A new study compares ChatGPT’s ability to analyze meals with that of human dietitians to see if AI can accurately analyze what’s on your plate.
Does AI Know What’s on Your Plate?
AI is everywhere—from helping you choose what to watch on Netflix to writing emails. But can it also tell you how many calories are in your meal? A new study from researchers in Dublin, Ireland, put ChatGPT-4 to the test to see how well it can analyze food from meal photos and estimate its nutritional value. The results? It’s good at recognizing what’s on your plate but still has a long way to go before replacing human dietitians.
The Experiment: Can ChatGPT-4 Guess Nutrients?
The study, published in Nutrients, examined whether ChatGPT-4 could accurately identify food and estimate its nutritional content using just images. Researchers used 38 common Irish meals, each photographed in small, medium, and large portions. ChatGPT was tasked with identifying foods and calculating calories, protein, fat, carbohydrates, and essential vitamins and minerals. Seven human dietitians also analyzed the images for comparison.
How Well Did AI Perform?
ChatGPT-4 excelled at identifying foods, scoring an impressive 88.6% accuracy. It correctly recognized 93% of the food items in the images but missed or misidentified 7%. While this might seem pretty solid, the real challenge came when estimating portion sizes and nutritional values.
Out of 114 meal images, ChatGPT underestimated the weight of 87 meals—over 76% of the time. For small meals, its guesses were fairly accurate, but for medium and large meals, it significantly underestimated food weight. And that meant its calorie and nutrient estimates were often way off.
- Calories: Almost accurate with only a 0.1% difference from the actual values.
- Protein: Underestimated by 2.7%.
- Carbohydrates: Underestimated by 6.5%.
- Fats & Vitamins: Some big misses—polyunsaturated fat (-9.1%), calcium (-27.8%), potassium (-49.5%), folate (-38.6%), and vitamin D (-100%).
What About Human Experts?
Interestingly, dietitians also faced similar challenges when estimating food portions and nutrition based only on images. Even among human experts, there were inconsistencies in assessing calories and nutrients. The study found that ChatGPT’s estimates for protein and energy were quite similar to those made by dietitians, but it was less reliable when calculating carbohydrates.
The AI even noted the limitations of its estimates, pointing out that cooking methods, hidden ingredients, and food brands could affect the accuracy. Dietitians made similar observations, showing that even for humans, assessing nutrition from pictures isn’t foolproof.
AI vs. Dietitians: Who Wins?
While ChatGPT-4 is pretty smart, it’s not ready to replace a trained nutritionist. It can recognize most foods and make reasonable calorie guesses, but it struggles with portion sizes and specific nutrient breakdowns. The biggest issue? Underestimating large meals, which could lead to incorrect dietary recommendations.
However, this study shows AI’s potential in nutrition tracking. With further improvements—like better portion estimation, integration with food composition databases, and training on more diverse diets—ChatGPT could become a helpful tool for dietary assessments in the future.
The Bottom Line
If you’re counting calories or tracking your diet, ChatGPT might give you a rough idea, but don’t ditch your dietitian just yet. The AI is great at recognizing what’s on your plate but still needs some fine-tuning to provide fully accurate nutrition estimates. For now, if you want precise results, you’re better off consulting a human expert
