Accuracy of polyphenol content information in berries: A comparative analysis of ChatGPT and Phenol-Explorer

Abstract

<h3>Background</h3><p>Polyphenols are widely occurring bioactive compounds in fruits and are extensively investigated for their potential health effects. The growing prominence of artificial intelligence tools in nutrition science necessitates evaluating their capacity to provide accurate biochemical data.</p><h3>Aim</h3><p>This analysis aims to assess the reliability of two models ChatGPT-4o mini (free version) and ChatGPT-4o (paid version) in predicting polyphenol compound concentrations and their potential use in nutritional research and health applications.</p><h3>Methods</h3><p>Seven different berries were selected for the study, and their anthocyanins, flavonols, phenolic acids, lignans, and stilbenes were queried in three different sessions using both ChatGPT-4o mini (free version) and ChatGPT-4o (paid version). The responses were compared with those from Phenol-Explorer, and the evaluation was based on relative accuracy (%).</p><h3>Results</h3><p>No significant difference in relative accuracy (%) was found between ChatGPT-4o mini (41.36 ± 34.74) and ChatGPT-4o (46.23 ± 34.01) models (<em>p</em> > 0.05; Cohen's <em>d</em> = −0.107). In ChatGPT-4o mini, the highest mean accuracy was observed for total polyphenols (68.01 ± 25.00%; significantly higher than flavonols, <em>p</em> < 0.01), followed by anthocyanins (58.95 ± 32.68%). In ChatGPT-4o, anthocyanins showed the highest accuracy (65.36 ± 38.17%; significantly higher than flavonols, <em>p</em> < 0.01, and stilbenes, <em>p</em> < 0.001) followed closely by total polyphenols (65.72 ± 20.93%). Accuracy for flavonols, phenolic acids, and stilbenes was lower than for other compounds.</p><h3>Conclusion</h3><p>This study shows that ChatGPT-4o mini and ChatGPT-4o exhibit varying accuracy in predicting polyphenols, with higher accuracy for common compounds l</p>