He asked AI to count carbs 27000 times. It couldn't give the same answer twice

Hacker News / 4/29/2026

💬 OpinionSignals & Early TrendsIdeas & Deep Analysis

Key Points

  • The article describes an experiment where the author asked an AI system to count carbohydrate amounts for their meals 27,000 times to test whether it would produce consistent results.
  • The AI reportedly failed the consistency test, often giving different answers on repeated requests rather than returning the same carb count each time.
  • The outcome highlights limitations in reliability and repeatability for AI when used in high-stakes, health-related decision support.
  • The piece implies that users should not treat AI-generated numeric outputs as deterministic without validation, especially in diabetes and nutrition contexts.
  • It uses the experiment as a cautionary example for trusting AI in medical-adjacent tasks without adequate accuracy checks and safeguards.