Large language models tend to be overly cautious when judging recipes for diabetes.

People with diabetes need meals that follow specific health rules, but checking recipes takes work. Researchers tested whether large language models, which are artificial intelligence programs trained on vast amounts of text, can decide if a recipe is safe for someone with diabetes. To do this, a model must break a recipe down into ingredients and cooking steps, recall dietary rules, and apply those rules to the dish. The team tested the models using different ways of asking, giving them varying amounts of medical guidance. They found that most models acted cautiously, avoiding labeling recipes as suitable so they would not cause harmful mistakes. Models that could reason with medical guidelines did better. Overall, models named Mistral-7B and Llama 70B performed best.

Why it matters

People with diabetes need to carefully manage their diet to stay healthy. If AI can reliably analyze recipes, it could provide quick, safe guidance, but being too cautious might unnecessarily restrict their food choices.

Questions answered