An Outside Observer's Prediction on When AI Will Perform Surgery

My prediction: Within the next 40 years, AI-powered robotic systems will perform surgical procedures autonomously — starting with anatomically bounded procedures and advancing toward complex interventions. It won't happen because surgeons are inadequate. It will happen because the technology is compounding, the surgical workforce is shrinking, and the generation being born today doesn't share the cultural resistance of the one raising them. The fight over when will be less about capability than about liability, lobbying, and institutional inertia. The question worth asking now is what the transition looks like, not whether it arrives.

---

I Should Say Upfront What I Am and What I'm Not

I barely made it through high school. I never went to university. My only medical experience is roughly two years as an EMT at a North Florida hospital in the early 2000s — which gave me proximity to medicine, not expertise in it. Over that time I watched cardiologists perform cardiac catheterizations more than once — threading a guidewire through the coronary vasculature under fluoroscopic guidance, injecting contrast to visualize the anatomy and find the occlusion. I was a civilian with a front-row seat.

What I do have is a career spent moving across industries — technology, consumer goods, real estate, emergency medicine on the margins — and a habit of noticing patterns. How industries resist change. How new capabilities arrive looking crude before they arrive looking inevitable. How the people most invested in the current system are usually the last to see what's coming.

This piece grew out of a conversation with a surgeon whose skepticism started the whole discussion. We disagreed about whether autonomous AI surgery would happen in our lifetime. I said yes. The answer was no. At some point I made the argument that humans make mistakes across every industry, regardless of skill or training — because we're human. We have bad days. We have emotions. We come in tired, distracted, or carrying things we can't set down at the door. A forklift operator does. An air traffic controller does. A surgeon does too. Not because of incompetence, but because of biology.

That isn't a comparison of surgical skill to operating heavy equipment. It's an observation about the universal condition of being human.

The best surgeons in the world are still human.

---

Google Cardboard, and the Shape of a Transition

In 2014, Google shipped a cardboard box with two plastic lenses. You folded it, shoved your Android phone into the front, and held it up to your face. The resolution was terrible. The frame rate was nauseating. It looked like someone had taped a video game from 1998 over your field of vision.

I put it on anyway. And even through the pixelation and the motion sickness, I understood immediately what I was holding. Not the product — the idea. The sense that something that would matter enormously was just beginning to exist in recognizable form.

A decade later, Apple shipped the Vision Pro.

From a cardboard box and a phone screen to a $3,500 spatial computing device with micro-OLED displays, eye tracking, and a processing architecture designed from scratch for three-dimensional interaction. The distance traveled isn't incremental — it's categorical. If you had looked at Google Cardboard in 2014 and concluded that VR was a gimmick, you would have been right about the product and catastrophically wrong about the idea.

That's the frame I keep returning to when I think about autonomous surgical systems. We're somewhere in the Google Cardboard era. The technology exists, it works in controlled conditions, and in some ways it's still crude. But the idea is there. If you've spent enough time watching technology compound, you don't need to see the finished product to recognize what's coming.

The pattern repeats across industries. When the iPhone launched in 2007, Steve Ballmer — then CEO of Microsoft — told USA Today: "There's no chance that the iPhone is going to get any significant market share."⁷ A senior Microsoft marketing director publicly predicted Apple wouldn't come close to the 10 million units Steve Jobs had projected for 2008. Apple sold over 11 million.⁸ By 2015, the iPhone held 42% of the U.S. smartphone market. The point isn't that Ballmer was stupid — he wasn't. The point is that even people with deep expertise and every incentive to see clearly can miss a trajectory that's unfolding in front of them. That's not a story about phones. It's a story about how transitions actually look from the inside.

My own career has mostly been in tech, and I've watched what AI tools have done to the output capacity of people in design, development, research, and writing — my own included. Fields I spent years in are being reshaped, not because the people in them are being replaced, but because the nature of the work is changing faster than most people anticipated. Surgery isn't the same as design. But the underlying dynamic — new capability arriving before the culture is ready for it — is the same. And it isn't just operating rooms. It's most jobs that involve pattern recognition, precision, and high-volume repetition.

---

What's Already Happening in the Lab

The public conversation around AI surgery still treats it largely as a future scenario. The research disagrees.

In January 2022, a team at Johns Hopkins University built a robot called STAR — the Smart Tissue Autonomous Robot — and used it to perform the first autonomous laparoscopic surgery on a live pig.¹ Not surgeon-assisted surgery. Not teleoperated surgery. Autonomous surgery, with the robot planning, adapting, and executing in real time on soft tissue without a human hand guiding it. The STAR outperformed expert human surgeons on intestinal anastomosis — the reconnection of two severed ends of bowel — a procedure requiring sustained fine motor precision.

Three years later, the same lab went further. In July 2025, they published results from a successor system called SRT-H that performed a complete laparoscopic cholecystectomy without human intervention.² The procedure required 17 discrete tasks: identifying anatomical structures including the cystic duct and artery, placing surgical clips with precision, dividing tissue with scissors, and adapting in real time to variability in the surgical field. The system learned by watching videos of Johns Hopkins surgeons performing the procedure on porcine cadavers. After training, it achieved 100% task accuracy — and held that performance when researchers altered its starting position mid-procedure and added contrast dye that changed the visual appearance of the tissue.

The principal investigator, Axel Krieger, summarized what's new: "This advancement moves us from robots that can execute specific surgical tasks to robots that truly understand surgical procedures. This is a critical distinction that brings us significantly closer to clinically viable autonomous surgical systems that can work in the messy, unpredictable reality of actual patient care."³

Diagnostics is further along the curve. In a multicenter study of over 1,500 patients, an autonomous AI system achieved a 99.1% sensitivity rate for abnormal findings on chest radiographs, compared to 72.3% for radiologist reports.⁴ As of late 2023, the FDA had cleared over 1,000 AI-powered medical devices, with diagnostic imaging as the dominant category.⁵ A Harvard Medical School study examining 140 radiologists found that AI assistance changed diagnostic performance in ways that varied substantially between physicians — a signal that human-AI integration is still being worked out.⁶ But directionally, in pattern recognition tasks on medical imaging, AI is already operating at or near specialist level in defined domains.

---

Pattern Recognition Cuts Both Ways

In the conversation that started all of this, a point came up directly from personal experience in surgical training and at conferences: the specific danger of pattern recognition.

Surgeons encounter a clinical presentation — a set of images, a tissue configuration, a constellation of findings — that matches something they've seen many times before. The experienced mind, almost automatically, concludes it's seeing the same thing again.

Except sometimes it isn't.

That cognitive failure mode has a name. It's a form of anchoring bias layered on heuristic pattern matching, and research on overconfidence and working memory — explored in depth by Veritasium in their study of why people are most confident precisely when their judgment is most compromised — finds that as cognitive load increases, confidence estimates become less accurate, not more.⁹ The more a mind has to track simultaneously, the more it falls back on shortcuts. For a surgeon managing a complex intraoperative picture, that's not a character flaw. It's a documented feature of how human cognition works under pressure.

Robotic systems don't carry that accumulated weight. A well-designed AI evaluates each case against its training data without the psychological residue of the previous ten thousand cases. Each encounter is, in a meaningful sense, new.

Of course, AI has its own failure modes — different ones, which is the point. Tesla's Autopilot has been involved in fatal accidents and remains under regulatory scrutiny. It is not a perfect system. But Tesla's Q3 2025 safety data showed one crash per 6.36 million miles driven with Autopilot engaged, compared to roughly one per 702,000 miles for conventional U.S. driving.¹⁰ Whatever the methodological caveats, the directional trend is real: over the decade since autonomous driving features were introduced, the failure rate has dropped substantially. Failure modes get identified, quantified, and corrected through iterative retraining. The trajectory is downward error, not flat.

That's the meaningful comparison for surgical AI. The question isn't whether early autonomous systems will be perfect. They won't be. The question is whether their failure modes are identifiable, correctable, and trending in the right direction — while human failure modes, documented for decades, have not meaningfully changed. Based on everything we know about how machine learning systems mature, the trajectory favors one over the other.

---

30 Cases vs. 3,000

That same conversation produced a direct challenge: why would anyone choose to trust a robot over a human surgeon?

Consider a skilled surgeon early in their attending career, performing laparoscopic cholecystectomies. Residency complete. Fellowship done. Credentialing in order. They've performed 30 gallbladder removals. Every one successful. A 100% success rate.

Now consider an AI surgical system that has performed 3,000 gallbladder removals. Also successful. Also 100%.

At 30 cases, the human surgeon is still on the steep part of the learning curve — building the intuition that only repetition creates, still encountering anatomical variations they haven't seen before. At 3,000, the AI has seen a far broader range of presentations, unexpected findings, and intraoperative complications — and every one of those cases has been incorporated back into the model.

Now the system does case 3,001. Then 3,002. Then 30,000. The human surgeon, across a full career, might perform 1,500 to 2,000 cholecystectomies. The machine's case count doesn't plateau. It doesn't stop accumulating at retirement. And what it learns isn't siloed to one operator in one hospital. When a system encounters an unusual anatomical variant in Pittsburgh, that information can strengthen the model operating in Phoenix, São Paulo, and Singapore simultaneously.

This is the compounding argument. Not that human surgeons are inadequate — but that machines accumulate procedural experience in a fundamentally different way, one that doesn't respect the biological limits of individual human careers. The early adopter who chooses that system at case 3,000 is getting the benefit of more aggregate experience than any individual surgeon could accumulate in a lifetime.

The same dynamic is already visible in creative fields. A musician creating a full studio album typically spends six to twelve months from pre-production through mastering.¹¹ An AI system can generate thousands of songs in the time it takes to read this sentence. That's not a marginal improvement in speed — it's a categorical change in what's possible. What survives that disruption isn't the output; it's the provenance. I think we'll see "Made by Human" labels the way we have organic food certifications. The consumer who cares about human origin will pay a premium. The consumer who just wants a bowl will pay less for the stamped version. Both markets coexist. But the economics are never the same again.

That logic reaches medicine too. Not because the human element stops mattering — but because the allocation of human attention shifts toward what machines genuinely cannot do.

---

The Shortage That Changes the Argument

I should say, before going further: I'm aware I'm an outsider making an argument about a field I don't practice in. Everything I've said so far and everything I'm about to say is filtered through that. What follows is data I can read, not clinical experience I have. I'm confident in the reading. I'm also aware of the distance between reading and doing.

There's another argument for autonomous surgery that has nothing to do with robots being better than surgeons. It's about the gap already opening between surgical demand and surgical supply.

The Association of American Medical Colleges projects a shortage of between 19,800 and 29,000 surgeons by 2030.¹² A peer-reviewed study in The American Journal of Surgery projects that general surgery alone could face a deficit of over 25,000 surgeons by 2050 if current trends hold, with nine surgical specialties facing workforce shortfalls by 2030.¹³ The Health Resources and Services Administration projects cardiothoracic surgery will face a 31% shortfall by 2035 — the largest projected shortfall of any physician specialty evaluated.¹⁴

The driver isn't just retirement. It's demographics. The U.S. population is aging, surgical demand is rising, and the training pipeline — constrained by residency caps and a decade-long medical school moratorium that artificially suppressed physician supply — isn't keeping pace.

That reframes the entire debate. The question stops being "why would we let a robot operate on someone?" and becomes "what happens to patients when there aren't enough surgeons to go around?" In rural communities, the shortage is already real. The South faces a projected deficit of over 10,000 surgical FTEs, with rural areas dramatically more exposed than urban ones.¹⁵

Autonomous systems don't just compete with human surgeons for patients who already have access to excellent care. They extend surgical capacity to the patients who currently don't.

---

What's Actually Holding This Back

If the technology is progressing this fast and the shortage is already measurable, why aren't hospitals piloting autonomous systems today?

The most obvious reason is the one worth stating first: these are human lives. You can iterate on an autonomous driving model with simulated miles. You can train a surgical robot on porcine cadavers. But at some point, a human being has to be the first patient, and the gap between "it worked on a pig" and "it's ready to operate on a person" is not a gap that responsible medicine closes quickly. The technology isn't fully there yet. I'd argue it's closer than most people realize, but I'm not claiming it's ready for a living patient tomorrow. That caution is appropriate, and it's the right reason for the pace to be deliberate.

But past that — past the real and legitimate caution about patient safety — there are three structural obstacles that have nothing to do with whether the robots can do the work.

Liability has no legal architecture yet. When an autonomous system is involved in a surgical complication, who bears responsibility? The manufacturer? The hospital? The surgeon who reviewed the pre-operative plan? The AMA's president has publicly stated that the organization is already "seeing lawsuits" around AI-assisted surgical devices, and physicians are actively lobbying to shift liability toward manufacturers.¹⁶ Surgical AI will go through its own version of the reckoning autonomous vehicles are still navigating, and until that legal framework settles, hospitals and insurers won't know what they're agreeing to.

The lobbying infrastructure is enormous and historically effective. The AMA spent approximately $21 million on lobbying in 2023 alone¹⁷ and increased that figure by 21% in Q1 2025.¹⁸ Their positioning is carefully chosen: they insist on the term "augmented intelligence" rather than artificial intelligence, and have built their regulatory stance around the principle that AI should "support the work of physicians, rather than replacing them."¹⁹ That's a reasonable position on the surface. It also, conveniently, preserves a physician billing code in every procedure. When the AMA successfully blocked universal healthcare in the 1930s, they shaped the architecture of American medicine for a century.²⁰ They are not an organization that loses slowly.

Medicine has always been among the last sectors to adopt new technology. In 2008, only 7.6% of U.S. hospitals had implemented even a basic electronic health record system.²¹ Facebook had already launched. The iPhone was already in pockets. A peer-reviewed study described healthcare as "decades behind other industries" on information technology adoption.²² EHR penetration only crossed 80% of hospitals in 2019 — more than a decade after the iPhone shipped. That pattern isn't an accident. It reflects a field that moves deliberately, for understandable reasons. But deliberately is not the same as never.

---

The Generation That Will Cross the Chasm

Everett Rogers mapped technology adoption in 1962: from the 2.5% of early innovators to the 13.5% of early adopters, and then the chasm — the gap where most promising technologies stall before reaching the mainstream.²³

Healthcare AI in diagnostics sits somewhere in the early adopter phase now. Autonomous surgery hasn't reached commercial deployment — it's still in the lab, which precedes even the innovator stage. But the chasm isn't permanent, and it tends to close faster than incumbents expect.

Max Planck understood the mechanism:

"A new scientific truth does not triumph by convincing its opponents and making them see the light, but rather because its opponents eventually die, and a new generation grows up that is familiar with it."

The medical students entering training today grew up using AI tools as a matter of course. The patients who will undergo surgery in the 2050s and 2060s won't carry the cultural resistance older patients do. They've already handed consequential decisions to algorithms — what to watch, who to meet, how to navigate. Ceding a bounded, validated surgical procedure to a well-trained robotic system won't feel like surrendering to a machine. It will feel like using the right tool.

---

What the Transition Actually Looks Like

The first commercially deployed autonomous systems will be narrow, supervised, and anatomically defined. Dermatological procedures are, in my opinion, the natural starting point — not because I have any real depth of knowledge about the relative complexity of surgical procedures, but because it's something I've had done to myself repeatedly, and I understand it from the patient side. I've had more than twenty moles removed from my body over the years — nearly all of them biopsied for potential malignancy, all of them benign. On more than one occasion, I let residents perform those procedures. If there was any procedure where a less experienced hand should practice, a surface excision on a healthy body seemed like a reasonable place. If the technology existed today, I'd be happy to be one of the first people to let a robot remove one.

Before each procedure, the system would present a pre-operative plan to the supervising physician: here is my assessment of the lesion, here is my proposed excision depth and margin, here is my instrument selection. The physician reviews, approves, and the system executes.

Think about how a CNC machine calibrates before cutting. It runs a probe across the material surface to map where the Z-axis begins — to know precisely where the material is before it commits a cut. An autonomous surgical system will do something similar: scan the target tissue, build a real-time spatial model, verify the planned path before it moves. Human anatomy isn't uniform. But the sensing and spatial mapping technology to accommodate that variability already exists in prototype form.

As outcomes data accumulates — across thousands of cases, then tens of thousands — the scope expands. Autonomous suturing. Tissue manipulation in laparoscopic settings. Eventually, procedures that look like what I once watched in that cardiology suite: real-time vascular navigation, catheter advancement through complex coronary anatomy, stent deployment guided by live fluoroscopy and AI-interpreted contrast imaging — without a human hand on the wire.

The business model will probably follow a per-procedure structure layered into insurance billing — a usage fee for surgical capability, where the robotics company earns a credit each time the system operates. Liability shifts toward manufacturers as the legal framework develops. Prices fall as volumes increase.

---

What Surgery Actually Is — And What Survives

The surgeon whose skepticism started this conversation is exceptional at the work. And surgery, viewed from the outside, involves far more than the procedure itself.

Before a patient ever reaches the operating room, the surgeon has reviewed prior imaging, laboratory results, and the patient's full medical history. They've consulted with anesthesiology and, depending on the case, with other specialists. They've weighed whether to operate at all — a judgment call that blends clinical data with experience and is rarely as clean as it looks from the outside. Intraoperatively, they're reading anatomy that doesn't always match what the imaging showed, making real-time decisions when something unexpected presents itself, managing the room, and communicating with a team.

And then there's the part that gets underestimated: the conversation before the cutting. The ability to sit with a patient who is frightened, who doesn't understand what is about to happen to their body, who needs someone to translate the clinical reality into something they can hold. Surgery is a physical act that begins as an emotional one. The informed consent. The explanation of risk. The accountability a patient needs before they let someone open their body. That doesn't go away. If anything, in a world where procedural execution becomes increasingly mechanical, it becomes more central.

So the surgeon's role doesn't disappear. It changes. A growing category of technically bounded, procedurally defined surgical tasks gets handed to systems that don't get tired and don't carry the weight of the previous case into the next one. The surgeon's attention shifts toward what machines genuinely cannot provide: judgment in novel situations, presence in difficult conversations, accountability to individual patients.

That's not a lesser role. In the long run, it may be a more purely human one.

---

The Argument I'm Actually Making

I'm not a doctor. I barely made it through high school and never went to university. I learned most of what I know by moving across industries and paying attention to what they have in common.

And what they have in common, in my observation, is this: every major technological transition looks impossible until it looks inevitable. The people who insist it can't happen are usually right about the current version of the technology. They're almost never right about the trajectory.

I put on Google Cardboard in 2014. I didn't need the product to be good to understand what the idea was worth. That's the only kind of forecasting I know how to do — pattern recognition across industries, from an outsider who's had a front-row seat more than once.

This is one of those moments. I'm confident about that. And having thought about it, I'm also aware that confidence isn't the same as certainty.

← All writing