Don't Let AI Lead to New Types of Med Errors

Top Takeaways

• Watch for AI-specific risks, such as automation bias, hallucinations, or voice-to-text issues.
• Review AI-generated documentation carefully to prevent errors from spreading into patient records, med orders, etc.
• Track AI-related errors and near misses to identify trends and strengthen safeguards over time.

You’ll have to take care not to trade old errors for new ones as the use of artificial intelligence (AI) increases.

We know AI offers powerful capabilities and can catch things that humans may miss. But it can also introduce new failure points that are very different from traditional errors...which makes them tricky to catch.

Be aware of AI-specific risks and help build guardrails into how you and your team use these tools...not just into the tools themselves.

For example, recognize automation bias with AI...the tendency to accept an output without verifying it, or to assume no alert means no problem...especially when you’re slammed or AI has been reliable so far.

This is particularly important since AI can hallucinate and generate clinical guidance, responses, or even references that sound confident and polished...without the red flags that signal something’s off.

Treat all AI-generated info as a starting point for clinical review...not evidence. And keep in mind, you’re accountable for any AI output you use. So confirm details and recommendations with trusted resources, guidelines, etc...especially when risk to the patient is high.

Avoid copying and pasting AI-generated text into the EHR or progress notes without thoroughly reviewing it first. Once an error is embedded in a record or system, it becomes harder to detect and correct.

Similarly, read AI-generated text in full before it moves downstream. Ambient documentation (AI that drafts notes by listening to patient encounters), voice-to-text order entry, and AI chart summaries can carry mistakes into patient records and med orders...such as sound-alike mix-ups (hydroxyzine vs hydralazine) or details that were never discussed.

Check AI-generated meeting summaries (P&T, etc) for accuracy too.

Reconcile records before trusting AI-generated or auto-filled data. For example, AI tools that populate admission med histories, discharge med lists, or care plans may perpetuate discontinued therapies, outdated info, or other inaccuracies if records aren’t reviewed and updated first.

Also guard against alert fatigue...it can compound quickly with AI and may lead staff to dismiss warnings, including the ones that matter. Work with informatics, P&T, and vendors to test new AI-driven alerts in “silent mode” before go-live and refine thresholds...rather than allowing “click-through” behavior to become the norm.

Stay alert for potential safety issues due to patient use of AI...especially during rounds or discharge counseling. Ask patients what information they have and where it came from...review the content together...and address any inaccuracies or missing context.

Help learners and colleagues develop a healthy dose of skepticism about AI-generated recommendations. Encourage them to ask not only “Does this sound right?” but also “How do we know it’s right?”

Advocate tracking AI errors and near misses as their own category in your safety event reporting system. This can help spot trends, flag problems after system updates, and improve safeguards before issues reach patients.

Dive into our Artificial Intelligence in Pharmacy Practice resource and our Clinical Capsules podcast episode for more best practices.

• FAQ: Artificial Intelligence in Pharmacy Practice