During diabetes care visits, AI-powered ambient scribes (such as Dragon CoPilot, Ambience, Heidi AI, Freed, and other EHR-embedded scribes)³ can assist with documentation by summarizing key elements of the conversation. Because this significantly reduces time spent at the computer, providers can focus more on the individual. Based on my experience, this reduction in administrative burden also enables deeper shared decision-making regarding management adjustments, health behavior changes, barriers, and concerns.

Following the visit, ambient scribes can organize information into a clear, personalized after-visit summary that reinforces key recommendations, action steps, and education provided. However, accuracy remains a concern; large language models (LLMs) generating discharge summaries were shown to contain errors, though the risk of harm was assessed as low¹.

From a dietitian’s perspective, LLMs can also help create individualized nutrition resources tailored to food choices, culture, and clinical needs. Instead of providing a generic meal plan, tools like ChatGPT, Gemini, and EHR-embedded AI can generate meal ideas, recipe modifications, or personalized food lists. A recent study using ChatGPT (GPT-5 version)⁴ found that ChatGPT could approximate carbohydrate targets but had limitations in accuracy when balancing multiple dietary requirements simultaneously (such as carbohydrate percentage, total calories, and meal planning constraints).

LLMs’ effectiveness and accuracy can improve with well-designed prompts that provide clinical context, specific goals, and defined constraints, but clinical review and oversight are recommended. Prompts may often require refinement, and LLMs simply react to input words, leading to hallucinations5.