At the time it aired, it was just clever. But we’re learning in 2026 that the question is real. As creators and physicians, we’re confronted with Coolidge’s problem: we now have to prove something that used to require no proof at all.

Subscribe now

Look at me. I’m human

Here on Substack, writing is now scored for machine participation. This Wired article profiles how writers are now crafting their words to look most certainly human. The stakes aren’t remotely comparable, but it reminds me of Jeremiah Denton, the American Viet Nam POW who famously blinked “torture” in Morse code during a forced television interview. A small signal meant to show that a human being was still in there.

I haven’t paid Substack’s bullshit detector much attention, but for years before AI I confess that I was preoccupied with the idea that people would see me as a fraud. So I figured it was time to lean into some of this human mimicry.

As an example, I’ve been experimenting with visual metaphors on Instagram. These are clean, minimalistic graphics about medicine that ooze sarcasm and hopefully make a point. I do these by hand and build in a subtle degree of “wiggle” in my lines as a kind of human watermark.

What’s important about this moment is that for the whole of civilization, creative output like this was human. Paintings, books, architecture, and the important parts of medicine. We were the makers of things. I never would have believed that I’d see a moment when human work needed authentication.

But here we are. And increasingly, people don’t like it.

The emergence of a countercultural revolution

Look around and you see pushback. College students walking out on commencement speakers. The Summer of Ludd in July, a weird, ’60s-esque repudiation of AI in New York City. Threats against tech leaders seem to happen nearly weekly. Tinder is now (you’ll need to sit down for this one) organizing in-person meetups. And people would rather have land than take a payday for a data center.

In the medical community you can feel it. There’s real skepticism of AI in clinical practice and concern around creeping ambient surveillance. My partner refuses to use a scribe. Check out the comments on my Instagram posts. There’s a palpable libertarian vibe running through a profession that’s already had its autonomy eroded.

The funny thing is: when you look closely at these walkouts and protests, it isn’t clear what people actually want. This countercultural moment isn’t actually for anything that I can see. Sure, strip away the surface layer and you find fear and loss of control.

We know we’re losing something. We somehow can’t name it.

Finally fighting technological determinism?

The best way I can frame this is as a defiance of technological determinism. Technological determinism is the idea that technological progress is inevitable and that it shapes who we are and what we do. And so our job is to sit back, download every update, and use what’s created for us. Technology decides what comes next.

So maybe all of this is an evolving anti-deterministic position that we haven’t figured out how to articulate just yet. Maybe the wiggly line matters because it says more than Bryan drew this. It says I actually chose where the line should go.

That may be what we’re trying to preserve: the right to choose where the line goes.

I last did this in 1987 (sans scribe) during ‘Physical Diagnosis 1’ where I had to narrate for a proctor what I was doing and feeling. So it was weird hearing the jargon pass over my lips after all these years.

This has created an unexpected avenue for dialog with patients and parents.

Subscribe now

A few thoughts about my experience.

Abdominal exam double Dutch

The hand waving that we’ve always done during a physical was one thing…narrating it with our weird lingo adds a whole other layer of mystery. I have found it daunting for some of my parents. This can be good and bad. Our profession’s opaque language serves as a nice privacy layer. It can create a barrier while at the same time offering some sense of security that what they see me doing may actually have a purpose.

It’s interesting that I haven’t been challenged with explaining too much of what I dictate, which obviously could create some friction to the flow of a busy clinic day.

I forget to tell parents that I’m about to ‘dictate’

Sometimes I’ll forget to tell families that I’m about to narrate my exam for the scribe and they think I’m talking to them. This only screws up my rhythm and potentially the recorded transcript as they cut in to tell me that they have no idea what I’m saying. If I have a translator and I don’t prepare them for this part of the visit, they’ll translate the liver edge is soft, smooth, and one centimeter below the right costal margin.

This has led to a couple of weird cross-cultural moments.

Curious pre-meds

I had a 17-year-old patient interested in medical school. She wanted to understand parts of what I was saying. It opened an interesting discussion about both ambient listening and the limitations and process of a physical assessment. It was a cool conversation that I think fostered more excitement about becoming a physician than she had when she came into the room.

Similarly, I had another teen interested in AI and computer programming who had heard about ambient scribes but had never seen it in action. After I stopped my recording and the note populated, we sat in front of the computer and I shared the result as well as some of my frustrations. He was jazzed up seeing his story come to life like that. He was able to imagine how my points of friction could be solved by changes in the LLM’s code.

It’s great to help the next generation see a different future.

Not appropriate for broader exam room consumption

Findings that should be discreet could come into the open before other family members. I’ve had to intentionally navigate some awkward situations here. I assess a lot of children with significant malnutrition and I have to admit that I’ve subtly attenuated some of my language in front of families. This may be the only downside or risk that I’ve assumed since starting this. In most cases the terminology that we use is fairly cryptic which allows it to go ‘unnoticed.’ But not always.

Of course, I can type things manually after the note populates but as you may understand, at the end of a full day it’s possible that I could forget to add these.

I don’t proofread my exam

I used to proofread the exam but The Machine (my term for any technology that does things for us) captures it pretty well and I find my time is better spent on polishing a good impression. This is what referring docs and the record-reading patients care about.

I’m more intentional

I’ll say that I’m more intentional about the extent of my exam. I used to have templates that I pasted automatically but the scribe documents only what I dictate which keeps things tighter, I think. And, yes, it’s keeping me honest.

Interesting times. For 30 years the physical exams of tens of thousands of patients lived quietly in my head. And now it’s free for everyone to see. Or hear.

I would love to know your experience dictating your exams.

But this week we learned that the most important elements of healthcare AI aren’t coming through hospitals. But rather through patients.

You can get the deets in this new Pew Survey report, Americans and AI 2026: Chatbots, Smart Devices and Views on Impact that examined how we’re using chatbots. Pew is the gold standard when it comes to understanding how people are using and trusting technology. I’ve relied on their analysis for years.

This report is important for anyone trying to understand medicine’s deeper cultural shifts. It’s easy to read and their clean graphs tell a nice story if your attention span doesn’t allow for reading longer reports.

Subscribe now

A few top-line findings:

• Half of American adults now say they use AI chatbots; About quarter use them every day.

• One in five adults says they use chatbots for medical advice.

• One in five uses them for diet and fitness information.

• One in ten uses them for emotional support or advice.

So increasingly the patient has already had an ‘encounter’ before the IRL visit begins.

From information to an answer

Honestly, I’m trying to process this. My entire career (except for maybe a year or two in the 90s) has involved helping parents manage what they found on Google. Years of inkjet printouts and other artifacts from the web. All of it was information edging up loosely against knowledge.

Now patients can go a level deeper. Patients can arrive with an ‘answer.’

While it may not be a revelation to most of you, this is a major shift. LLMs don’t just retrieve health information, they compresses it and turn uncertainty into a paragraph. It shapes a kind of narrative inference before I ever enter the room.

Recently, my most interesting experiences have been with the isolated parents who know how to engineer a prompt in a way that allows them to move beyond information/knowledge retrieval and closer to how clincial judgment take shape. It’s wild to see. They don’t have medical background, but they can peek inside the kimono (or in this case, the lab coat)....

This element may have been understated in the report: Pew found that 60% of American adults read those AI summaries that now have taken over the top of Google search results. This matters because search is becoming less like exploration and more like delegation. And searching patients are less often reading sources.

So it’s likely that a higher percentage of adults than Pew reports are getting their information processed through AI synthesis at the top of their search engine page. The ‘use chatbot’ findings may be wildly underestimated if we pay attention to this key stat.

Bottom line for clinicians: The information environment we’re experiencing with patients is changing. The old question was, “What did you read online?” The new question may be, “What did the machine already tell you?”

This may matter for documentation. When a patient changes a medication or restricts a diet that becomes part of the clinical story. And it may eventually become part of the medical record. In last week’s Nature NPJ Digital it was suggested that physicians could be responsible for not correcting the wrong information that comes from chatbots like ChatGPT. It’s a little crazy and worth reading. I also did an Instagram reel here.

Patients left to their own devices

Pew found that 37% of U.S. adults have a smartwatch. That means millions of people are wearing health sensors every day. Of course, we still don’t know what these numbers really mean.

FWIW, Oura and Whoop announced this week that they are doing more to connect patient’s data with their doctors. Sheltering behind ‘wellness’ status and masquerading as medical-adjacent devices, widget companies of all stripes pitch the technology that captures the data, and we as health professionals inherit the interpretation problem. This drives me batshittake crazy which led to another Instagram mini-rant.

I think the most concerning finding was the number of adults under 30 using chatbots for emotional support. The technooptimists put a happy face on this and call it ‘access’; I call it emblematic of how we’ve failed the younger generation. Fight me.

Helping patients navigate what they already know

So like search in the 90s and aughts, the public isn’t waiting for medicine to give it permission. They’re using it. And it creates a weird new reality that isn’t getting enough dialog.

So where does that leave someone like me? The work shifts from helping patients navigate discovered facts to untangling how the patient understands the story created about them.

All of this requires new habits. And two come to mind:

• Ask what the patient prompted.

• Ask what the chatbot said and what they believe.

The real story isn’t that patients are asking chatbots for medical advice. It’s that a more serious clinical encounter now begins before the physician knows it has started. And it’s really different from when patients first gave me inkjet printouts in 1996.

Thanks for reading. Happy Father’s Day to all my dad readers.

I hear this occasionally: “My doctor gave me his cell phone number so I can call him whenever I have a problem.” While some practice situations necessitate this (rural, solo practice, or concierge) in other circumstances I don’t think this is fully thought through.

Here’s why I don’t give my cell phone number to patients

While your mileage may vary, here are some of the reasons why 24/7 access is not a good idea for me and my patients:

Subscribe now

• I’m at dinner and had two glasses of wine. I shouldn’t be dispensing advice in that circumstance. And if I did, you shouldn’t want it.

• I’m in Tasmania. This happened to me while speaking in Australia. A problem arose with a patient who had my number. Let’s try Dr. V. But it’s 2:30 AM Oz time. And once I’m called, my mind’s involved. And I’m up.

• I have no idea who you are. Despite the fantasy that our doctors remember everything about us, I don’t remember everything about my patients. While I can remember a lot with questions and prompting, it’s not the best way to handle things. And here’s a tip: never trust a doctor who says he remembers everything about you.

• I can’t hear what your saying. When I’m on call I try to avoid stadiums, amphitheaters, pep rallies, pool parties, etc. When you have a problem you want someone available who is ready to hear your problem.

• I’m not in the ‘on call zone.’ Ohio family practitioner Mike Sevilla traditionally began his call weekends on X by announcing, “Let’s do this America.” Like Dr. Sevilla, most of us on call are primed to answer, admit, refer, and advise. It’s ‘our weekend.’ At home I keep Epic open and try to enter my phone notes as soon as I complete the call. I’m just in that mode of doing and documenting. In a way, I’m in the moment.

Are there time where I give my cell phone number to patients?

Absolutely. Direct contact represents a great solution for patients and me in certain cases. The classic situation is Friday afternoon when I am returning phone calls and not on for the weekend. If I’m dispensing an odd compounded medication that I know is problematic to fill, I don’t want my on-call associate to have to try to sort out what I wanted while the family waits at the pharmacy. Another case is the critical MRI at the end of the day when I’m playing phone tag with a family.

In all these cases I finish up with, ‘If you have any other questions this weekend or tonight, Dr. Chu is on call and he’s great.’ This makes it clear that I’m not accessible 24/7.

So while lots of my patients have my cell phone number, they all know our protocol for emergency calls at night and on weekends. I’m never called on my cell outside of these circumstances.

It’s important to understand that our best care doesn’t happen when we’re available all the time.

Coverage is good because it puts us in the right mindset and on-the-spot for emergencies and problems. More importantly it allows those not on the spot to disconnect and be more mindful and connected on Monday morning when they return to their patients.

Patients deserve our best attention and that doesn’t come when we’re always available.

Do you share your number with patients?

Quick programming note: I have launched a Liminal MD Instagram page for quick takes. If you use IG, I’d appreciate a look and a follow.

This week OpenAI released ChatGPT for Clinicians. This is effectively ChatGPT for doctors and advanced practitioners. It’s directly in the lane occupied by DoxGPT and OpenEvidence, so if you’ve been using either of those tools, the use case will feel familiar.

ChatGPT for Clinicians does not come with HIPAA assurance unless your organization has a business agreement in place. For most of my prompts, honestly, this isn’t a big deal. I’m not entering protected health information. I’m asking for drug dosing, differentials, background, quick refreshers, or help thinking through a clinical problem.

So what’s different here? Maybe not as much as you might think, at least not yet. OpenAI is positioning ChatGPT for Clinicians as a broader clinical solution: ie, care consults, writing and documentation, and medical research. So less “medical search” and more general-purpose clinical reasoning and productivity platform. But in its current form, it still feels closer to OE and Dox than to some new clinical operating layer.

If you use the public ChatGPT, the interface will look immediately familiar. That’s a huge advantage for OpenAI given that most physicians are chatbot naiive.

OpenEvidence was early to define this space. Doximity then dropped in with DoxGPT and has been leveraging their physician network. OpenAI’s debut may foreshadow what’s ahead. Because the real battle here is not over who gives doctors the best answer. It’s who owns the front door to clinical reasoning.

So here are a few early reads from inside the clinical day.

Subscribe now

Answers are becoming a commodity

In the short term, you’ll hear a lot about which model gives the best answer. And that matters. I depend on these tools. When I ask about a drug dose, a differential, or a workup, I want the answer to be right.

But despite the initial differences that will be screen-captured as gotchas, I suspect all three players (OE, Dox, ChatGPT) will adjust and gravitate toward a mean that makes early differences nearly negligible to the average physician user. There isn’t a lot of variation in the differential and front-line eval of a child with their first episode of acute pancreatitis. And in 2026, coming up with accurate drug dosing shouldn’t be considered disruptive.

I’m not suggesting the answer doesn’t matter. Just that the answer alone may not be enough of a differentiator for these competitors.

Some doctors are dug in. Most aren’t.

In the trenches with my peers, I’ve seen some lock-in to OpenEvidence and DoxGPT. The original OpenEvidence users are a steadfast and passionate group. There’s a kind of first-mover loyalty there. The second wave of physicians entered the clinical LLM space with DoxGPT, which carries trust for a lot of doctors.

But neither OpenEvidence nor Doximity has yet created a clear defensive moat that can hold water. OpenEvidence has real medical search cred. Doximity has built-in distribution. OpenAI’s angle is still emerging, but its consumer advantage is hard to overstate.

I don’t know what kind of uptake UpToDate is getting. I haven’t seen much use in the workrooms I frequent, which may say something or nothing.

It’s not what you say. It’s how you look

Once answers begin to converge, presentation becomes a bigger deal. This sounds superficial until you use these products in the middle of a clinical day.

I like the concise answers from DoxGPT. But there are subtleties in OpenEvidence’s interface design that are just easier on my eyes — it’s nice. I don’t want to get too far ahead of myself, but my earliest prompts into ChatGPT for Clinicians have yielded some painfully long responses. I think that matters alot but it’s really early.

I do feel there is a context challenge here that these organizations aren’t considering. I often use these tools in the hallway outside the exam room. Other times, I want to go deep on my desktop with an impossible dilemma. For me these are two very different needs that need to be addressed in design.

And once the answers start to look similar, these companies will look for other ways to claim difference. As these products try to trounce one another, expect add-ons. Maybe some of it will be useful but it could lead to performative feature bloat. When the core use case begins to commoditize, you have to get creative.

Trust is about to change

Trust was an early battle cry for the two incumbent models. Trust the evidence. Trust the physician review. etc etc. And yes, trust absolutely matters.

But I’m not sure either of these tools are wildly distrusted by users. Among the doctors using them, the question has evolved from, “Can I trust this at all?” to “How much do I trust this for this specific task?” And I think this last question is changing by the month. Or the week at this point.

Workflow, workflow, workflow

Workflow is central to value for doctors. I’ll use a lesser product if I can get what I need with less friction. That’s one of the most important truths in clinical technology, and I’m pretty sure no one in Silicon Valley appreciates it. The best product doesn’t always win. The one closest to my work usually does.

Right now, all of these tools still live outside our clinical platforms. You go somewhere else, log in, submit your question prompt, read the answer, and then figure out what to do with it. Getting that work back into Epic remains a huge pain. And yet I still use these tools because they are transformative. That’s enough value for me.

Big picture: I suspect the healthcare information ecosystem will evolve around a chatbot interface. The patient’s entry point to care will move in this direction. I work in the largest pediatric system in the United States, and I have no doubt that patients will eventually begin their child’s journey with some GPT-like product that helps them understand symptoms, navigate options, and get to the right service.

For doctors, the same thing is coming from the other direction: a clinical interface that lives among the chart, inbox, order set, history, patient, and plan. That’s where it gets interesting.

From answers to management

So I suspect that the next frontier is tighter, patient-specific guidance based on what is known about a patient and their history. For the moment, these tools are still outside-the-clinic-door quick checks. They support or confirm what I know or suspect. Drug dosing, of course, is big for me.

But I want something closer to my work: a voice-first interface that can draw from the chart and the literature while I’m face to face with the patient, without me asking.

That is where the category starts to move from answers to management.

The HealthBench advantage

One foundational advantage for OpenAI may prove to be HealthBench. HealthBench Professional is built around clinician dialogue across consults, documentation, and research. OpenAI says this makes it better at evaluating the messy, practical work of clinical reasoning and communication. We’ll see.

If HealthBench becomes more than benchmarking theater, it could matter. The real test is not whether a model can retrieve the right article or regurgitate the right guideline. It is whether it can help inside the practical chaos of clinical work.

Clinical work is not just about finding the source. It’s deciding what matters in context, explaining it, documenting it, and moving the next step forward. If OE is literature-native, OpenAI wants to become workflow-and-reasoning-native.

I’ve previously suggested that OpenAI is well positioned to create a big, comfy blanket over Epic’s chaos. That still seems plausible. But plausible is not inevitable. Healthcare has a habit of turning elegant visions into compliance theater, gutless bullshit, and worthless dashboards. And not necessarily in that order.

Whether this actually makes clinical work easier, my kids better, and my mommas happy is yet to be seen.

Who will own the front door to clinical reasoning?

The freestanding black-box answer machine for doctors may turn out to be the hula hoop phase of medical AI: briefly fascinating and fun.

These tools will get smarter. They will move from answering general questions to managing context, memory, workflow, and care navigation. The chatbot will not disappear, but the standalone answer bot probably will.

The battle will not be who can generate a better paragraph on pancreatitis. It will be who becomes the trusted cognitive layer for doctors.

I see and work around this gap every day. In fact, with nearly every encounter I actively work to manage it. But the funny thing is I didn’t really know I was managing it until just last week when I read Alexandra Sifferlin’s new book, The Elusive Body. It was here that this space between what patients think and how I think came more clear.

Central to Sifferlin’s book is what she calls the diagnosis crisis. This the reality that more patients than we like will go un-, or mis-, diagnosed during their search for an answer. She does a wonderful job telling the story of diagnosis. And along the way she captures what folks like me do every day and how hard it really is. And that’s before we pile on friction from the insurance industry.

I had the chance to sit down with Alexandra Sifferlin on the Liminal MD podcast. Our conversation is a great supplement to The Elusive Body.

You can also listen on Spotify:

<iframe frameborder=”0” height=”200” scrolling=”no” src=/__u/liminalmd.substack.com/&#8221;https://playlist.megaphone.fm?e=RLIIO1815596488&#8221; width=”100%”></iframe>

She finishes her book with a discussion of AI, which will improve parts of the diagnostic process. Pattern recognition at scale will bring real gains. But Sifferlin’s title is apt in ways that go beyond what diagnostic tools can fix. The body is elusive not just because our instruments are imperfect, but because illness is deeply embedded in a person. And what we’re after isn’t always visible in patterns of symptoms and signs.

The expectation gap will outlast whatever we build to bridge it. Narrowing the gap requires us to admit that we’re navigating uncertainty together. That’s where the encounter should start.

That caption belongs to a woman on TikTok, sharing her lab results with whoever happened to be watching. It’s the subject of a recent New York Times article (gift link), and the comments that followed tell you everything you need to know about where medical interpretation has landed.

“Your blood work is amazing. Especially your A1c and cholesterol.” Someone else offered concern about the absence of a primary care physician. Her response: “Not all of us are that blessed.”

For those not in the clinical space, the 21st Century Cures Act passed in 2016 required that patients receive unmediated access to their results the moment they were finalized.

But it seems the Cures Act has spawned a new social genre: Videos of people opening scan results, reading labs, reacting in real time to information they may or may not be equipped to understand.

Of course, some of this is performance. The algorithms reward emotional revelation and creators know it. But that doesn’t make the anxiety less real, and it doesn’t change what the genre tells us.

Subscribe now

What this genre exposes is the abolition of a clinical function. The function that was abolished is interpretive translation: the act of taking a result and locating it inside a specific person’s history, risk profile, and life circumstances. That function requires context that the comment section can’t supply.

“Your A1c and cholesterol are amazing” is not interpretation. It’s well-meaning pattern recognition. And it’s not that it isn’t worth anything. But it’s categorically different from what it’s replacing.

None of what I’m discussing here is a technology problem. The Cures Act and the portals didn’t create it. But what this did do is lay bare how little infrastructure exists to help patients understand what they’re looking at.

————

Medicine made a trade: transparency for interpretation. It was framed as a patient rights victory, and in many ways it was. Patients deserve their data. But the interpretive layer didn’t migrate with the law. It stayed behind. And now people are opening their scan results alone in their backyards, on a beautiful day, with their cameras on, because there’s nowhere else to take what they’re feeling.

The 21st Century Cures Act didn’t abolish access. We’d been losing that decades before. What we abolished was the pause.​​​​​​​​​​​​​​​​

The films were dense and smooth and real. Acetate film had a feel and odor uniquely its own.

And each was marked with a crooked lead label that indelibly proved it belonged to that patient. Every well-trained physician would match the label to the case to make sure they were reviewing the right film.

For most of my career medicine was about things we could see and hold.

Subscribe now

—————

Fast forward to 2026. Our medical journals are publishing studies about things that aren’t real: Deep fake radiographs. Images created by prompt. An RSNA study published last week, The Rise of Deepfake Medical Imaging: Radiologists’ Diagnostic Accuracy in Detecting ChatGPT-generated Radiographs (free), found that radiologists were able to correctly identify fake radiographs only about 75% of the time. Even LLMs, the very agents that fabricate them, struggle to identify the bogus images.

The world I’ve worked in always relied on a tacit agreement that the things we’re dealing with are real. This agreement was never discussed because it didn’t need to be. Why would something in the medical record not be real?

This study points to a vulnerability deeper than radiology. Think path slides, EKGs, endoscopic images, or anything image-based.

—————

In medicine there are things like a scalpel, hospital gurney, or a patient’s hand. These are the objects of medicine. And then there are the growing number of artifacts that live in the aether. They are the triggers of experience.

And these things we can touch are the building blocks of reality. “Objects stabilize life,” wrote Swiss-German philosopher Byung-Chul Han. “The digital order deobjectifies the world by rendering it information.” Something happens to knowledge when the object disappears.

A film was always part of that reality. I could hold it in my hand. I could even verify its relative recency based on its odor.

Medicine built the diagnostic apparatus on the assumption that images are evidence. Last week that assumption dissolved.

In fact, when you listen to those making these tools there’s a manic sense that we’re approaching a kind of escape velocity. Our shiniest, best days are just around the corner.

In healthcare the promise of the inflection point is that technology and computation will finally fix our problems. This solutionist mindset imagines not only making the system smoother, but smashing it to smithereens for a fresh start.

James Bridle in New Dark Age - Technology and the End of the Future labeled this as accelerationism:

Acceleration itself is one of the bywords of the age. In the last couple of decades, a variety of theorists have put forward versions of accelerationist thought, advocating that technological processes perceived to be damaging society should not be opposed, but should be sped up - either to be commandeered and repurposed for socially beneficial ends, or simply to destroy the current order.

Last week Graham Walker dropped a great essay on the realities of abundance and the challenges it will lay bare in patient care. Technology, he suggests, will get faster and faster, but the touch points that involve judgment and synthesis will get choked.

I got to thinking about these non-negotiable choke points in human care. These are sometimes called bottlenecks — a kind of pejorative term for the things that throttle Silicon Valley’s Promethean vision of escape velocity.

Subscribe now

Physical stuff

Let’s start with the most basic reality: Care for humans will always be grounded in the body itself (and mind). And no matter what the longevity folks sell us, or how well we optimize the system around the body, we will always heal at a neanderthal rate. Flesh is the ultimate bottleneck.

Beyond our ability to heal ourselves, there are things that need to be done physically — in person, on a gurney, in an OR, or wherever. Think IVs, infusions, operations, and the like. A 10-day course of oral antibiotics is physical and it definitely doesn’t scale.

Telemedicine showed us that when it comes to human bodies with lacerations and palpable masses, there’s a limit to what can be delivered virtually. So as long as we’re in the business of caring for humans, we’re restricted by the physical embodiment of the person.

Judgment

Given the wild variability of the human mind there will always need to be some level of orchestration that involves judgment.

Judgment is contextual and relational. It requires knowing this patient, not the distribution of patients who look like this patient. It takes what’s unspoken (grief, fear, ambivalence) and weighs it against what’s been measured. It considers that the patient’s stated preference and their deeper interests are often not the same thing.

I’m not sure where this bottleneck gets resolved. Probably during some kind of synchronous human-to-human exchange at select points in the patient journey. Or as we used to say, a visit.

Trust and legitimacy

Knowledge being nearly free and accessible doesn’t mean everyone’s going to believe it. And when it comes to health technology we bring all kinds of baggage with us. Part of this baggage is fear, doubt, and suspicion.

The diagnosis of a simple problem by an LLM, for example, may be spot on but still not leave someone empowered to act. The bottleneck may be less about information and more about the credentialed background and baked-in human approval that makes a recommendation actionable for a patient. This isn’t judgment as much as faith in the machine.

It’s worth noting that people can distrust an algorithm even when it outperforms a doctor. This has been called algorithmic aversion. A single error can blemish a solid track record of accuracy. What it means is that human adoption of AI health tools may not follow performance in a straight line. I suspect credentialed doctors will likely carry the diagnostic burden well after it’s technically necessary.

—————

I can imagine a near-future entry point to the healthcare system that takes place entirely through a reliable, legitimized AI ... something that intercepts and treats the simple things and escalates the sticky stuff to some kind of interpretive core where the important decisions are negotiated. Or places where physical intervention is done.

In this scenario the human moments become exceptions rather than the rule. Which is okay. It offloads the swipe right things to the machine while allowing us to leverage our greatest skill as guides in the healthcare journey.

All of this is part of an imagined system with unlimited access to medical knowledge, perhaps diagnostics, but rationed access to human judgment and wisdom.

If anyone’s interested, I’ve been doing my illustrations by hand on Procreate. I’m having as much fun sketching as I am writing….

I’m not going to make this kid’s digital footprint any wider by linking out. And I won’t try to browbeat the professional educators at Mayo. They know what they’re doing. If you want to see it, y’all are industrious enough to dig it up.

I do have some thoughts on the matter, however.

Subscribe now

And I have some standing here. A few years back I was appointed Director of Digital Professionalism at Baylor College of Medicine. I did all the things: built a four-year curriculum, shaped policy, trained staff and post-grads, etc. Along the way I handled my share of digital deviants. It was a ride, for sure.

So when I say I’ll offer a more measured perspective than the ranting Instagram influencers, I mean that less as a boast and more as a promise.

————

What’s interesting is we don’t see this kind of professional transgression like we once did. At the dawn of the internet and the early days of social media, doctors stumbled as they figured out what to do in the new public arena. Less so now. I suspect that fledgling users have more voices and power users to watch and emulate.

As the number of public physicians rose, we professionally went into a weird, paranoid period. When Twitter was in its early meteoric rise there were obsessive concerns about what was “professional” and “unprofessional.” In fact, there’s a rich body of early medical literature in journals like JAMA that did things like plot foul language over time by medical professionals. Today those papers read like something from The Onion. Once all the doctors began cursing, the studies became futile and the pearl clutching subsided.

I suspect those building careers in professionalism didn’t want to see the data.

————

The system has a responsibility to educate trainees on how to carry themselves in very public places. Managing a digital presence as a professional is different from managing one as a frat boy. When trainees fail, it can (not always) reflect our failure as educators. I don’t know this young man’s history, and how Mayo handles it will depend on factors opaque to all of us. As it should be.

That said, a tight curriculum won’t fix stupid. There will always be those with judgment so bad that a good education won’t prevent disasters. This is true in all aspects of medicine.

Intent matters too, and here’s where this case gets harder to excuse. Many problems physicians face online happen because of the speed of digital. There are those moments where we’re moving too fast, share something impulsive, and realize afterward it wasn’t a good idea. We’ve all done this. But those narrow-frame lapses are different from conceiving, filming, editing, posting, and promoting a video clip. This is a process with plenty of steps to reconsider what you’re doing.

One of the persistent blind spots for new trainees is context collapse. We calibrate our behavior to our context … how you carry yourself in Vegas with college friends is different from how you handle yourself in church. Most of us learn this code early. What’s invisible to many trainees is that public digital venues mash up all those contexts simultaneously. In one social space you might find your former fraternity brothers, your rabbi, your medical school dean, and your patients. You simply can’t perform for all of them at once.

And even when intent is good, perception trumps reality. Gallows humor lands differently in public than it does in the OR lounge. If you’re unsure, get a gut check from a colleague before you post.

One more thing worth saying: you probably think you have no digital footprint. Do a vanity search on Google or ChatGPT. You’ll see. The only digital properties you fully control are the ones you build — LinkedIn, Doximity, your hospital bio. Articles, news reports, and conversations someone has about you are completely out of your hands. Which is all the more reason to manage what you can.

————

At the end of the day, being a doctor is a privilege. The tension between individual expression and responsibility to community remains one of the core challenges of the modern physician.

And my role at Baylor? I trained myself out of a job. Founded in an unfounded fear, it didn’t bear out for a full-time salary. Texas Children’s The Woodlands was under construction, and I got pulled into one of my first leadership stints. That’s behind me now. But apparently it’s not far enough behind me that I don’t still have opinions about it.

For fun you can check out an online resource that I made a few years back — The Public Physician. Some of it’s dated but kind of interesting.

————

I promise in the next few weeks to send out a graphic showing the history of docs online.

I was able to grab Dr. Lantos for a conversation:

When I spent time in the Indian Health Service with the Zuni Pueblo I witnessed the aura in the medicine man. It was a vibe created in the eye of the Zuni patient fulfilled by the presence of the sage. In some cases, blowing wafts of incensed smoke into the tribal member’s face added to the mystique. His aura materialized in healer presence, patient hope and the fulfillment of expectations.

Early on I believed that doctors had some ethereal capacity. Scientifically trained medicine men with a powerful mashup of deep knowledge and wisdom. In junior high school I was convinced that the doctor could identify my most embarrassing personal habits and shortcomings just by physical exam. For a period it was terrifying.

For those who believe, there are things that have changed to lift the physician aura.

Subscribe now

1. Medicine is transactional

The aura was historically concentrated on one person, your doctor. And continuity made that possible. But shift work hospitalists, specialist-driven care, and care teams distributed the relationship across so many people that no single clinician could accumulate the trust and presence that aura requires. It’s hard to conjure a solid aura during a 15-minute visit with someone you’ll never see again.

And as Lantos lays, the objectification of the patient that began over a century ago was the beginning of the end. When we diagnose based on markers and scans, you’ve got a transaction.

2. Information is democratized

The monopoly on knowledge that defined the physician began to dissolve in the 90’s. It began with the internet and has evolved now to include access to some loose version of clinical reasoning through LLMs.

I think it was in the early 2000’s at Health 2.0 in Boston that I heard patient advocate declare that the priesthood was dead. This mindset during the peak of the patient empowerment movement came from breaking the longstanding sense of exclusion that came with a monopoly on information and decision making that often excluded the patient.

3. Medicine is transparent

The physician aura that most of us understand is dependent on opacity. But as Walter Benjamin (the philosopher cited in Lantos’ JAMA editorial) suggests, the reproduction of a work of art risks exposing the fallibility of the original piece. But traditionally when things went wrong in the hospital, they often stayed inside the profession. Our imperfections had no real platform for amplification.

The patient safety movement (malpractice culture, public reporting of outcomes, To Err is Human) systematically exposed medicine’s fallibility at scale. And it’s hard to brand yourself around mystique when your complication rates are on a dashboard.

Of course, there are benefits that come from an unbridled trust in another person. Think placebos. There are things we don’t understand that arise from faith and hope, be it in God or through the hands of a tribal healer.

Alternatively, there are the costs of an aura like paternalism and opacity of the facts.

******

The involution of the physician aura is not a judgment or declaration of failure. Technology that can synthesize information into knowledge is outsourcing some of what we once did. This is just a reality.

As far as the future of the aura, we have to sort out where we figure in. We have to identify a role integrated with tools that now do something close to what we once did.

In any field a skilled sense maker and translator will always carry a certain aura. This docent-like role for us could be someone with a unique understanding of how things might be connected and what they could mean.

No matter how advanced things become, healthcare will always be rooted in the care of the physical human. Unlike our digital thinking partners, humans are imperfect and weird and needy. And the journey we go through calls for someone who can bear witness with us and lend some level of wisdom. Someone to weigh the things an LLM may reflect but never understand.

Ultimately, I suspect we’ll evolve into a new aura.

Maybe you’ve been to a similar meeting.

Subscribe now

Through my career I’ve noticed something I couldn’t quite name — a creeping sense that something in medicine had shifted, something to do with how we talk about value. Then I read an essay in Aeon by the philosopher Julian Baggini, and he gave it a name: instrumentalisation. This is the tendency to strip things that are valuable in themselves of their intrinsic worth, and reduce them to means toward some further end. Medicine has fully absorbed this logic. We can’t defend the value of anything that doesn’t plot on a graph.

Once I heard this word, it described what I was seeing.

As an example, in the last couple of decades, we started talking about rapport with patients as though its value depends on outcome data. You’ve seen the framing: a good therapeutic relationship correlates with better adherence. Empathy improves clinical outcomes.

I understand why this language exists. In an instrumentalist system you have to translate everything into its terms or risk it being dismissed as soft or unserious. But the translation comes at a cost. When you justify being good to your patients because it correlates with better outcomes, you’ve accepted that the relationship needs extrinsic justification. Kindness, on its own terms, apparently isn’t enough.

But do we really need an NIH-funded study to tell us that patients do better when their doctor gives a damn? What does it say about us that the study was considered necessary?

This kind of thing is also found in empathy training and research. As if empathy is a skill you can acquire in a workshop and deploy in twelve-minute increments. The instrumentalist frame turns what should be a basic feature of human engagement (caring about the patient in front of you) into a technique with an evidence base and a revenue cycle implication.

*****

The second part of this is that once we decide that only measurable things matter, the measures then take over. This leads to a kind of next-level instrumentalisation.

Consider the A1C. It was developed as a proxy for glycemic control—a useful one, I’ll add. A window into what glucose has been doing over the past few months, aggregated into a single number that could guide treatment. But somewhere along the way, the proxy became the goal. We stopped asking whether a patient’s glucose control was serving their life and started asking whether their A1C was at target.

This distinction matters. Patients with erratic glucose, frequent hypoglycemia, or limited life expectancy sometimes ended up with tightly controlled A1Cs and worse lives. The ACCORD trial offered a hard warning about what happens when you chase a number past the point where it still tracks the thing you actually care about. A mortality signal, in a study designed to show benefit. No one disputes that glycemic control matters. What the instrumentalisation of diabetes does is sever the instrument from the outcome, so that hitting the number feels like the same thing as helping the patient. It isn’t always.

Patient satisfaction scores followed a similar arc. The original intent was right: capture whether patients felt heard, informed, respected. I’d argue that this is genuinely important, and not because it correlates with anything. But as we all watched, the score became the target. Physicians started getting graded and compensated on numbers that had no reliable correlation with clinical quality. And the blind pursuit of those measures created perverse incentives, like giving patients what they want, not what they need. The opioid crisis has a satisfaction score subplot that we still haven’t fully reckoned with.

The patient satisfaction cottage industry didn’t cause the epidemic, but the instrumental logic that made satisfaction a compensation metric helped create a world where saying no to a patient carried a professional cost.

*****

There’s an old philosophical distinction between intrinsic and extrinsic value that medicine used to understand, at least implicitly. Some things are valuable because they lead to something else. And some things are valuable in themselves. A good conversation with a patient where something is understood that wasn’t understood before has intrinsic value for both doc and patient.

I want to be clear that the problem isn’t measurement. I want proven benefits for the interventions I prescribe, and I trust the apparatus of evidence that helps me sort what works from what doesn’t. The problem is what happens when measurement becomes the only language we have for value.

I’m concerned that we’ve lost the language to defend things that matter but can’t be measured.

Good morning,

I’m trying to get a sense of who’s reading this newsletter and what you want more of. This survey takes about a minute. Please take a moment to help me learn more about you. I can’t get better unless I hear from you.

Start Survey

As always, thanks for being a subscriber and I look forward to hearing what you think.

Bryan

Slowly but surely, we’re entering a world where every discussion and interaction we have will be part of a permanent intelligent memory that is always accessible and unlocks all kinds of insights and connections.

I came on this post (via the unflappable Spencer Dorn) as I’ve been hammering out my relationship with an AI scribe. Sorting that out and reading Belsky’s idea has me seeing the future of the medical record.

Let me unpack two elements of Belsky’s argument about summarized living that feels relevant to medicine.

Subscribe now

1. Summarized living will remove bias from work and life

Can you imagine how many experiences in daily life — from employment interviews and performance reviews to team meetings and spousal debates (uh oh) — are riddled with biased and bad recollections? Perhaps relying on our natural memory will someday become unthinkable, much like navigating the world in the era before anything was written down.

This is spot-on for me.

Reading the AI summaries of my patient encounters has made it clear to me that I’m unable to recall some of the details even if I sat myself down in the workroom immediately after the visit. Maybe more importantly, my 24-year-old self probably wouldn’t do any better.

Belsky makes the point that when we compare what we remember to a factual record of what actually happened we can begin to recognize our own biases.

2. We will converse differently for better summarization

I’m already doing this, although I didn’t expect to.

I’ve fulfilled the promise of the AI scribe by pushing my keyboard aside and simply talking to patients. But the conversation isn’t what it used to be. I find myself asking questions differently, deliberately summarizing aloud, and subtly nudging parents toward phrasing that I know will translate cleanly. I am, in subtle ways, shaping the encounter for the summary.

As I steer the dialogue toward something the system can digest, I feel a kind of complicity with the scribe.

This isn’t a criticism. The scribe just isn’t that smart. I think of our scribe as like a good sub-intern from a mid-tier medical school who thinks she’s a little better than she actually is. So I have to gently help her along.

This kind of behavior is seen in the early iterations of new technology where we compensate for what it can’t do. This has been called fauxtomation. This is when humans quietly cover for what automation cannot yet do.

He says some other stuff in his essay, but I’ll leave those for the moment.

A renewed focus on the impression

So how has AI enabled summarization tech changed where I spend my time?

True to what’s been shown in scribe studies, summarization technology creates more satisfaction than bottom-line time savings. For me, the benefit not captured in these value studies is a better note because I’m able to prioritize what referring doctors want when they send a patient my direction.

And some of this relates to the kind of work I do in my north Houston healthcare ecosystem, so your mileage may vary. But my day job centers on 1) identifying unknowns, and 2) managing chronic disease.

For example, accounting for an asymptomatic preteen with dramatic weight loss and is not just a data problem. It’s physiology, psychology, maternal anxiety, and referral expectations glomed together.

So, the impression is where I work this through. This involves not only the likely pathophys, but the outcome of shared decision making and the thick layer of psychosocial goo around mothers and their ailing children.

While I’ve always found the impression to be my most important deliverable, I now have the bandwidth for really nailing the impression. And this is what I’m trying to do

I do it with straight dictation — I don’t use the scribe software. I also now put my impression at the top of my note.

Summary and synthesis | The two parts of the record

Looking ahead, I see two distinct modes emerging in the medical record: summary and synthesis.

Summary is what AI does well. And this is what a scribe does. It compresses, extracts, and reduces. It takes a meandering history from an exhausted mother and makes it legible. This is the creation of signal.

But summary won’t stop with the history.

The EHR is evolving as the exhaust pipe of the human health experience: labs, genomics, sequential imaging over decades, message threads, device data, ambient transcripts, etc. So, as our system collapses under its own data weight, information summarization will become critical. Belsky’s summarization is the only viable response to that kind of scale.

Synthesis, by contrast, is totally different.

Synthesis is what I do in the impression. It is where all of the far reaching stuff is integrated into meaning and direction. For that boy with unexplained weight loss, it’s the integration of physiology, family dynamics, probability, ambiguous celiac serology, and risk tolerance into a coherent approach. Synthesis is less about what happened and more about what it means — and what we are going to do.

As the EHR evolves into a clinical operating system, summary may become its dominant function. And synthesis will be its interpretive core.

The rise of the interpretive core

Until we have fully autonomous clinical systems (ie, end-to-end clinical experiences navigated with bots and agents, independent of human involvement), there will be a uniquely human interpretive core.

Sure, AI can simulate synthesis rhetorically. And synthesis by humans can involve the help of smart machines. That’s good. It can surface associations, challenge anchoring, and broaden a differential.

And for the foreseeable future, the physician owns the consequences and the interpretive core.

Bred in a system where handlers beat the individuality out of young skaters, Liu defied everything to defend her passion for movement, energy, and expression. She knew the only thing she could control was her process and mindset. And with seemingly no regard for outcome, she captured gold and the heart of the world

To understand this, go further and watch her competitors. Many were tense and consumed with either 1) the things that could go wrong, or worse, 2) the fear of never seeing the podium.

What made Liu’s appearance so arresting is how rare something like this is to see. When someone operates with this level of passion unencumbered by the outcome of what they’re doing, it draws the attention of the world.

Subscribe now

So how does this relate to medicine?

I got to thinking that individual expression may be dying/dead in medicine. Not because we lack passion, but because we’re part of a system that demands compliance over craft. Physicians are central to the industrial success of every health system’s operation.

We’re sort of like Liu’s competition.

It starts early for us. Raw compliance is the foundation of medical training. We goose-step from pre-med to residency avoiding any move that could fail. This carries into medical practice where, after the first plaintiff demand letter, we lean into defensive care. At that point, after becoming conditioned to chase any number of end-games, we find ourselves entrenched in a system rigged to make the outcomes out of our control.

And the problem isn’t the outcomes; they’re important. It’s just that they’ve become central to our professional identity. When our worth is tethered to standardized surveys and productivity dashboards, the craft is hard to keep alive.

But can we find Alysa Liu’s brand of energy and individual expression in medicine?

I think so. Some of it depends on your frame of reference.

Finding medicine’s rare, passionate voices

Medicine has spawned many figures passionately anchored to their work. Oliver Sacks writing about his patients comes to mind. The seriousness of neurology didn’t erase his sense of wonder. Paul Kalanithi, author of When Breath Becomes Air, while terminally ill continued to write and operate when possible. He had a unicorn devotion to craft and meaning that most of us will never grasp. Closer to home, Will Flanary (alias Glaucomflecken) has gone all in with the creation of a cast of Instagram characters that pull back the layers of medicine’s weird culture and expose healthcare’s deep dysfunction.

I’m fortunate to have had heroes during my training who brought this level of passion into my orbit. None celebrities. Some clinical teachers in community hospitals in spaces far removed from The Medical Center.

Exploring medicine’s corollary to Liu’s brand of skating is an important thought experiment for all of us.

We can only hope we’re courageous enough to identify our purpose and act on it as she did.

I put together a list of (somewhat organized) thoughts around what this tells us.

1. Technology is doing things we didn’t think possible

I know this is obvious, but deep fakes demonstrate how far this crazy technology is pushing things. But constant change will soon be the norm in medicine. This may be more disarming for a late-career physician like me than a Gen-Z doc. Or maybe not.

Subscribe now

2. When we start saying ‘embodied experience,’ we’re in deep

We’ve hit a point where embodied experience is being used in reference to medicine. Embodied experience refers to the concept that bodily presence and physical interactions with the world are fundamental to how we understand and engage with our environment.

Of course, medicine has always been rooted in embodied experiences. Digital systems, including deep fakes, operate by minimizing or eliminating them.

3. There’s value in scarcity

The question isn’t how we prevent impersonation. It’s whether the embodied encounter regains value precisely because it cannot be faked. As these deep fakes and surrogate things like ChatGPT Health proliferate, embodied medicine will become premium. In a synthetic environment, reality starts to carry weight.

of The Convivial Society was ahead of his time on Twitter in January of 2023 when he wrote:

Maybe the potential upshot of the emerging synthetic media environment will be that we are actually thrown back on embodied experience as the focal point of our sensemaking.

If you’re a regular reader of Liminal MD, you know I’m big on the emerging value of things that don’t scale.

4. The physician voice is a commodity

Let this sink in: Your vocal pattern is now trainable data. And tone, cadence, and inflection are part of what patients interpret as credibility.

Oh, and don’t forget that ambient listening and scribing systems already learn from your every move/click/plan.

The physician voice is digitally extractable.

5. The human face is the last unsecured credential

Through most of medicine’s history, physician identity was anchored to our presence. You walk into the room, and the patient sees you and hears you. That visual confirmation has functioned as a kind of authentication.

Sure, we’ve built elaborate systems around licensure and credentialing. But at the point of service, the patient trusts the face (and the stethoscope?). When the clinician’s face and expressions can be faked, this credential dissolves.

Of course, this shift is bigger than just medicine. But in medicine we feel it because our roles and work are grounded in trust.

Soon your virtual credibility may require third-party cryptographic validation. The face that once anchored trust becomes an unsecured interface until cleared by code.

That’s all I’ll say about solutions. They’re way over my head.

6. We’ve built a culture where simulation is good enough

For me, the uncomfortable part isn’t that deception exists. It’s that simulated doctors may be fine for a lot of folks. Because if a fabricated physician can reassure or influence behavior, then the performance of authority is doing the work.

In my recent Liminal MD podcast episode with Dr. Charlotte Blease, I challenged her on whether we’d ever want to see end-of-life discussions happen with highly tuned chatbots. She suggested that this should be a patient’s choice. I can’t disagree. But when we choose to talk to a chatbot during our final hours, something’s going on.

Nearly a decade ago Sherry Turkle in Alone Together called this our robotic moment:

It refers to our state of emotional—and I would say philosophical—readiness. I find people willing to seriously consider robots not only as pets but as potential friends, confidants, and even romantic partners. We don’t seem to care what these artificial intelligences “know” or “understand” of the human moments we might “share” with them. At the robotic moment, the performance of connection seems connection enough. We are poised to attach to the inanimate without prejudice. The phrase “technological promiscuity” comes to mind.

The human core in patient care may already be thinner than we think.

And this didn’t begin with AI.

We’ve been normalizing mediation for some time now. Telehealth made physical absence routine during COVID. Patient portals have made asynchronous care good enough. Chatbots pull us even further from relational care. And each of these tech-mediated steps is framed around convenience and access. All true. But each step has also moved medicine further from IRL care and closer to an interface. A deep fake doc is really just an extrapolation of these things.

To be clear, none of this is a repudiation of technology. We have to remember that whenever we mediate a patient encounter we lose something, even if small. And yes, it cuts both ways. We gain things as well.

Alot of this is about shifting expectations. In a health system designed for scale, good enough is becoming an acceptable standard. And now we may have landed on a version of medicine where being convincingly doctor-like can substitute for being a doctor.

Once that new standard hardens, human authenticity becomes optional.

7. We still trust doctors

For all this craziness there’s something reassuring here: When scammers want to borrow your face it means you still have something going on. And it seems consumers still pay attention to doctors (or things that look like doctors).

And never forget that humans love a narrator. Just something to keep in mind.

The answer was sobering.

And I have to say up front that this is one of the most provocative studies around patients and AI that I’ve read in a while. You should check it out. I’ve had so many conversations with patients about what they are finding on ChatGPT, and I’ve been trying to understand where it falls short. This study helped me see it more clearly.

What’s interesting is that it lays bare the things I actually do in an exam room. Technology teaches us things about ourselves and how we work. LLMs and the challenge of public problem solving are a nice example of this.

Participants were given detailed medical scenarios and asked to use commercially available chatbots to decide what to do next. Call an ambulance? Stay home? Schedule an appointment? They were also asked to identify the likely diagnosis. The results were compared to a control group who used whatever they would normally use at home, which meant Google.

Bottom line: The chatbots did no better than Google. And the authors concluded that none of the models tested were ready for deployment in direct patient care.

Participants chose the correct course of action less than half the time. They identified the correct diagnosis only about a third of the time.

This clashes with the prevailing public narrative. These systems pass licensing exams and outperform poor slobs like me on carefully constructed diagnostic scenarios. By every benchmark, they look remarkable.

And they are remarkable in many ways. But for those of us who do this for a living, we know that good medicine is nothing like answering board questions.

So what happened? Here are three key thoughts/takeaways.

⸻

1. The failure is translation, not intelligence

This is critical: When researchers fed the full medical vignette directly into the model, accuracy jumped to 94 percent. The system could reason through the case when all relevant information was present. But when participants interacted naturally, leaving out key details or describing symptoms imprecisely, performance tanked.

There were no truer words than from Nassim Taleb on Twitter in 2023:

An expert is someone who knows what not to be wrong about.

And in our line of work knowing what not to be wrong about is everything.

That’s why doctors spend years learning which details matter. We are trained to detect the salient features of a patient’s story. Location, timing, severity, modifiers, aggravating and alleviating factors, associated symptoms, etc. We know that worst headache of my life is different from terrible headache. We know that sudden onset carries weight. We know that neck stiffness and photophobia can be ominous in the right context. And when discussing febrile infants, the word irritable gets taken very seriously by smart pediatricians.

Patients don’t know this. And why would they?

Subscribe now

⸻

2. Small words change big outcomes

This study exposes a structural reality of generative AI: large language models are really sensitive to input framing.

Half of the observed errors in this study were attributed to users omitting relevant details. If a system is designed for public use (which ChatGPT is not, in its current iteration, IMHO), it can’t assume perfect prompting. It has to compensate for lousy storytelling.

Because if a model is dependent on the perfect input, it is not doing anything even remotely related to what a doctor does. For example, I create mini hypotheses during my history and I circle back and probe repeatedly.

Another finding deserves attention. Even when given complete information, the models struggled to consistently distinguish between urgent and non-urgent scenarios. Central to triage is the ability to calibrate an appropriate response. Not just to recognize a condition, but to determine how fast to move and how worried to be.

⸻

3. Knowledge is not judgment

It is important to be clear about what this study doesn’t show. It doesn’t show that large language models are useless in medicine. It doesn’t negate their utility for clinicians or even patients. It doesn’t refute the possibility that future iterations will improve.

What it shows is that intelligence alone is insufficient for what matters most.

What we are watching is the collision between answer engines and messy human input. The models are good at churning outputs when the inputs are structured and complete. But as we see in this study, patients do not give us structured and complete. They jump around, tell stories, and bring their own biases to the exam room. Or in this case, the prompt field.

The bridge between bad questions and helpful answers is judgment.

And judgment involves more than retrieving the correct differential. It involves calibrating risk and escalating when needed.

The real challenge for AI health leaders is building systems that understand how decisions are made and when uncertainty should trigger escalation. That requires deep engagement with the people like us who do this work daily.

⸻

For now, patients using chatbots for medical advice are navigating a powerful but clearly unstable interface. When I cut my teeth during Web 2.0, I argued that physicians have an obligation to help patients understand the limitations of search. It seems what’s old is new again.

Until AI systems can operate responsibly and consistently within the space of human ambiguity, patients need to understand the difference between intelligence and wisdom.

You can read the full study here. It’s open access, the graphics are excellent, and the design is easy to follow. As I like to say, don’t trust me. Read it yourself.

AI slop is a sort of liquid nothingness that will pour through culture, searching for whatever’s fungible and displacing it.

Searching for whatever’s fungible and displacing it.

This resonated because medicine has been slowly making me fungible for years.

It’s no secret that I can be swapped out of the clinic schedule with another pediatric gastroenterologist and not a lot will change. The clinic will run and the billing will clear. Sure, there’s unique value that I deliver; the problem is that what I bring isn’t valued quite enough to make me irreplaceable. My skills as a master-level clinician might be acknowledged, but it ends there.

Because as a clinician I’m interchangeable. That’s what fungible means.

Operational interchangeability is the first step toward cognitive interchangeability.

And this makes sense. Physician fungibility is critical to the operational success of industrialized care. It’s baked into our time-boxed visits, standardized orders, and guideline-constrained decisions. 21st century physicians like me are defined by productivity, throughput, capacity, and coverage. Healthcare is a product. And we deliver it.

AI will replace what we’ve commoditized

AI is perfectly positioned to work inside of an industrialized model of care delivery.

So circling back to Carr, this is his point: AI will displace whatever we’ve already decided is interchangeable. Applied to medicine, the decision has all but been made. Once a role is defined as standardized throughput, it’s only a matter of time before someone asks whether software can deliver the same output.

And if you think your system wouldn’t replace you with an advanced LLM, you’re living in an alternate universe.

So have we decided that doctors are fungible? I think we have. But the decision is reversible.

This isn’t about what technology can do. It’s about what’s valued in healthcare.

Subscribe now

The linchpin doctor

Let me take a quick diversion on an exception that I call the linchpin doctor. These are those of us who deliver something truly unique — a unicorn skill. Every hospital needs them. It could be a technical skill or an emotional intel superpower that moves things in a way that few others can. Some docs pull it all together in a way that’s imperceivable and almost impossible to replicate. We’ve all worked with them.

I originally heard about this concept in Seth Godin’s excellent book Linchpin — What Makes You Indispensible?. For young professionals starting out in medicine, I recommend it. For fun, this is what I wrote about this book 16 years ago 👀 — my ideas have evolved a little.

But linchpin docs are the exception.

Tech determinism and docs that deserve to be replaced

I suspect I’ll get some iteration of this perverted epithet in the comments: Any doctor who can be replaced by AI deserves to be replaced.

This idea is rooted in a mindset of raw technological determinism. It sees illness as a sequence of inevitable transactions. It assumes machines will do your work if the technology is there. There really is no choice. Silicon Valley has spoken, coded, and delivered.

And sure, transactional care will (and probably should be) offset to LLMs and the like. Diagnostics will continue to be turned over to smart machines. In fact, we conceded years ago that a CT can identify acute appendicitis in a child better than your physical exam. So we’re already ceded parts of medicine to machines. The question is where we stop.

Bedtime stories and hospice conversations

But let’s look at values, tech determinism, and your child. Or your dying mother.

As an example, consider the release of a bot that can perfectly mimic your voice and read bedtime stories to your toddler. As the machine does this frighteningly well, you are free to be productive catching up on email.

You might argue that bedtime stories are different from clinical work. But both involve more than transaction.

Talk to any parent, no matter how busy, and almost universally they cherish this experience with their children. They value this time more than the convenience. Most will make the decision to keep this kind of experience no matter what technology is available.

Or take your mother struggling to make sense of her life in the final days of a battle with cancer. A Silicon Valley startup has a therapeutic hospice bot that can carry on these conversations. But is the management of your mother’s existential moment something we want to turn over to an LLM?

Maybe. But it depends on what you value.

Human determinism and the confrontation of our values

Nearly a decade ago Sherry Turkle in Alone Together — Why We Expect More From Technology and Less From Each Other wrote that every new technology forces us to confront our values. And this is precisely where we’re at.

By human determinism I mean this: the deliberate choice to preserve human roles even when automation is possible. This next generation of hospital administrators, doctors, and even patients will struggle with deciding what we think is important.

What hospitals and patients need from doctors will change. And how we are valued will change. We are in a major sorting out period right now. I suspect human values will override the inevitability of technology in select cases.

I have no idea what our days will look like but they will be very different.

Right now we need to draw lines where we believe they should be drawn. Both as patients and professionals we need to bring a level of human determinism to where it’s needed.

My advice: As machines get smarter, look for the thing that’s scarce.

Do we have an obligation to replace docs with machines?

In her new book Dr. Bot: Why Doctors Can Fail Us and How AI Could Save Lives, Dr. builds a serious case for the moral obligation to automate large parts of what we do (Hold your rage until you read it). You can listen to my conversation with her on The Liminal MD Podcast. And please subscribe where ever you listen.

But here’s the harder question.

If what we deliver is standardized, protocol-bound care, and if we define our value in RVUs and coverage grids, why would we object when software performs it more safely, more consistently, and at scale?

We can’t industrialize ourselves for two decades and then suddenly claim artisanal status.

If we want to resist replacement, we have to decide what’s not for sale. What’s not interchangeable. What requires a human being in the room.

And we have to build the case for it.

In case you missed it, I wrote a piece on tech interfaces and their power to shape how we see the world.

Subscribe now

Epic releases its ambient scribe

Epic dropped its long-telegraphed AI Charting product ($) and it’s aimed at the ambient scribe market. Epic’s CMO Jackie Gerhart bristles at the word scribe because they’re framing this as something more active: not just notes, but orders and downstream workflow. STAT has the clean reporting. has the deep dive. He’s bearish on the future of free roaming scribes. Me, less so.

🔺 Scribing is a feature, not a product. Once Epic can deliver “documentation + orders + coding + whatever’s next” inside the EHR, standalone scribe companies will fold faster than beach chairs at high tide. Yes, big systems have contracts. But I suspect it’s just a matter of time.

Millions ‘mourn’ the retirement of GPT-4o

OpenAI is retiring GPT-4o, the ChatGPT model infamous for its sycophantic, affirming tone. The backlash has been revealing. Thousands of users describe it as a friend, a romantic partner, even a spiritual presence. I can’t decide if this is weird or sad

Okay, it’s sad.

GPT-4o was optimized to make us all feel amazing. It reflected feelings, validated pain, and reduced friction. For the lonely or isolated, you can see the draw.

🔺 This dependency may be the most concerning part of the AI revolution. The most dangerous failure mode isn’t bad answers. It’s synthetic intimacy without any kind of accountability. These things deliver a relationship that feels like care, but can’t intervene or hand off when things go sideways.

If you’re interested, I had a conversation not long ago with Daniel Oberhaus about his wonderful book The Silicon Shrink. It’s a sharp look at the digital mental health boom and a nice companion to this moment. Note: my podcast is now The Liminal MD — the logo apparently didn’t populate with the link.

Super Bowl Special — The loose evidence for helmet extenders

As a football fan I thought this was interesting…If you watch the Super Bowl, you may notice players wearing those cartoonish helmet caps. The NFL promotes them as a safety advance. The manufacturer’s claim is less rosy: they blunt routine hits but do less to prevent the more serious concussions.

🔺 This is endorsement filling an evidence gap. The Guardian Cap may offer marginal protection, but the league’s narrative does the heavier lifting. That confidence flows downstream, where parents and coaches assume that “what the pros wear” must be proven.

Plastic surgeons suggest delaying gender surgery

The American Society of Plastic Surgeons now recommends deferring gender transition surgeries until after age 19, citing limited evidence on outcomes and concern about irreversible harm. This lands amid political pressure, legal scrutiny, and growing institutional caution. The AMA also recontoured its position this week.

🔺 We’re watching a shift toward delay as a default. When evidence is contested and risk expands, specialty societies and systems will gravitate toward restraint. Less because of a resolution of the science, but because uncertainty now carries consequences.

I Love this quote

The market for something to believe in is infinite. — Hugh McLeod quoted in this Seth Godin post on Feb 5th.

Who said there’s no future for textbooks?

I shared this on Substack notes earlier in the week while on inpatient service. I weigh in there on smaller issues so be sure to follow me there for compelling images like this.

—————

Technology requires that we connect with it in order to use it. An interface is that spot where you work, like a touchscreen or keyboard. We can even connect with our voices. Every technology has an interface. It’s how we connect to our tools.

The problem is that our interfaces have become our core lens on the world. They no longer mediate reality for us; they replace it.

If you swipe through Instagram you’re processing someone else’s reality. You’re forced to see things their way. Through the screen of your phone you hear someone else’s narrative. It’s not something you’ve created or designed for your needs. And it turns out that a relatively small number of people quietly steer culture, attention, and even morality.

James Bridle wrote in The New Dark Age:

Google set out to index all human knowledge and became the source and arbiter of that knowledge; it became what people actually think …. software is unable to distinguish between its model of the world and reality — and once conditioned, neither are we.

Jim Morrison was ahead of his time when he suggested, whoever controls the media controls the mind.

Subscribe now

Now ChatGPT is evolving as our new frame of reference to the world. I’m amazed at what it can do for me. But at the same time it has a frightening grip on how I understand things.

In medicine we run the risk of mdGPT (aka, OpenEvidence and DoxGPT) defining what we see and understand about the clinical world. Will we scrutinize literature or will thinking machines deliver a kind of forgone conclusion? For patients this represents a new way of understanding themselves. But again it runs the risk of being their true North about their health.

Of course, we’ll never be able to abandon our connection to information, entertainment, and art. But when you turn away from your screen you earn a lens on your own world. When you become comfortable, even for a short time, with the sound of your own thoughts or the ideas that emerge on the blank pages of a notebook, you create your own narrative.

We just need to pay more attention to what we pay attention to and understand how it shapes or distorts our view. And when we’re untainted by what Jaron Lanier calls the persuasion labyrinth, we can begin to find our way back to reality.

Reality can be scary, of course. Which is why our phones can be a kind of opium — an escape from the monotony and pain of our surroundings and space. And be it alcohol or an app, occasional escapes are okay. But when it becomes your default means of seeing and moving through the world, you’ve got a problem.

Here’s the hardest pill to swallow: Show me your granddaughter’s phone and I’ll show you how she sees the world. And perhaps who she becomes. She has no pre-interface world to compare it against.

What’s between you and the world? What are your interfaces?

As always, thanks for making it this far. I’ll catch y’all this weekend….