Physicians do not need to be convinced that thinking matters.

That is the job.

We think through incomplete information. We weigh risks that do not fit neatly into an algorithm. We explain uncertainty to patients who are often scared, overwhelmed, or just trying to make the best decision they can. We decide when something is routine and when it is not. We decide when a question needs a quick answer and when it needs a deeper look.

That is the part of medicine I still find most meaningful. It is also the part that I think gets squeezed the most.

Not always in obvious ways. It is not usually one giant task that steals all of our attention. It is the accumulation of small interruptions around the actual clinical work. A medication that needs to be checked. A study that needs to be found. A note that needs to be cleaned up. A patient explanation that needs to be made clearer. A clinical question that sits just outside the lane you live in every day.

None of those things are beneath us. And none of them are unimportant. But they all pull from the same limited reserve.

Attention.

That is what I think gets missed when we talk about physician time. Yes, time matters. Of course it does. But in medicine, the bigger issue is often what happens to our attention inside that time.

Because good medicine is not just about having information. It is about knowing what to do with information. It is about knowing whether a medication interaction matters for this patient. Whether a study applies to this situation. Whether the answer needs to change the plan or simply guide the conversation. Whether the patient understands what we are recommending and why.

That is judgment. And judgment requires attention.

That is also the lens I have started using when I think about AI tools in clinical practice. I have used general AI tools enough to see why physicians are curious about them and why they are cautious. A fast answer is not automatically a useful answer, especially if I cannot quickly tell where it came from or whether it applies.

That is one reason Doximity Ask, formerly DoxGPT, has become the AI tool I use most clinically. It is not because I want AI to make decisions for me. I do not. It's because Ask gets me to the right information faster, while leaving the actual judgment exactly where it belongs. With the physician.

So when I think about AI in medicine, I am not asking whether it can think for me. I am asking whether it can protect more of my attention for the thinking that actually requires me.

But there is a second part too. And it may be even more important.

A tool only helps protect attention if I can trust what it gives me. If I get a fast answer but then have to spend the next several minutes wondering whether the source is real, whether the information is current, or whether the answer actually applies, I have not saved time. I have just moved the work somewhere else.

And maybe worse, I have added a new kind of mental burden. That is why accuracy and transparency are not separate from efficiency. In medicine, they are part of efficiency. A tool only really saves time if it also makes verification easier.

That is the standard I think clinical AI has to meet. It should give physicians more room to think well. And it should be trustworthy enough that we can actually use that room.

Looking things up is not the problem

I think we sometimes talk about efficiency in medicine in a way that misses the point.

The goal is not to never look anything up. In fact, looking things up is often part of being careful. Medicine changes. Patients are complex. New treatments come out. New data comes out. And no physician, no matter how well trained, has every answer in their head at every moment.

The problem is not the lookup itself. The problem is when a simple lookup turns into a string of interruptions that pulls you away from the patient or the plan you were trying to think through.

I notice this a lot with breast reconstruction patients. These patients are often navigating many things at once. Chemotherapy. Radiation. Endocrine therapy. Surgical timing. Anxiety about recurrence. Expectations around reconstruction. Questions from family. A question that sounds simple at first may not be simple once you put it in the context of the patient’s full situation.

For example, a patient may ask whether a medication affects healing or whether a treatment changes the timing of surgery. I may know the general answer, but I still want to check the specifics. I want to understand the medication. I want to know if neutrophils, wound healing, or surgical timing are relevant. I want to see whether there is a source worth reviewing. And then I want to explain the answer in a way that is accurate without overwhelming the patient.

That is the real work. It is not just finding a fact. It is figuring out what the fact means for this patient.

So when I think about tools that might help, I am not thinking about replacing that process. I am thinking about making the process smoother. If a tool helps me get to relevant information faster, and lets me verify where that information came from, then I can spend less energy digging and more energy applying judgment. That is where the value is.

The wrong use of AI is outsourcing judgment

This is where I think physicians need to be careful.

There is a version of AI use in medicine that I am not comfortable with. It is the version where the tool gives an answer, the answer sounds good, and the physician accepts it without really thinking through whether it applies. That is not support. That is outsourcing.

And I do not think that is what most physicians actually want.

Most of us are not looking for a tool to be the doctor. We are looking for help with the work surrounding the doctoring. We want to get to the source faster. We want to check a medication without opening three different tabs. We want a patient explanation to start from something clearer than a dense paragraph of medical language. We want documentation to be more organized without losing the actual thinking behind it.

That is a much more realistic role for AI. And it is also a safer one.

If I use an AI tool to check a medication interaction, I am still the one deciding whether it matters for the patient. If I use it to get oriented around a chemotherapy-related question, I am still the one deciding whether I need to go deeper, call oncology, or review the literature myself. If I use it to help draft patient-facing language, I am still the one making sure the explanation matches the plan.

That responsibility does not move and I do not want it to move. What I do want is a tool that helps me spend more of my mental energy on the part of the work that actually requires a physician.

Where a tool actually starts to earn a place

That is the lens I use when I use AI tools. The question is not, “Can this give me an answer?” Lots of tools can give an answer.

The better question is, “Can this help me get to a useful, checkable answer faster, while keeping me responsible for the decision?”

That is a much higher bar. And it is the bar that matters clinically.

For me, this is where DoxGPT (now Ask) has started to earn a place. The useful moments are usually pretty ordinary. A medication question comes up during clinic. A patient asks something adjacent to my specialty, but not something I manage every day. I want to review a source before counseling someone. I want to turn a dense medical explanation into something a patient can understand. I want to tighten documentation after the clinical thinking has already happened.

These are not glamorous use cases. But they are real ones.

And in medicine, real matters more than flashy.

That is where Doximity Ask, formerly DoxGPT, can be useful.Not because it replaces judgment, but because it reduces the friction around judgment. The drug reference matters. The citations matter. The ability to move from an answer to the source matters. And physician review through efforts like PeerCheck matters because medicine is full of context that a general answer can miss.

This is also where speed matters, but only in the right way.

If a tool is slow or clunky, I simply will not use it during the moments when it could help. That is just the reality of clinical practice. The day keeps moving. Patients keep coming. Notes keep waiting. So a tool has to be fast enough to fit into the day.

But speed alone is not enough.

A fast answer that I cannot verify is not a solution. It is another task. A fast answer that sounds confident but does not show its work is not reassuring. It is concerning. And a fast answer that makes me wonder if I need to redo the search myself has not really saved me anything.

That is why I think clinical AI has to be judged differently than general AI.

In medicine, the answer is only useful if I can understand where it came from and decide whether it applies.

Accuracy is what makes the time savings real

This is the part that I think gets missed when people talk about AI saving time.

In medicine, accuracy is not separate from efficiency. It is part of efficiency.

If an AI tool gives me a fast answer but I cannot tell where the answer came from, I am not done. I now have a new task. I need to figure out whether the source exists, whether the information is current, whether the recommendation is reasonable, and whether it applies to my patient.

That is not saving time. It is shifting the work.

A tool only saves physician attention if it also makes verification easier. That is one of the reasons I am more comfortable with a medical-specific tool like DoxGPT (now Ask) than I would be with a general chatbot for clinical work. I want citations. I want to see the source. I want drug information that is grounded in something more reliable than a polished paragraph. I want to be able to move from the answer to the underlying reference and decide for myself whether it is useful.

And I think physician oversight matters too.

Doximity’s PeerCheck is interesting because it is designed to bring physician review into AI-generated answers. That matters because medicine is not general. A broad answer may be fine for a broad question, but real clinical questions often live in the details.

A chemotherapy-related question before reconstruction is not just a generic oncology question. A wound healing issue in a radiated patient is not just a generic surgery question. A medication issue in a complex patient is not just a drug lookup. It is that drug, in that patient, at that moment, with that plan.

That is why I do not want AI tools that just sound confident.

I want tools that help me verify faster.

The physician still has to be the physician

The most important guardrail is that the physician still has to be the physician.

If I use Ask to check a medication, I still decide what applies. If I use it to summarize or organize information, I still decide what matters. If I use it to help explain something to a patient If I use it to clean up documentation, I still make sure the note reflects what happened and what I actually think.

That is not a small distinction.

Because the goal of clinical AI should not be to make physicians less central. It should be to protect physician attention for the places where we matter most. That is really the standard I care about.

Does this tool make me a more efficient doctor without making me a sloppier one? If yes, I am interested. If no, I am not.

Disclaimer: This post was written in partnership with Doximity.