A narrative review of AI in diabetes care argues that separating prediction tools (ML/LLMs) from automation/control methods (e.g., automated insulin delivery) is clinically crucial because evidence standards, safety risks, and governance needs differ substantially across these categories.

Why It Matters To Your Practice

• Diabetes care involves frequent, high-stakes decisions amid major day-to-day physiologic variability—exactly the setting where algorithmic tools can either reduce burden or amplify risk.

• More continuous glucose monitoring, connected insulin delivery, and longitudinal EHR data are increasing the number of AI-enabled tools entering routine workflows.

Clinical Implications

• Match oversight to function: automated insulin delivery/control systems generally require different validation and monitoring than ML/LLM-based prediction and recommendation tools.

• Adoption decisions should weigh clinical readiness and strength of evidence differently for: (1) closed-loop control, (2) risk prediction/pattern recognition, and (3) natural-language applications (documentation, messaging, triage support).

• Implementation planning should include safeguards proportional to harm potential, especially as tools seek greater autonomy and interoperability across devices and data sources.

Insights

• “AI” is not one risk category: control algorithms and machine learning models can fail in different ways and therefore need different governance.

• Evidence is strongest in established automated insulin delivery; emerging areas include wearables, “digital twin” frameworks, and LLM-enabled workflows.

• High-frequency physiologic data can improve consistency of decisions, but also increases the chance that spurious patterns or data-quality issues drive incorrect outputs.

The Bottom Line

• For diabetes mellitus (DM), clinicians should evaluate AI tools by what they do—predict, recommend, or control—because that function determines the appropriate evidence bar, safety monitoring, and implementation guardrails.