Researcher finds four AI models give wildly inconsistent carbohydrate estimates for the same food photo, with differences large enough to cause dangerous insulin dosing errors.

A study submitted 13 food photographs to OpenAI GPT-5.4, Anthropic Claude Sonnet 4.6, Google Gemini 2.5 Pro, and Google Gemini 3.1 Pro Preview, sending each photo 500 times to each model with identical prompts and settings. Across 26,904 queries total, every model returned different carbohydrate estimates for the same photo in repeated queries.

Claude Sonnet 4.6 showed median variation (coefficient of variation) of 2.4% with a median insulin swing of 0.9 U, while Gemini 2.5 Pro showed 11.0% variation and 4.7 U median swing. On a single paella photo, Gemini 2.5 Pro's estimates ranged from 55g to 484g across 500+ queries — a 429g range equivalent to 42.9 units of insulin at a 1:10 ICR (insulin-to-carbohydrate ratio).

On five images with the strongest reference values, Claude kept 100% of queries in the safe or moderate insulin error zone (under 2 units). GPT-5.4 had 37% of queries causing clinically significant insulin error (>2U). Gemini 2.5 Pro had 12% of queries causing errors exceeding 5U, the threshold associated with severe hypoglycaemia requiring third-party assistance.

Model confidence scores showed poor correlation with accuracy: Claude's confidence had r = -0.01 (zero correlation) with actual accuracy, and Gemini models reported confidence above 0.9 for 86–76% of all food items regardless of correctness. The DTN-UK stated earlier this year that generic LLMs must never be used as autonomous advisory calculators for insulin delivery.