Can generative AI reliably estimate the environmental impact of diets? Evidence from four dietary scenarios
Nutrition and Food Science, ss.1-13, 2026 (ESCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1108/nfs-03-2026-0117
- Dergi Adı: Nutrition and Food Science
- Derginin Tarandığı İndeksler: Emerging Sources Citation Index (ESCI), Scopus, Hospitality & Tourism Complete, Hospitality & Tourism Index, Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Health Research Premium Collection (ProQuest), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Sayfa Sayıları: ss.1-13
- Anahtar Kelimeler: Artificial intelligence, Carbon footprint, Generative AI, Life cycle assessment, Sustainable diets
- İstanbul Gelişim Üniversitesi Adresli: Evet
Özet
Purpose – The purpose is the study was to evaluate the accuracy and consistency of artificial intelligence (AI) models in estimating environmental parameters of different diet types. Design/methodology/approach – A model performance evaluation study was conducted using four generative AI systems (ChatGPT-5.2, Gemini-3, Grok-4.1 and DeepSeek-V3.2). Four standardized dietary scenarios were developed (current diet, model diet, lacto-ovo vegetarian diet and vegan diet), each providing 2000 kcal/day. Carbon footprint (CF), water footprint (WF), energy footprint (EF) and land use were calculated using life cycle assessment–based databases and used as reference values. Each AI model was queried three times using the same standardized prompt and mean outputs were analyzed. Findings – Significant differences between AI-generated estimates and reference values were observed across most dietary scenarios and sustainability parameters (p < 0.05). ChatGPT-5.2 demonstrated the best overall performance, particularly for WF [mean absolute error (MAE): 124.19; absolute error: 6.11%] and EF (MAE: 5.46; absolute error: 13.82%). DeepSeek-V3.2 showed the lowest prediction error for CF (MAE: 0.25; absolute error: 12.18%) and land use (MAE: 1.12; absolute error: 23.99%). In contrast, Grok-4.1 exhibited the highest deviations, with CF errors reaching 48.38% and substantial overestimations of land use (absolute error: 199.78%). Large deviations were particularly evident for WF and land use across several models. Originality/value – To the best of the authors’ knowledge, this study provides one of the first comparative evaluations of generative AI models in estimating environmental sustainability indicators of dietary patterns. By benchmarking AI-generated outputs against life cycle assessment–based reference values across standardized diet scenarios, the study offers new evidence regarding the reliability and limitations of AI in sustainable nutrition research.