How to Start Using AI Safely as a Physician

The best way to learn AI is to use it, with guardrails

There is a simple principle that came up repeatedly in the session:

“The best way to learn generative AI is just to start using generative AI.”

That advice is correct. It just needs context.

Not all use cases carry the same level of risk. The safest way to start is outside of clinical care. Low-stakes use builds familiarity. It helps you understand how these tools behave before you rely on them in higher-stakes environments.

To see how clinicians are being introduced to this approach, watch Introducing AI Residency for Clinicians - Part 1

Resources:

• Session slides
• Dr. Michael Hobbs' AI 101 Guide

Start with low-stakes applications

Before using AI clinically, use it for everyday tasks. Ask it to plan a trip, draft an email, or summarize an article or explain a topic you are unfamiliar with.

These interactions help you build intuition. You start to recognize when outputs feel off. You begin to understand how different prompts produce different results. You notice where the model struggles.

That intuition is what allows you to safely identify issues later in clinical contexts.

Learn how to prompt effectively

AI outputs are highly dependent on inputs.

“What you put in is very impactful to what you get out.”

Vague prompts tend to produce generic responses. More detailed prompts tend to produce more useful ones.

A simple structure can help:

• Assign a role
• Provide relevant context
• Specify the desired format
• Set boundaries or constraints

Adding context reduces the likelihood of hallucinations. It also improves relevance. At the same time, more detail increases responsibility. The more context you provide, the more you need to ensure it is accurate and appropriate to share.

Protect patient data consistently

One of the most important habits to build early is protecting patient data.

“Keep protected health information out of your prompts.”

Unless your institution has a formal agreement with an AI vendor, you should assume that any data entered into these tools is not protected.

That means:

• No identifiable patient information
• No protected health information
• No shortcuts

This is less about the tool and more about the habit. Scanning your prompt before submitting it should become automatic, just like checking a medication dose before ordering it.

Understand where responsibility sits

AI does not change liability. The clinician remains responsible for decisions made using AI outputs.

“When AI gets something wrong, the clinician owns it.”

This is consistent with how medicine already works. You are responsible for decisions informed by guidelines, colleagues, or trainees. AI is no different. It is another input, not a replacement for judgment.

Use AI as an assistant, not a decision-maker

The most practical way to think about AI is as a highly capable assistant.

“Think of it like the world’s best assistant… really good at many skills.”

It can:

• Gather information
• Summarize content
• Draft documentation
• Suggest possibilities

It cannot:

• Take responsibility
• Understand context fully
• Replace clinical judgment

That distinction should guide how it is used.

Bottom line

AI is already part of the clinical environment. The difference between using it safely and using it poorly comes down to habits.

• Start in low-stakes settings
• Develop strong prompting skills
• Protect patient data
• Verify outputs consistently

Over time, these habits turn AI from a potential risk into a useful tool.