Czy ChatGPT nadaje się do liczenia kalorii? Test dokładności, prompty i ukryte pułapki

The Viral Trend: "ChatGPT is My Dietitian"

Over the past year, fitness influencers across TikTok and YouTube have popularized a new life hack: instead of paying for a dedicated nutrition app or weighing food on a scale, simply snap a picture of your lunch, paste it into ChatGPT, and prompt: "How many calories and macros are in this plate?"

At first glance, the result looks miraculous. Within three seconds, ChatGPT outputs a beautifully formatted table breaking down protein, carbohydrates, fats, and total calories. It feels conversational, effortless, and free.

However, when clinical researchers and sports dietitians put large language models (LLMs) through rigorous double-blind testing against certified bomb calorimeters and laboratory gram scales, the results revealed systematic, dangerous blind spots.

The 10-Meal Clinical Accuracy Test: ChatGPT vs. The Scale

In a benchmark evaluation conducted using real-world meals prepared across both home kitchens and commercial restaurants, researchers compared GPT-4o Vision estimates against exact digital scale measurements:

Why ChatGPT Systematically Underestimates Calories

Notice the stark pattern in the table above: for simple, unadorned home foods (plain chicken breast, white rice, raw fruit), ChatGPT performs reasonably well. But for restaurant meals, stir-fries, and mixed pasta dishes, it underestimates calories by up to 37%. Why?

1. No 3D Volumetric Spatial Modeling

Dedicated food logging apps like Cal AI use computer vision networks specifically trained to estimate 3D point clouds and depth voxels from reference objects (like plate rims and silverware). ChatGPT does not compute spatial depth meshes; it treats the photo as a 2D text token association, frequently guessing that a 12-ounce dense steak is an 8-ounce cut.

2. The "Home Cook" Low-Fat Bias

ChatGPT was trained on broad internet text, including thousands of amateur recipe blogs that claim: "Sear the steak with 1 teaspoon of olive oil." In the real world, professional chefs finish that same steak with 2 to 3 tablespoons of butter (200–300 kcal of pure lipid fat). Because ChatGPT cannot see melted butter absorbed into muscle fibers, it defaults to the low-fat recipe text it memorized.

3. The Hallucination Vector

If ChatGPT is uncertain about a portion size, it is programmed by its RLHF safety alignment to generate a polite, plausible-sounding answer rather than admitting uncertainty. It will confidently tell you your creamy dressing is 60 calories when laboratory analysis proves it contains 240 calories.

The 5 Critical Features ChatGPT Lacks vs. Dedicated Trackers

Where ChatGPT DOES Shine in Nutrition

While you should never trust ChatGPT to maintain your daily caloric deficit, it is an extraordinary co-pilot for high-level nutritional strategy:

• Meal Prep Brainstorming: "Give me 5 high-protein dinner recipes under 500 calories that use chicken breast and canned lentils."
• Recipe Swaps: "How can I replace heavy cream in this soup recipe to reduce total fat by 70%?"
• Dietary Education: "Explain the biological difference between soluble and insoluble fiber in simple terms."

The Bottom Line

Use ChatGPT as your creative nutritional brainstormer, but leave daily food logging, portion estimation, and caloric accounting to dedicated computer-vision tools built specifically for metabolic precision. Relying on conversational AI to calculate your deficit is the fastest way to end up in a frustrating, unexplained weight loss plateau.