If the problem is so evidently impossible then the LLM itself should recognise this, state that the problem isn't solvable, not* provide what's certain to be an inaccurate result, and suggest better approaches to arriving at a reasonable answer.

That said, it's notable that diabetes education materials often suggest estimating glycemic loads by rough portion size / plate ratios. Which is to say that absent accurate weight measurements (themselves subject to variations in ingredients, moisture levels, etc.) current clinical recommendations are themselves pretty rough.