The carbon footprints of recipes in a health and wellbeing mobile app: a cross-sectional study (Preprint)

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides methodological insights for quantifying recipe-level emissions and highlights data challenges in AI-generated dietary interventions.

🏢実務担当者:Food companies and app developers can use these findings to design AI recipes that lower carbon footprints and meet sustainability expectations.

🏛政策担当者:Supports policies promoting sustainable diets and carbon labeling, and underscores need for transparency in AI-generated food recommendations.

<sec> <title>BACKGROUND</title> Recipes play a key role in shaping dietary behaviours and habits, forming an integral part of meal planning and cooking routines. AI-generated recipes are increasingly popular, yet their content remains poorly characterised. This represents a critical gap in understanding how AI can be used to recommend recipes to consumers that are both healthy and sustainable. </sec> <sec> <title>OBJECTIVE</title> To quantify the greenhouse gas emissions (GHGEs) associated with recipes available in an AI-enabled commercial health and wellbeing app and compare the GHGEs between different recipes categorised by their primary protein source (plants, dairy/eggs, seafood, poultry, or red meat) and meal type (snack, breakfast, lunch, or dinner). This included examining the distribution of recipes across low, medium, high carbon footprint categories and identifying the food groups contributing most to emissions. </sec> <sec> <title>METHODS</title> Using a cross-sectional design, ingredient data from the app’s recipes were matched with GHGE data in carbon dioxide equivalents (CO2eq) sourced from life cycle assessments of global food production. Emissions per ingredient were aggregated to calculate total emissions per recipe, portion, and 100 g. Emissions per portion were classified using cut-offs adopted by the National Health Service (NHS) in England. Emissions for each protein source category were summarised as medians and interquartile ranges. Food groups contributing most to emissions in the highest-emission recipes were identified for each protein source category and displayed visually. </sec> <sec> <title>RESULTS</title> Across 205 recipes, the median (IQR) emissions (kg CO2eq) per recipe, portion, and 100 g were 1.22 (0.69-3.09), 0.51 (0.24-1.29), and 0.21 (0.14-0.44), respectively. Almost half were classified as low- or very low-emission per portion (&lt;0.50 kg CO2eq), yet a quarter were very high-emission (&gt;1.20 kg CO2eq). Plant recipes (n=105) had the lowest average emissions, followed by dairy/egg (n=40), seafood (n=34), and poultry (n=21) recipes. Red meat recipes (n=5) had over fourfold higher average emissions than seafood and poultry recipes. Most dinners (84.3%) contained seafood or meat and had more than double the emissions of other meals. Animal proteins contributed most to the highest-emission seafood, poultry, and red meat recipes, while plant and dairy/egg recipe emissions were more evenly distributed across food groups. </sec> <sec> <title>CONCLUSIONS</title> AI-generated recipes from health and wellbeing apps may yield environmental co-benefits, guiding consumer choices to support planetary health. Further adaptation of large language models, however, is required for dinner recipes to minimise the highest emission recipes. Methodological insights include the feasibility of quantifying recipe-level emissions, the complexities and assumptions of environmental data, and transparency challenges in using AI for generating recipes in dietary interventions. </sec>

• openalex https://doi.org/10.2196/preprints.109250first seen 2026-08-13 05:09:26