Last month we looked at how our clinical encounters may be improved by AI, with AI scribes and EKG interpretation, as well as improvements to our EMR. This month, we’ll look at how other online AI tools can impact our role as physician and/or ED leader.
Chat GPT–Large Language Models
If you watched the Super Bowl earlier this year, you saw that AI is coming for you. About half of the population has used a large language model (LLM) like ChatGPT, Gemini, or Claude, but only about 20% of the population uses it on a weekly basis. About a year ago, I started to explore how frequently it’s used among those in my world, so I started with my son. His answer was, “Dad, I’m a college student. I’m an expert.”
On reflection, maybe I’m not getting my value in tuition. I also spoke to three professionals outside of medicine. One was an entrepreneur with a large social media presence. I asked how he created so much content, and he said it was all done by AI.
Another contact lives in the computer science world and spends his day coding. I know very little about coding, but to probably no one’s surprise, AI is great at creating code. A third is president of an international engineering company and she used it routinely for her email communication and business proposals.
All of this made me realize I need to learn how to incorporate it into my life. My early tests were in rewriting brief emails and creating announcements and social media posts. On the personal side, I used it to plan a good bit of a family trip to London last summer. It was amazing.
I was at a conference recently given by attorneys and one of the speakers said it wasn’t uncommon for him to rewrite memos or briefs upwards of 20 times. Of course, if I’m getting paid by the hour, maybe I would do that too. But for those of us who spend a good amount of time at a desk, with too much work and not enough dedicated office time, LLMs can save time when it comes to some of our writing.
Although I don’t write a lot of letters of recommendation, I’ve saved time on some recent ones. On my last one, I found that I had several paragraphs I really liked and then a rough outline of a few more thoughts.
With enough prompts, instructions, and revisions, I got a letter that I thought captured the person well in much less time than it typically takes me to write one. A colleague also told me ChatGPT produced such an outstanding letter for another person, he had to revise it and told ChatGPT to tone down the praise and don’t make his writing so erudite.
I also find it can be helpful when I can’t quite achieve the tone I’m looking for. Sometimes I want an email to be more positive and upbeat. Other times, I need more action and am looking for a tougher tone. I realize I ramble sometimes in emails, so getting a “tighter” email with 20% less words can make the finished product more impactful.
Getting the Most out of your LLM
There are entire books and internet sites dedicated to maximizing your output when you use an LLM. The secret is in the prompt. The better your prompt, the better your outcome. Use more details. And then continue to add more prompts.
Start by introducing who you are—such as, “I’m a medical director of a high volume, high acuity Emergency Department.”
Then give detailed instructions about what you’re doing, including any data it may need–I’m writing a letter of recommendation for our lead scribe…..insert key lines from a job description, insert phrases about your interactions with the scribe.
The more background information you can offer, the more specific the writing can be.
Tell it exactly what you want to do (write a glowing letter for recommendations for medical school, for this scribe who I consider to be among the best candidates I’ve ever worked with)
Since my team and my bosses are reading this, I have to clarify that I probably use ChatGPT on 5-10% of my total writing. If writing is a tool, I think of it as sharpening my spear. On the other hand, it’s also great for putting a bunch of ideas down and asking for a finished product.
Think of that as starting with a quiver of arrows put into ChatGPT and you then pick the best arrow. Whether it’s ChatGPT or an AI scribe creating your medical decision making, it requires oversight and editing until you are satisfied with the finished product.
My favorite high school teacher made it clear that the secret to good writing is rewriting and that is true with LLMs as well. Pretend you’re the editor sending it back to the writer and have a conversation about specific changes, clarifications, or improvements you’d like the writer to make. Don’t settle for one revision. I had a colleague tell me that he will frequently use two different AI products, asking for feedback and revisions on something, and toggle between the two until he has a product he likes.
Another great use of AI is for slide presentations. How many of you are asked to give a brief medical update to your nursing team on a medical subject? LLMs can create these in minutes as well. To date, I’ve actually never given a lecture with AI slides but I’ve experimented by creating several 10 slide lectures to understand the process and what the outcome can be.
I’ve personally had better success with Google Gemini than ChatGPT, but both can provide anything from an outline to a finished product. Another colleague thinks NotebookLM is the best product for this.
Whether you’re writing or producing a slide deck, consider prepping it with studies and background information from Open Evidence and then ask your AI tool to include that information in your presentation.
Open Evidence vs UpToDate
We can all agree that Dr Google is rarely helpful? ChatGPT may be better from a patient perspective, if they can set the prompts up correctly. The question becomes how do we use LLMs to help with patient care. I’ve been using UpToDate for as long as I can remember. It can easily answer a straightforward question, it’s built into Epic, and I get CME. UpToDate is probably the gold standard for finding peer reviewed information online, but it can be a bit bulky to get through. So after watching some of my team use Open Evidence, I created an account and am starting to use that. Open Evidence is an AI literature search engine for clinicians and is built for point of care decision support.
You need an NPI number to register and gain access. It pulls data from JAMA, the NEJM, The National Comprehensive Cancer Network, Mayo Clinic, and partners with ACEP. ChatGPT can perform a search, but it’s really designed to explain, summarize, and educate, and not give specific best practice recommendations. ChatGPT can be useful in the clinical setting for helping create a differential diagnosis. It can even be helpful when it comes to finding the right talking points to have a conversation with a patient and/or their family about a situation.
Whereas Open Evidence is best when you’re looking for real clinical decisions, for example a dosage of a medicine or the sensitivity of ultrasound versus CT scan for diagnosing ovarian torsion or MRI for diagnosing a TIA (one of our inhouse recent discussions). Open Evidence has been a huge aide as I’ve had to rework some of our clinical policies—i.e. rewriting our EKG at triage policy and rewriting several of our trauma clinical practice guidelines. Open Evidence can also be great to find published literature that can then be uploaded to another LLM (like ChatGPT or Gemini) to be used in writing or slide deck creation. Keep in mind, Open Evidence is still AI with the potential to hallucinate (so check those references if something doesn’t seem right).
Dr CaBot, developed by Harvard Medical School researchers, is an AI system that reproduces expert diagnostic reasoning, working through a differential diagnosis, and ultimately showing a final diagnosis as a medical case presentation. It produces written and slide based presentations that emulate CPC formats. This can offer great teaching currently and we’ll see how it evolves into a way to manage complex cases in real time.
I’ve had a couple of friends recently who are going through some complex medical issues. One used ChatGPT to load in his symptoms, get a differential diagnosis, and then continued to add to the conversation after assessing interventions. Honestly, some of the conversations he had with ChatGPT were better than any doctor encounter he could have had.
It helped with his differential diagnosis and treatment options. Another friend has colon cancer. All of the diagnostic data, including detailed CT reports, were loaded in Open Evidence and a literature-based treatment plan was designed to serve as a jumping off point with their oncologist for treatment.
Future Questions
One of the questions that still needs to be answered is how much AI is too much? Do we want or need more built in assistance from medical specific AI sources that links AI to the chart—AI reviews chart, interprets labs,– and is designed to provide clinical decision support if you’re not asking for it—is it too annoying? Should AI be incorporated to create an evolving differential diagnosis as we go through our work up? Most of us already ignore many of the pop ups we get.
How about identifying an abnormal lab and making a recommendation, which may save you from a critical miss, but may be really intrusive with pop ups. Over the years, we’ve had a handful of missed labs at my ED (anything from a positive HCG to a significant bandemia) and have identified ways to make clinicians aware so it hopefully doesn’t happen again. These are only slightly intrusive. But should there be a computer reading over our shoulder identifying every abnormal (think hypoglycemia) and linking it to an order set? That seems a little much.
Consider an AI scribe that can provide immediate patient satisfaction feedback to the physician based on hitting certain discussion points or the use of empathetic statements. While it’s a great way to teach, should we all be getting notifications on every patient we see?
Even linking something like Open Evidence to our charts to generate prompts for order sets or a differential diagnosis, but on what kind of patients is that the most helpful? Our AI scribes are going to generate medical decision making notes so linking them to a true medical LLM is the way to go. And I will benefit from the note from a documentation and billing point of view.
Compliance
As AI moves into the clinical environment, the compliance question isn’t whether physicians can use it — it’s how we use it without shifting decision-making away from the clinician. Most ED tools will function as clinical decision support, similar to ECG interpretations or drug-interaction alerts. AI should inform our thinking, not replace it. The chart still needs to reflect physician judgment: we synthesize the information, we make the decision, and we document the reasoning. The medicolegal risk won’t come from using AI — it will come from appearing to defer to it.
Departments will also need clear governance around these tools. Physicians should be able to override recommendations, document disagreement, and understand that risk scores are inputs rather than conclusions. Enterprise-approved platforms, protected data workflows, and ongoing performance monitoring will matter more than the sophistication of the model itself.
If implemented thoughtfully, AI doesn’t remove responsibility — it clarifies it. The physician remains the decision-maker; the technology simply improves the visibility of risk. Ultimately, physicians remain responsible for the content and are therefore responsible for the hallucinations that can occur with AI.
The consent process surrounding AI scribe utilization in healthcare continues to evolve. Much like when human scribes first entered the exam room, most patients appear comfortable with the use of AI documentation tools, and acceptance will likely continue to grow as these systems become more commonplace across healthcare settings.
At the same time, some patients may be uncomfortable with the idea of their conversation being recorded or processed by an AI system, particularly during discussions involving sensitive personal or medical issues. Healthcare organizations must therefore navigate not only HIPAA and data privacy concerns but also varying state laws regarding single-party versus multi-party consent for audio recording. Most experts now recommend transparent disclosure, documented patient consent, use of HIPAA-compliant vendors, and a clear opt-out process.
Conclusions:
AI is not going away, and neither are the questions surrounding how best to incorporate it into emergency medicine. I had a doc that didn’t use Dragon for the first year (or two) that we had it. He just wasn’t quite sure how it worked or if he could trust it. Not every doc is tech savvy and wants to be an early adopter of technology. (A five minute lesson with my doc and he felt like his life was changed). Like every major technological shift before it—from electronic health records to bedside ultrasound—the challenge is not whether we adopt it, but how thoughtfully we do so.
Used appropriately, AI has the potential to reduce administrative burden, improve efficiency, enhance education, and support clinical decision-making, while allowing physicians to spend more time focused on patients rather than screens. At the same time, we must remain vigilant about oversight, accuracy, patient privacy, and preserving the human elements of care that define our specialty. The future of AI in emergency medicine will not be determined by the technology alone, but by the physicians and leaders who shape how it is implemented at the bedside. And the best news is that AI is rapidly advancing and the AI technology you’re using now is likely to be the worst AI you ever use.
