AI Carb Counting Tested 27,000 Times: Models Cannot Give Consistent Nutritional Answers

Diabettech·low signal

A researcher asked AI models to count carbohydrates 27,000 times and documented that no model could provide the same answer twice for identical foods, per a detailed blog post now trending on Hacker News (238 points, 297 comments). The study highlights a critical reliability gap for health-adjacent AI applications where consistency matters as much as accuracy. For builders deploying AI in health/nutrition contexts, this is a concrete demonstration of non-determinism risks.