Accuracy of ChatGPT , Gemini, Claude and DeepSeek in Carbohydrate Counting
Authors' conclusion
ChatGPT most accurately approximated clinicians' carbohydrate calculation among the tested AI models and fulfilled strict clinical equivalence criteria. Although the other models tended to underestimate carbohydrate content, their mean deviations remained within clinically acceptable limits.
Study context
- Population
- Meals equally distributed among breakfast, lunch, dinner and snacks; identical standardized English prompts
- Participants
- 124 meals
- Comparison
- ChatGPT, Gemini, Claude and DeepSeek vs clinician-calculated reference values
- Outcome
- Equivalence between clinician and AI-generated carbohydrate calculations within an error margin of ±5%
Methodological detailsequivalence (±5% margin), Two One-Sided Tests, Bland-Altman analysis
- Equivalence (±5% margin)
- Two One-Sided Tests
- Bland-Altman analysis
Performance
As reported in the abstract
Equivalence with clinicians' carbohydrate calculation (±5% margin)
- Only ChatGPT met the predefined ±5% equivalence criterion
- Gemini and DeepSeek achieved equivalence within a ±10% margin
Further findings
| Outcome | Values |
|---|---|
| Carbohydrate content per meal | Clinicians median 30.32 g · ChatGPT median 30.75 g · Gemini median 30.40 g · DeepSeek median 29.75 g · Claude median 29.25 g |
Agreement with clinicians (bias, limits of agreement, correlation)
- ChatGPT showed the smallest bias, narrowest limits of agreement and highest correlation with clinicians
- Claude displayed the largest negative bias and widest dispersion
Original publication
Accuracy of ChatGPT , Gemini, Claude and DeepSeek in Carbohydrate Counting. Diabetes Obes Metab. Published 13 April 2026. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →