2.6 million robotic surgeries, none of them autonomous

TL;DR. More than 2.6 million procedures were performed with the da Vinci system in 2024, across thousands of installed units worldwide. It is by a wide margin the most-deployed robot in medicine and the public's mental image of AI in surgery. It is a teleoperator. The surgeon controls every movement in real time from a console; the system decides nothing. And the largest evidence synthesis, the COMPARE study in Annals of Surgery, pooled 230 studies covering over a million robotic procedures and reported real benefits, and it was co-authored by scientists from the manufacturer. The study is registered, peer-reviewed and follows reporting standards. It is also the case that the most comprehensive evaluation of a device was produced with the participation of the company that sells it, which is the external validation question in a domain where the evidence is otherwise strong.

---

Status: established. Primary sources: the COMPARE study, Ricciardi and colleagues, Annals of Surgery 281(5):748-763, May 2025, PROSPERO-registered and PRISMA-following, and the manufacturer's own announcement of it. Both are cited below.

---

The robot does not decide anything

Worth stating plainly because the language obscures it.

In robot-assisted surgery the surgeon sits at a console and moves controls. The instruments inside the patient follow those movements. The system scales motion, filters tremor, and provides stereoscopic vision and instrument articulation beyond what a human wrist manages through a laparoscopic port.

It does not plan the operation, choose where to cut, recognise anatomy, or act on its own. Remove the surgeon and nothing happens.

This is teleoperation, and a very good implementation of it. The value is real: motion scaling and tremor filtering genuinely improve fine manipulation in a confined space, and the seated console position reduces surgeon fatigue over long procedures.

But it is not autonomy, and it is not artificial intelligence in any sense that survives contact with the definition. The most-deployed robot in medicine, the one most people picture when they think of AI in surgery, contains no decision-making at all.

That matters for the same reason the industrial robot figures matter. The successful physical systems are not doing the thing the discussion is about.

The evidence is substantial

Unlike the sepsis model, this technology has been studied extensively.

The COMPARE study, published in Annals of Surgery in May 2025, systematically searched three databases across twelve years and pooled 230 studies from 22 countries: 34 randomised controlled trials, 74 prospective studies and 122 database studies.

The comparison covered seven oncologic procedures and, remarkably, more than a million procedures in each arm: 1,194,559 robotic, 1,095,936 laparoscopic or thoracoscopic, and 1,625,320 open.

It reported advantages for the robotic approach on conversion to open surgery, blood loss, transfusion rate, length of stay, readmissions and reoperations.

That is a serious piece of work, PROSPERO-registered in advance, following PRISMA reporting standards, with formal risk-of-bias assessment using ROBINS-I and RoB 2.

And the evaluator

The meta-analysis was conducted by scientists from Intuitive and Massachusetts General Hospital. Intuitive manufactures and sells the da Vinci system.

This is disclosed, not concealed, and the manufacturer's own announcement states it. Industry participation in clinical research is normal, frequently necessary, and the alternative, evidence produced only by parties with no access to the technology, is worse.

It is also precisely the structure the sepsis case turned on. That article argued a validation is useful in proportion to two things: the evaluators are not the developers, and the population was not selected by the developer.

Here the first condition is partly unmet. The second is largely satisfied, since the pooled studies were conducted by many independent groups over twelve years, and the 34 randomised trials in particular were not the manufacturer's to design.

So the honest reading is not that the finding is wrong. It is that the most comprehensive synthesis available was co-produced by an interested party, and that an equally comprehensive independent synthesis does not exist. Those are different claims and only the first is usually reported.

What the underlying evidence looks like

Beneath the pooled figures there is a pattern worth naming, and it is not unique to this device.

The strongest advantages are perioperative: less blood loss, shorter stay, fewer conversions to open surgery, faster recovery. These are real, measurable, and matter to patients.

The advantages on longer-term outcomes are harder to establish. Cancer recurrence, survival and functional recovery require years of follow-up and much larger trials, and the evidence base thins considerably once the horizon extends past thirty days. The COMPARE study is explicitly a thirty-day outcomes analysis.

And cost runs the other way. Capital cost is substantial, consumable instruments account for a large share of per-procedure expenditure, and setup adds roughly half an hour against laparoscopy in early cases. Break-even depends heavily on operating room utilisation and surgeon volume, which means the economics are institution-specific in a way the clinical figures are not.

None of that contradicts the clinical findings. It is what the clinical findings do not cover.

Three things this establishes

Deployment scale is not a claim about autonomy. Millions of procedures with a teleoperator says nothing about what an autonomous system could do, and the shared vocabulary invites exactly that inference.

Extensive evidence and independent evidence are different properties. This field has far more of the first than most, and less of the second than the volume suggests. A reader counting studies will conclude the question is settled; a reader checking authorship will find the largest synthesis has an interested co-author.

And the measurement horizon shapes the conclusion. Thirty-day perioperative outcomes are where robotic assistance looks strongest and where evidence is most abundant, because they are cheap to measure. The outcomes patients care about most are the ones the evidence base covers least, which is a general property of surgical research rather than a criticism of this device.

What it does not establish

That robot-assisted surgery does not work. The perioperative advantages are consistently reported across independent groups, including in randomised trials the manufacturer did not run.

That the COMPARE study is biased. Manufacturer involvement is a reason to want independent replication, not a finding of error. Pre-registration, PRISMA adherence and formal bias assessment are exactly the safeguards that make an interested party's work assessable.

That the technology could not become autonomous. Autonomous suturing and tissue manipulation are active research areas. The claim here is about what is deployed, which is teleoperation.

And nothing about comparative harm. This article makes no claim that robotic surgery causes harm relative to alternatives, and the evidence reviewed points the other way on the outcomes it covers.

What is unresolved

Whether an independent synthesis of equal scope would reach the same conclusions. Nobody has produced one, and the question is answerable only by someone doing it.

Long-term oncologic outcomes. Survival and recurrence at five and ten years across these procedures are not what the thirty-day literature measures.

Whether the perioperative advantages justify the cost at typical volumes. Break-even analysis depends on utilisation, and hospitals with low volumes may not reach it.

And whether newer platforms perform equivalently. Comparative studies of emerging systems have largely benchmarked against an older generation of the incumbent, which is a moving comparison rather than a fixed one.

The counter-argument

Calling teleoperation a finding is pedantic. Nobody working in surgery believes the robot operates itself, the term "robot-assisted" is standard, and objecting to the public's misunderstanding of a technical term is not a criticism of the technology.

Manufacturer involvement is the price of comprehensive evidence. Assembling 230 studies across 22 countries requires resources academic groups rarely have, and the alternative to industry-supported synthesis is usually no synthesis at all. Treating disclosed participation as a flaw penalises the parties who fund the work.

The randomised trials are the answer. Thirty-four RCTs conducted by independent groups sit inside the pooled analysis, and their results do not depend on who assembled them. Pooling is a methodological act, not an evidentiary one.

And the cost objection is about health systems, not devices. Whether a hospital can justify the capital outlay is a procurement question. It says nothing about whether the technology helps the patient on the table, which is what the clinical evidence addresses.

The short version

More than 2.6 million procedures were performed with the da Vinci system in 2024, making it the most-deployed robot in medicine and the public's image of AI in surgery.

It is a teleoperator. The surgeon moves controls at a console and the instruments follow. It scales motion, filters tremor and articulates beyond a human wrist through a port, all of which is genuinely valuable. It decides nothing. Remove the surgeon and nothing happens.

The evidence is substantial. The COMPARE study pooled 230 studies from 22 countries, including 34 randomised trials, covering more than a million procedures in each arm, and reported advantages on conversions, blood loss, transfusions, length of stay, readmissions and reoperations. It was pre-registered, follows PRISMA, and includes formal bias assessment.

And it was co-authored by scientists from the manufacturer, which is disclosed and is also the structure the sepsis article identified: a validation is useful in proportion to the evaluators not being the developers. The honest reading is not that the finding is wrong, but that the most comprehensive synthesis available was co-produced by an interested party and no equally comprehensive independent one exists.

Beneath it, the pattern is ordinary. Perioperative outcomes are where the advantages are clearest and where evidence is most abundant, because thirty days is cheap to measure. Survival and recurrence need years and much larger trials. The outcomes patients care about most are the ones covered least, which is a property of surgical research rather than of this device.

Common questions

Is robotic surgery actually autonomous? No. In robot-assisted surgery the surgeon sits at a console and moves controls in real time, and the instruments inside the patient follow those movements. The system scales motion, filters hand tremor and provides stereoscopic vision and articulation beyond what a human wrist achieves through a laparoscopic port. It does not plan the operation, identify anatomy or act on its own. Remove the surgeon and nothing happens. It is teleoperation, and a very good implementation of it.

How widely is it used? More than 2.6 million procedures were performed with the da Vinci system in 2024, across thousands of installed units worldwide, with over 200 hospitals in Italy alone and 25 years of use there. It is by a wide margin the most-deployed robot in medicine.

What does the evidence say? The COMPARE study, published in Annals of Surgery in May 2025, pooled 230 studies from 22 countries over 12 years, including 34 randomised controlled trials, 74 prospective studies and 122 database studies. It covered seven oncologic procedures with more than a million procedures in each comparison arm and reported advantages for the robotic approach on conversion to open surgery, blood loss, transfusion rate, length of stay, readmissions and reoperations at 30 days.

Who conducted that study? Scientists from Intuitive, which manufactures the da Vinci system, together with Massachusetts General Hospital. This is disclosed in the manufacturer's own announcement. The study was registered with PROSPERO in advance, follows PRISMA reporting standards and includes formal risk-of-bias assessment, which are the safeguards that make interested-party research assessable rather than dismissible.

Does manufacturer involvement mean the results are wrong? No, and this article does not claim that. Industry participation in clinical research is normal and frequently necessary, since assembling 230 studies across 22 countries requires resources academic groups rarely have. The point is narrower: the most comprehensive synthesis available was co-produced by an interested party, no equally comprehensive independent synthesis exists, and only the first half of that is usually reported. The 34 randomised trials inside the pooling were conducted by independent groups and their results do not depend on who assembled them.

What does the evidence not cover? Long-term outcomes. The COMPARE study is explicitly a 30-day analysis, and the evidence base thins considerably beyond that horizon. Cancer recurrence, survival and functional recovery require years of follow-up and much larger trials. Perioperative outcomes are where robotic assistance looks strongest and where evidence is most abundant, largely because 30 days is cheap to measure, which means the outcomes patients care about most are the ones covered least.

What about cost? It runs the other way. Capital cost is substantial, consumable instruments account for a large share of per-procedure expenditure, and setup adds roughly half an hour against laparoscopy in early cases. Break-even depends heavily on operating room utilisation and surgeon volume, so the economics are institution-specific in a way the clinical figures are not. That is a procurement question rather than a clinical one, and it does not bear on whether the technology helps the patient.

Why does this belong in a series about robotics and AI? Because it is the clearest case of the vocabulary doing work the technology does not. Millions of procedures performed by the most-deployed medical robot, and the robot decides nothing. Deployment scale gets read as evidence about autonomy, when the system demonstrates something different and genuinely valuable: that a human operator with better instruments outperforms the same operator with worse ones.