AI and medication safety
By ChatGPT Health. Educational synthesis prepared with AI assistance. Sources checked September 28, 2026. Clinically reviewed by Harvey Castro, MD, MBA, on September 28, 2026. Approved as educational content. This page is not a clinical protocol.
Medication safety depends on a complete medication history, reliable reference information and a qualified professional who can act on a concern. Evaluate each AI task separately.
Separate the tools
A prescribing alert, a medication reconciliation system and a general chatbot perform different jobs. Record whether a tool detects an unusual order, summarizes a list, checks an interaction or proposes a dose. A result for one task cannot validate the others. Ask who owns the final decision and how the original medication record remains available.
What one prescribing study measured
A 2019 prospective study in one hospital ward evaluated an outlier detection system alongside existing decision support over 16 months. It alerted on 0.4% of orders; clinicians judged 85% of alerts valid, and 43% led to later order changes. This measured alert usefulness, not a randomized reduction in patient harm. Read the primary study.
Questions for a clinical evaluation
- Does the system reconcile prescriptions, nonprescription medicines, supplements, allergies and recent changes?
- How are missing information, duplicate entries, units and outdated laboratory values handled?
- Can a pharmacist inspect the evidence behind an alert and document an override?
- Does testing include age differences, renal impairment, pregnancy and transitions between care settings when those are within scope?
- Measure missed errors and clinically inappropriate alerts as well as alert counts. Include patient harm and staff burden where feasible.
For patients and caregivers
Use AI to organize questions for a pharmacist or prescriber. Do not start, stop or change a medicine or dose based only on chatbot output. Bring the original medication list and explain what the tool said so a professional can check the context. Avoid entering identifying health information into a consumer chatbot. AMA guidance on patient use of AI chatbots.
A useful next step
For a proposed medication tool, complete the worksheet with the pharmacy team. Define an error review process, an escalation owner and a threshold for pausing the pilot before the tool influences real prescribing.
Sources
Explore related topics
Put the evidence to work
Download the evidence and safety evaluation worksheets or compare the study designs and their limits.
Editorial standards · Suggest a correction. Please do not send identifiable patient information.