corporations’ narratives.

Therefore, this study evaluates whether healthy eating recommendations generated by four AI chatbots align with the DGBP and whether responses include characteristics commonly found in food industry communication, such as references to branded products, health claims, or nutrient-focused messages. To do so, we adopted an exploratory design and developed two indicator-based instruments, which are made available for use in future research.

2. Materials and Methods

2.1. Study Design

This exploratory study examined whether healthy eating recommendations generated by AI chatbots align with the DGBP recommendations and, within the framework of the commercial determinants of health, assessed the presence of food corporations’ influence in the AI chatbots’ responses—either directly (e.g., presence of brand recommendations) or indirectly (e.g., food corporations’ narratives)—meaning that the information may originate from sources not necessarily affiliated with official bodies. This framework was informed by the literature on the commercial determinants of health and by Scrinis (2020), who discusses strategies used by the food corporations to shape knowledge about food and nutrition, thereby undermining scientific evidence free from conflicts of interest. It should be noted that this design identifies content that is consistent with food corporations’ narratives; it cannot establish that industry influence produced that content, since a nutrient-centered framing may equally originate from mainstream clinical nutrition or from the composition of the models’ training corpora.[13]

2.2. Data Collection—Interaction Procedures with the AI Chatbots

2.2.1. Sample

Four widely accessible AI-based conversational systems were included in the analysis: ChatGPT (GPT-5.2 Instant; Free plan; OpenAI, San Francisco, CA, USA), Gemini (Gemini 1.5 Flash; Free plan; Google LLC, Mountain View, CA, USA), DeepSeek (DeepSeek-V3.2; Free plan; DeepSeek, Hangzhou, Zhejiang, China), and Grok (Grok 4.20; Free plan; xAI, San Francisco Bay Area, CA, USA). All interactions were conducted using an anonymous browsing session in Google Chrome. An email account created specifically for this study was used to access the platforms that required an account. All interactions were conducted in Portuguese from Campinas, São Paulo, Brazil. These systems were selected based on their popularity at the time of study, without applying any criteria regarding quality or performance of these platforms [14]^.

2.2.2. Development of Interaction Questionnaires

We used two questionnaires to guide interactions with the four AI chatbots. The first was the validated GAB-1 instrument, which is an instrument to assess primary health care workers’ knowledge of healthy eating based on the DGBP [15]. The AI chatbots were instructed to respond independently to each statement, without following predefined response patterns, and to provide justifications based on their training data. The GAB-1 instrument comprises 21 statements covering the main messages of the DGBP. For each statement, responses were classified as “true,” “false,” or “unsure/I don’t know.” As the GAB-1 was developed for health professionals, it constituted the technical arm of the present design.

The second questionnaire was developed by the authors to simulate a citizen’s interaction with an AI chatbot on healthy eating topics, in order to assess consistency with the principles of the DGBP and identify the presence of food corporations’ narratives. This interaction guide was based on the semi-structured interview method and guided by the research question, allowing the study to maintain focus while enabling in-depth exploration of responses. Each question had a specific objective aligned with the main research question, ensuring systematization of the collected information and supporting the researcher (acting as a simulated user) in exploring the topic comprehensively [16]. Two authors, familiar with the DGBP and the commercial determinants of health, developed the interaction guide. It was reviewed for content validity by the project coordinator, who acted as an expert reviewer and contributed to the development of the DGBP, based on criteria of coherence with study objectives, clarity, relevance, and comprehensiveness [17].

This guide was based on the five chapters and the final section of the DGBP, including questions regarding the definition of healthy eating, recommendations on food choices and dietary patterns, meal organization throughout the day, cooking practices, eating environments and commensality, and common barriers to healthy eating such as cost and information. The questions followed a sequence of increasing complexity and included follow-up questions to deepen the themes addressed. For each theme, primary and secondary objectives were defined regarding the type of information expected from generative AI responses, and based on these objectives, questions were formulated for use in interactions with AI chatbots. Supplementary Table S1 presents the guide in detail.

To ensure that the structure and wording of the questions reflected those that could be used by the general population when interacting with AI chatbots, the questions were tested using Google Search, based on a previously proposed approach by Abroms et al. (2025) [9]. This process followed the suggestions displayed in two sections of the search engine— “People also ask” and “Others also searched for”— using an anonymous browsing window in Campinas, São Paulo, Brazil. The search was conducted in Portuguese in August 2025. For example, the question for Chapter 1 in the questionnaire was “What is healthy eating?”. This question was entered into Google Search, and the following questions were identified in the “People also ask” section: “How can you define healthy eating?”; “What does it mean to have a healthy diet?”; “What does it mean to eat healthy?”; and “What are the three pillars of healthy eating?”. In the “Others also searched for” section, the following related searches were identified: “Importance of healthy eating”; “A text about healthy eating”; “Summary about healthy eating”; “Essay about healthy eating”; “Five benefits of healthy eating”; “Research on healthy eating”; “Text about healthy eating with interpretation”; and “Essay on healthy eating (20 lines)”. From these options, the research team selected the question that best aligned with the chapter’s aim, which could be either suggested by Google Search or developed by the research team. The final questionnaire used in the chatbot interactions is presented in Supplementary Table S2.

2.2.3. Interaction Procedures

Each questionnaire was submitted once to each AI chatbot, on different days. Some questions of the second questionnaire, which was developed to simulate a citizen’s interaction with an AI chatbot on healthy eating topics, were slightly adapted to maintain a natural and coherent flow during the interaction with the AI chatbots. For example, in the interaction with ChatGPT, the question “What does it mean to have a healthy diet?” was sufficient to address all the aims of Chapter One, as the response included guidance on food choices. However, in the interaction with Gemini, two questions were necessary to evaluate all the items in Chapter One: “What does it mean to have a healthy diet?”and “What should I do to maintain a healthy diet?”. Before initiating each interaction with the AI chatbot, we included a brief user profile describing a 20-year-old woman seeking advice on healthy eating, to standardize the prompts and better reflect common user scenarios [18]. All interactions were conducted in Portuguese to reflect the linguistic and cultural context of the DGBP and to simulate typical information-seeking behavior among Portuguese-speaking users.

The full responses generated by the AI chatbots were recorded and saved in their entirety to preserve the original wording and structure of the outputs. Because AI chatbots are frequently updated, all prompts were submitted within the same data collection period (February and March of 2026) to minimize variation associated with model updates. The date of interaction with each system and the model version available at the time of data collection were documented to improve transparency and reproducibility. The complete list of prompts used in the study is provided in Supplementary Table S3.

2.3. Evaluation Framework

We used three instruments to assess AI chatbot interactions. For the interaction in which the GAB1 instrument was applied, we followed the template described in the original validation study, which the responses of “I don’t know” were classified as incorrect and the performance was assessed as the percentage of correct responses [15].

For the interaction simulating a citizen asking about healthy dietary recommendations, we developed an evaluation framework comprising two instruments based on predefined indicators: (1) misalignment with guideline recommendations, including elements reflecting food corporations’ narratives, either directly or indirectly, which was informed by the commercial determinants of health literature and by the work of Scrinis (2020); (2) alignment with DGBP recommendations, using indicators derived from each chapter of the guideline [11,13]. We use the term “misalignment with guideline recommendations” because the DGBP was developed based on scientific evidence without influence from the food industry or industry-funded studies [12].

Although this study was exploratory, we conducted a pilot phase to assess the two evaluation instruments developed for this study. We evaluated interactions with all AI chatbots using pilot instruments, and two nutritionists who were not members of the research team, had expertise in public health, and were familiar with the DGBP subsequently assessed inter-rater reliability. Inter-rater reliability was assessed using the kappa coefficient classified according to Landis & Koch (1977) [19], which indicated high agreement (κ = 0.81) for the first evaluation instrument and fair agreement (κ = 0.25) for the second evaluation instrument. Based on these results, we revised the instruments by developing a response expectation model to standardize the classification of indicators and introducing a partial-response category to assess alignment with the DGBP indicators. This category allowed responses that addressed only part of the concept covered in each DGBP chapter to be classified accordingly. All interactions were then re-evaluated using the revised instruments, and these results represent the findings reported in this study.

Finally, the instrument for assessing misalignment with guideline recommendations or reflecting food corporations’ narratives included 9 indicators with binary responses and used binary scoring to evaluate the presence (1) or absence (0) of content inconsistent with the guideline or reflecting food corporations’ narratives. The indicators covered aspects of the food corporations’ narratives about healthy eating, food choices, the presence of brand recommendations, and the sources referenced to assess potential conflicts of interest. This instrument is available on Supplementary Table S4.

The instrument for assessing alignment with DGBP recommendations included 14 indicators. It is important to note that the DGBP is a document in which the sections are interconnected. Accordingly, the chapters are intended to be read as a cohesive whole, and no differential weighting was assigned to the indicators. The combination of all indicators seeks to capture the main idea of the DGBP, which is presented across the five chapters and the final section. Indicator 1 refers to Chapter 1; Indicators 2, 3, and 4 refer to Chapter 2; Indicators 5, 6, and 7 refer to Chapter 3; Indicators 8 and 9 refer to Chapter 4; Indicators 10, 11, 12, and 13 refer to Chapter 5; and Indicator 14 refers to the final section.

We used a three-level scoring system in this instrument: (i) full alignment, when all relevant information from the guideline was present; (ii) partial alignment, when only part of the information from the guideline was present; and (iii) no alignment, when the response provided little or no information consistent with the guideline. Scores were assigned as follows: full alignment, 2 points; partial alignment, 1 point; and no alignment, 0 points. This evaluation instrument and the response expectations are presented in Supplementary Table S5.

2.4. Data Analysis

For GAB1, each AI chatbot’s response was calculated as the percentage of correct responses according to the answer key. Based on the criteria established by Reis and Jaime (2019) [15], scores ≥75% were classified as indicating consistent information aligned with the DGBP, whereas scores <75% were considered to indicate inconsistent information in relation to the DGBP.

For both instruments, descriptive statistics were used to summarize the responses produced by each AI chatbot, and differences across AI chatbots were assessed using Cochran’s Q test for the nine binary misalignment indicators and Friedman’s test for the 14 ordinal alignment indicators, with indicators treated as blocks and chatbots as the conditions compared. Where Friedman’s test indicated a significant difference, pairwise Wilcoxon signed-rank tests were applied as post hoc comparisons. In this analysis, each item was treated as a matched observation across systems, allowing comparison of the frequency of indicators while accounting for within-item dependence. Given the small number of items and their thematic (rather than psychometric) relationship, results were interpreted as exploratory. All analyses were conducted using a significance level of 5%. No adjustment for multiple comparisons was applied, given the exploratory nature of the analysis, and all analyses were performed in Stata 19.0 SE edition.

3. Results

Regarding interactions with AI chatbots based on technical questions (GAB1), all AI chatbots were classified as providing information consistent with the DGBP, as they correctly answered the GAB1 items and provided accurate justifications for each selected response. The only exception was statement 5 — “A diet predominantly composed of ultra-processed foods is cheaper than a diet predominantly composed of unprocessed or minimally processed foods” — for which ChatGPT selected the “I don’t know” option, arguing that this relationship may vary across different contexts. This response was therefore considered incorrect. Consequently, ChatGPT achieved 95.2% accuracy (20/21 items) in the knowledge domain assessed by the GAB1, whereas the other chatbots achieved 100% accuracy (21/21 items) (data not tabulated)

In the application of the questionnaire developed in this study, designed to simulate an interaction with a citizen, the AI chatbots exhibited distinct patterns of performance across the evaluated indicators. Regarding the indicator assessing content misaligned with the guideline or reflecting food corporations’ narratives, as shown in Table 1, all AI chatbots presented at least one response that met the criteria for this indicator, indicating content that was misaligned with the guideline or could reflect food corporations’ narratives. The proportion of scores on these indicators did not differ significantly across the AI chatbots (Cochran’s Q = 4.71, df = 3, p = 0.194), although Grok showed the lowest proportion of content misaligned with the guideline or reflecting food corporations’ narratives (ChatGPT 3/9, 33.3%; Gemini 4/9, 44.4%; DeepSeek 4/9, 44.4%; Grok 2/9, 22.2%). Five of the nine indicators (1, 5, 6, 7 and 8) were scored zero for all four chatbots and therefore did not contribute to the test statistics.

Overall, all AI chatbots scored similarly across the individual indicators. For example, all AI chatbots focused on recommendations regarding food choices within food groups primarily from the perspective of their nutritional content, emphasizing the composition of meals in terms of protein, carbohydrates, and fat, as mentioned in the interaction with Gemini: “The best snacks are those that combine fiber + protein or healthy fats”.

However, an exception was observed regarding recommendations involving UPFs as part of a healthy diet or as a strategy to save time. Only Gemini and DeepSeek included such recommendations in their interactions. For example, when asked about diet organization, Gemini suggested: “Emergency Drawer: Keep a ‘safety stock’ at work with 70% dark chocolate, good-quality protein bars, or seeds.” Similarly, when asked what to eat when seeking healthy food on the street, DeepSeek suggested: “Natural sandwich: It is your greatest ally! Ask for whole-grain bread (if available) with turkey breast or shredded chicken, plenty of salad (lettuce, tomato, grated carrot), and light cream cheese or cream cheese. Be careful not to overdo mayonnaise and ready-made sauces.”

When assessing alignment with the DGBP, the AI chatbots fully addressed fewer than half of the DGBP indicators assessed in their nutrition advice, as shown in Table 2. Alignment scores differed significantly across the AI chatbots over the 14 items evaluated (Friedman’s χ² = 11.07; df = 3; p = 0.011). Post hoc pairwise comparisons indicated that Grok scored higher than ChatGPT (Wilcoxon signed-rank, p = 0.011) and Gemini (p = 0.034), with a borderline difference relative to DeepSeek (p = 0.059). Grok achieved the highest score (20 of a maximum of 28 points, 71.4%), followed by DeepSeek (15 points, 53.6%), Gemini (14 points, 50.0%) and ChatGPT (12 points, 42.9%). Furthermore, Grok demonstrated the most comprehensive coverage, correctly addressing 42% of the main content of the Guide across all 14 items, more than twice the rate observed for ChatGPT and Gemini (Table 2).

Regarding the individual indicators, the AI chatbots showed the poorest performance on Indicator 1 — “Addressed the interrelationship between healthy and adequate diets and a socially and environmentally sustainable food system” — as only Grok addressed the sustainability of the food system as a factor influencing access to healthy foods. However, Grok did not explain that this relationship is mutual, as the healthy eating pattern proposed by the DGBP promotes a sustainable food system. Poor performance was also observed for Indicator 3 — “Presented the Brazilian Dietary Guidelines’ Golden Rule, ‘Always prefer natural or minimally processed foods and freshly made meals to ultra-processed foods,’ as the guiding principle for food choices.” For this indicator, the AI chatbots tended to present their own “golden rules” rather than the DGBP’s Golden Rule as the guiding principle for healthy eating. ChatGPT and Gemini did not present the Golden Rule at all, DeepSeek presented it only partially and Grok reproduced it in full. Table 3 presents these nuances in the discourse surrounding the different “golden rules” and the DGBP’s Golden Rule. Table 3 presents these nuances in the discourses surrounding the different “golden rules” and the DGBP’s Golden Rule.

In addition, the AI chatbots showed poor performance on Indicator 14 — “Presented the 10 Steps to Healthy Eating from the Dietary Guideline for Brazilian Population” — which summarizes the most relevant content of the DGBP. Most AI chatbots received a score of 1, except DeepSeek, which achieved 2 points. ChatGPT presented the steps in a disorganized order and omitted several key elements from the DGBP. For example, its initial steps — “Make meals regular and peaceful moments”; “Eat a variety of foods”; “Prefer unprocessed or minimally processed foods”; and “Limit ultra-processed foods” — differed from and did not accurately reflect the corresponding steps in the DGBP. Gemini presented ideas that were broadly consistent with most of the steps but omitted the step “Be critical about information, guidelines, and messages about food conveyed in commercial advertising.” Grok likewise presented the steps only partially.

4. Discussion

This study showed that AI chatbots correctly identified DGBP-consistent statements on healthy eating when engaged in interactions using a technical approach. Accordingly, when queried using professionally oriented terms that prompted a scientifically oriented interaction based on the DGBP, the AI chatbots were almost entirely capable of providing appropriate and assertive information on the promotion of adequate and healthy eating. Conversely, in the second interaction, which simulated a conversation with a citizen, the AI chatbots showed variation in the presence of content misaligned with the guideline or reflecting food corporations’ narratives and less consistent alignment with the DGBP, with nuances in performance across the AI chatbots.

Regarding technical expertise, similar studies that employed questionnaires to assess the competencies of AI chatbots found that their knowledge is comparable to that of a nutrition student and that, when asked to classify foods according to the NOVA system, they showed potential to support nutrition research on this task [20,21]. However, the authors argue that AIs lack precision in guiding individuals based on their specific circumstances and that response variations occur with changes in prompt wording, indicating potential biases in their answers and raising concerns that they may disseminate misinformation with consequences for public health [22,23].

In this direction, the present study found that, in simulations of everyday, non-professional interactions, AI chatbots may exhibit bias and ambiguity in the information provided. Although they generated several responses aligned with the DGBP, they also produced confusing or misaligned information that could mislead users about healthy eating, particularly when responses reflected narratives or recommendations commonly influenced by the food corporations’ narratives and potentially involving conflicts of interest [24].

Part of the ambiguity observed in the information provided to consumers can be attributed to the ongoing contestation between scientific evidence in nutrition and public health and the discourse promoted by the food corporations, both of which seek to influence public policy formulation and participation in contemporary food governance. In this context [25,26], the proliferation of competing information and narratives contributes to the maintenance and updating of what has been termed nutritionism [27,28]. This may help explain why, in the present study, all chatbot interactions discussed food groups primarily in terms of their macronutrient content rather than their place within dietary patterns and food traditions. It is important to note that when food systems are predominantly framed through a nutrient-centric approach, there is a risk of fostering the perception that foods—even those that are part of traditional dietary patterns—should be excluded to achieve a healthy diet, while reformulated food products may be legitimized as suitable substitutes in meals [28,29].

Furthermore, in routine health communication, nutrition professionals rarely translate dietary guidance into explicit quantities or proportions of macronutrients for patients [30]. Although such metrics underpin technical nutritional knowledge, they are generally conveyed through food-based recommendations, meals, and dietary patterns, which are more accessible and actionable in real-life settings. Therefore, AI chatbots should avoid predominantly framing dietary guidance in these terms, aligning instead with food-based approaches recommended in public health nutrition [10].

It is noteworthy that two of the AI chatbots (Gemini and DeepSeek) recommended the consumption of certain ultra-processed foods — cereal/’protein bars and turkey breast, which were central to past diet trends associated with the idea of healthy food. Despite the existence of substantial current evidence pointing to the long-term ineffectiveness and potential harms to physical and mental health of such strategies, the discourse suggesting possible benefits of these products remains prevalent [31,32,33].

In reality, fad diets generate cycles of consumption, and as marketing rapidly identifies consumer trends and perceptions, currently shaped by a fitness-oriented health culture, it sustains the overvaluation of a single nutrient based on real evidence (that protein increases satiety and aids muscle formation) to promote the sale of new products [34,35]. A clear contemporary example is protein bars [34,35]. Despite arguments from nutrition professionals that the nutritional quality and absorption of these proteins may be substantially lower than commonly perceived, due to the disruption of the food matrix and the ingredients used, such products have shown strong sales driven by nutritional claims [34,35]. In the present study, they were also recommended by Gemini.

Only one chatbot (Grok) addressed the importance of a healthy and sustainable food system in promoting adequate and healthy diets, and did so only partially, without acknowledging that this relationship is reciprocal. This gap is inconsistent with the principles of the DGBP, the Framework for Food and Nutrition Education for Public Policies, and the National Food and Nutrition Policy, raising questions about the ability of these tools to provide high-quality information aligned with public health and nutrition policies. Furthermore, neglecting food system–related aspects contradict established scientific evidence in nutrition and public health, which highlights the central role of the food system in shaping healthy dietary patterns [36,37]. Consequently, these tools often fail to convey to users that achieving a healthy diet depends on a healthy and sustainable food system.

The recommendations most frequently and thoroughly provided by the chatbots, and which were aligned with the DGBP, were those focusing on individual actions, such as examples of healthy meals through culinary preparations, the importance of water consumption, meal planning, and aspects related to the act of eating and commensality. While these recommendations are consistent with the DGBP, and although no chatbot framed responsibility as resting solely with the individual, the balance of the advice emphasized individual responsibility at the expense of the broader social, cultural, and environmental dimensions of food choices [38]. One approach that could help address this limitation would be the development of chatbots by health authorities, based exclusively on official government documents, which largely emphasize the importance of sustainability, the food system, and food sovereignty in promoting adequate and healthy diets. Such tools are technically feasible through retrieval-augmented generation, in which responses are constrained to a curated corpus of official documents, and would allow health authorities to guarantee both the provenance and the traceability of the information delivered to the public. Additionally, the Framework for Food and Nutrition Education for Public Policies emphasizes that partnerships between public and private sectors should not be established in the promotion of Food and Nutrition Education due to potential conflicts of interest in the information provided [36].

Although all AI chatbots cited the references used during the interactions, it was unclear which information was drawn from each source, highlighting the so-called “black box” nature of these tools. The lack of transparency regarding sources represents an additional obstacle to the safe use of these chatbots and further underscores the need for the development of official tools and/or other strategies to support the population in using these tools for nutritional or health guidance. Moreover, although all four chatbots warned that the advertising of ultra-processed foods seeks to present them as superior to other foods (Indicator 11), few explicitly addressed the commercial conflicts of interest that may affect the sources underpinning the information they provided.

This study is not without limitations. First, the interviews were conducted in Portuguese, and responses may differ in other languages. However, the focus of the study was to assess the alignment of the responses with Brazilian guidelines, which are based on scientific evidence free of conflicts of interest and have already influenced the development of several other dietary guidelines. Second, the questionnaire used to assess the alignment of nutritional information with the DGBP was not validated. However, no validated instrument was found in the literature to assess whether Food and Nutrition Education (FNE) aligns with the DGBP recommendations. To mitigate this limitation, we included in the evaluation process a validated instrument designed to assess understanding of the DGBP recommendations—although it was not specifically designed for evaluating FNE—and added a step in which external reviewers completed the evaluation of pilot questionnaire. This highlights the opportunity for future studies to develop or validate instruments to assess whether FNE aligns with the DGBP. Third, each questionnaire was administered only once to each chatbot; because large language models produce stochastic outputs, run-to-run variability could not be estimated, and the comparisons between systems should therefore be read as descriptive. Fourth, a single simulated persona (a 20-year-old woman) was used, so the findings do not indicate how advice might vary by age, sex, health status or socioeconomic framing. Fifth, the results are tied to the specific model versions available during the data collection period and may not generalize to subsequent releases. Sixth, the small number of items and the four systems compared limit statistical power, so non-significant results should not be read as evidence of equivalence. Finally, and most importantly for the interpretation of this study, the design identifies content that is consistent with food corporations’ narratives but cannot demonstrate that industry influence produced it; a nutrient-centered framing may equally reflect mainstream clinical nutrition or the composition of the models’ training corpora.

Furthermore, this study has several strengths, particularly the novelty of its dual design, which assessed both technical and lay-user interactions, and its exploratory approach. Exploratory approaches are commonly used to investigate emerging technologies and information environments in which standardized evaluation methods are still evolving. In the context of evaluating AI-based conversational systems and given the limited empirical research examining the nutritional information generated by these tools, this approach enables more systematic evaluation of their responses while also contributing to the development of tools for future research. This study provides two instruments that allow for the assessment of both alignment with the recommendations and principles of the DGBP and the presence of food corporations’ narratives in AI-generated responses. These instruments may support further research in this field and inform actions by public health agencies.

5. Conclusions

In this exploratory assessment, all four chatbots recognized DGBP-consistent statements when queried in technical terms, but the advice they generated for a simulated citizen covered fewer than half of the guideline’s core content, framed food choices mainly by macronutrient composition rather than processing of foods and largely omitted the link between healthy diets and a sustainable food system. All four produced at least one response inconsistent with the DGBP or consistent with food corporations’ narratives. Guideline-consistent knowledge is therefore retrievable from these systems without being reliably delivered to citizens. Given the single-run, single-persona design, these results are descriptive; still, they support evaluating approaches that constrain responses to official documents, such as retrieval-augmented generation applied to national dietary guidelines. The two instruments developed here are made available for such assessments. Finally, access and understanding the quality of information about eating is essential to support public policies, which seek to improve the availability of health eating information free from conflict of interest and aligned with national dietary recommendations.