In the movie Elysium, set in the year 2154, Matt Damon plays Max Da Costa, a criminal lowlife trying his damndest to leave Earth. Up above, the wealthy elite live on a ring-shaped space station, where disease and poverty have been eradicated. Most of the population, however, still lives on Earth, and they are struggling to survive each day. They can only dream of ever making it to the space station in the sky. But for Max, getting to Elysium is a matter of survival: He needs access to healthcare that can save his own life.
Even in the imaginary world of Elysium, 125 years into the future, high-quality, AI-driven healthcare is only for the privileged few. The ultrawealthy, who were able to escape an impoverished and overpopulated planet, live in tranquil luxury in space. This dystopian portrayal of healthcare’s vast inequities is unfortunately more realistic than we would like to believe. But our future doesn’t have to be this stark.
Imagine Sam, a man in his 50s who walks into his doctor’s office complaining of chest pain. First, the doctor asks Sam about his symptom history and risk factors; then doctor performs a physical assessment, taking vitals like heart rate and blood pressure. The doctor might request additional tests, such as an ECG, a blood test, or chest X-ray. All these data help the physician make a list of differential diagnoses, such as angina, heart attack, or just acid reflux. Based on training and experience, the doctor makes a diagnosis and proposes additional tests and treatment options.
Today, these functions are the domain of human expertise—doctors, nurses, and providers who undergo highly specialized education and on-the-job training. But that may not be the case for long. Within the next decade, Sam may be seeing an artificial intelligence doctor for his common healthcare needs and day-to-day management.
The transition to medicine supported and eventually delivered by algorithms —what I call autonomous health—promises to make healthcare more convenient, accessible, and affordable. Imagine no more waiting to see a doctor. In 2025, there was an estimated shortfall of 35,600 physicians and 25,200 surgeons. We have the highest per-capita spend on healthcare in the developed world, yet annual healthcare premiums for employers are going up 6.5-7% this year.
Increasing automation is the only path to mitigate the access shortages and increasing cost challenges in healthcare. LLM researchers have demonstrated incredible ability to distill medical knowledge. Many already visit ChatGPT and Claude for healthcare answers before they see their doctor.
But this also raises a flurry of questions. Foremost, can we ensure these technologies are safe and clinically validated? How do we develop a hybrid model of care, where humans and AI systems work together as a care team? What will the experience of the healthcare consumer be in this new world?
And then there are the bigger structural questions. Does the financial model of healthcare need to support a path towards increasing autonomy? What should happen to reimbursement rates for clinical visits? How will this impact today’s professional workforce? Will we need even more healthcare professionals, but perhaps with different skills than today’s doctors? Would AI-enabled care be reserved, Elysium-like, for the privileged few? Or might we see the inverse, where automation is common among the masses at lower costs, while healthcare with human expertise comes at a premium?
We can’t even begin to answer these questions until we have a frame of reference, some sense of where we’re going and where we want to be.
Consider the self-driving car. For years, manufacturers used terms like “autonomous,” “automated,” and “self-driving” in their press and marketing, interchangeably and without clear distinctions. This led to widespread confusion about the capabilities and limitations of competing automated driving systems among consumers and regulators.
The Society of Automotive Engineers proposed a solution. They defined a standard for driving automation systems for motor vehicles, which established different levels of automation, from driver assistance aids to fully automated driving. SAE J3016, adopted in 2016, became a cornerstone in the development of autonomous driving technology. Every automaker used it to benchmark their systems, and it gave governments and regulatory bodies a clear framework to create laws, set safety standards, and ensure the safe deployment of automated vehicles. It eliminated the friction that would have held the industry back.
In many ways, AI in healthcare is where self-driving vehicles were a decade ago. Tech companies, medical start-ups, hospital systems, and even insurance companies are now caught up in their own hype cycle. To provide scientific context amid all the noise, I propose four levels for autonomy in healthcare, as part of a framework that can help regulators, administrators, and all of us evaluate and adopt new technologies: No Automation, Light Automation, High Automation, and—perhaps at some point—Complete Automation.
Level 1: No Automation
Across the planet, the average person today experiences little to no automation in their care. Whether a person walks into a concierge doctor’s office in Fort Lauderdale or a rural clinic in Namibia, the diagnosis and treatment decision, even for a simple common cold, is made by a physician.
This is not to discount the widespread access to blood and lab testing, as well as advances in medical imaging. More than 4 billion imaging tests, such as X-rays, CT scans, and MRIs, are done every year across the world. These data are invaluable, but the interpretation and diagnosis, and the recommended treatment are still heavily dependent on human medical expertise.
Level 2: Light Automation
Light automation, or automated assistance for human doctors and medical providers, is akin to cruise control in your car. Adaptive cruise control maintains a set following distance from the vehicle ahead, and lane-keep assist gently steers the car back into its lane if it starts to drift. These features don’t replace human drivers, but they help prevent common mistakes that occur when a human driver may be distracted or fatigued, often due to the sheer monotony of driving.
Level 2 for healthcare is happening now. But unlike adaptive cruise control, these Level 2 doctor assistance systems are not managing mundane, monotonous, mechanical tasks. Instead, they help human doctors and nurses with diagnoses, usually by summarizing or processing large swathes of additional complex data.
Algorithms now play a role in assisting with diagnosis in imaging, such as radiology and cancer detection, and in the identification of risk factors based on genetic data and markers. Without computational analysis, for instance, it is physically impossible for a human doctor to sift through genomic data and confirm the gene expression of BRCA1 or TP53 [genes that repair DNA and prevent cancer] for a patient during a 15-minute appointment.
As of summer 2026, the U.S. Food and Drug Administration (FDA) had authorized more than 1500 AI or machine learning (ML) algorithms for market use. The vast majority of these, about 76%, are in radiology. Most of these make imaging refinements, such as improving image quality, but a growing number are now helping clinicians make better clinical diagnoses. These ML algorithms and models are assisting radiologists in making a better decision, not replacing them. Even now, all the AI and ML algorithms cleared by the FDA require a human to be in the loop.
If 99% of healthcare diagnostic and treatment decisions made every day on planet Earth are at Level 1: No Automation, the remaining 1% are at Level 2: Light Automation. But the share of care being provided with light automation is growing rapidly, with hundreds of new algorithms being approved by the FDA every year.
Level 3: High Automation
From a purely algorithmic perspective, the rapidly improving ability of AI is leading us toward Level 3: High Automation. With high automation, a meaningful percentage of diagnosis and treatment decisions are entirely automated and made by algorithms.
Today, doctors and healthcare professionals make hundreds of decisions every day. At a high level, there are diagnosis decisions, treatment decisions, and clinical interventions such as surgery. Over a year ago, MedGemini and MedPrompt, from Google and Microsoft respectively, achieved a competitive accuracy of 91.1 and 90.2, exceeding human experts (87.01) on U.S. Medical Licensing Examination (USMLE) benchmarks, a dataset of realistic medical exam questions that requires clinical knowledge and reasoning ability. In the last year, researchers have further shown AI agents, integrated with electronic health record (EHR) data, are exceeding the performance of human diagnostic decisions based on clinical data.
To understand the impact this could have on healthcare, let’s rewind. Over a decade ago, at my previous company, Ginger, a unique collaboration began to take shape. Dr. Sai Moturu, a computer scientist, joined forces with Dr. Mimi Winsburg, a leading psychiatrist, and other expert colleagues, to tackle one of psychiatry’s most complex challenges: choosing the right SSRI for a new patient’s first treatment.
Selective serotonin reuptake inhibitors (SSRIs) such as Prozac and Zoloft are antidepressants typically prescribed for anxiety, depression, and PTSD, often alongside psychotherapy or coaching. The decision to prescribe an SSRI is based on a patient’s clinical state, symptoms, lifestyle factors, other comorbid chronic conditions, related medications, and past response to an SSRI. For example, Prozac may have a higher risk of interactions with other medications the patient is taking, but it also has a longer half-life than Zoloft, making it better for patients who occasionally miss doses. A human psychiatrist evaluates the patient, considers these and other factors, and may suggest an appropriate SSRI drug to treat the patient’s symptoms.
By blending a machine learning approach with extensive clinical experience, this cross-functional team of experts built decision-tree models that incorporated clinical history and lifestyle data to match the judgment of a panel of experienced psychiatrists. The results were striking. In head-to-head evaluations across dozens of real-world cases, they found that their ML decision tree model selected the correct initial SSRI over 90% of the time. As in typical practice, dosages could be changed or medications would be updated based on a patient’s response, creating new branches in the treatment decision tree.
This is a relatively simple example of machines’ decision-making potential for treatments in behavioral health. But the potential impact on access to and costs of healthcare would be tremendous. Today, amid a shortage of providers, an estimated 55% of adults with mental health needs go without treatment. What would the economic benefit be if a quarter of cases of depression that need a psychiatrist but can’t find one could be directed to an algorithm?
Such a scenario could allow human psychiatrists to spend their time on more complex cases that benefit from their expertise. In 2025, the average psychiatry visit costs either the consumer or insurer between $75 and $200, depending on the reimbursement code, state, and available insurance coverage. Using current estimates for compute and energy, an algorithmic visit with a psychiatrist could be competitively priced at one-fifth or one-tenth the cost.
We are used to seeing medical residents on hospital floors, providing direct patient care under the supervision of an attending doctor. Under high automation, could algorithms play a similar role and alleviate the access and workforce shortage in healthcare?
My educated guess is that high automation will first focus on primary care, chronic condition management, or mental health, tackling the simplest of diagnoses. AI doctors will likely emerge first in incorporated telemedicine or virtual care, where the ability of human doctors to read cues, expressions, and touch is already constrained. They will integrate seamlessly with existing human clinical care and may escalate many of the cases presented to them to human doctors.
Sensory Equivalence is a Hard Technical Problem
I intentionally oversimplified the Level 3: High Automation scenarios by assuming sensory equivalence. That’s the concept that a human doctor and an algorithm are dealing with the same input data, from different senses. But that’s actually a big assumption to make.
Let’s return to the analogy of self-driving cars. Lidar [light detection and ranging] was an essential breakthrough that allowed cars to perceive the constantly changing world around them, reading the movement of cars, pedestrians, buses, and bicyclists. It provides precise depth perception, detecting objects’ size, shape, and distance far more accurately than cameras or radar sensors alone. While there are alternative approaches, including Tesla’s camera-based system, Lidar has played a crucial role in the development of autonomous vehicles. Without depth perception, deep learning AI models couldn’t learn to drive.
We are not great at communication when we see a doctor. We often withhold relevant health information because we feel shame or stigma about admitting to our actions or behaviors. We dismiss what we believe are unrelated symptoms. Add to these cultural differences, language barriers, and concerns about insurance coverage, and it is no wonder communication in healthcare settings is so challenging.
And yet, doctors are skilled at building a complete picture with incomplete data to make a diagnosis. From the moment they see the patient, it’s an interactive process. Humans have a remarkable ability to merge raw sensory data from sight, sound, context, and intuition.
Deep learning advances about a decade ago helped computers get remarkably better at face and image recognition, detecting microexpressions, gaze, and subtle body language. Current vision-language models are better at image segmentation and understanding depth perception, and are compatible with video. Similarly, haptic and tactile sensors are helping machines develop a bit more of the human sense of touch. Advancements in flexible electronic skins can detect pressure, temperature, texture, and vibrations. Perhaps at some point in the future medical robots, either at home or the doctor’s office, could achieve parity with human sensing ability. For the moment, though, there is still an inconceivable gap between human observational skills and the most advanced machine vision and perception technologies.
Level 4: Complete Automation
Level 4: Complete Automation is the final stage of autonomous health. Here, machines don’t just make common diagnostic and treatment decisions, but all the decisions. Only a small set of human doctors and healthcare professionals would be required, and their principal responsibility would be to oversee, improve, and train the algorithms.
This iteration of machines doing most healthcare work seems impossible to imagine, and it may perhaps never manifest. But if our society does ever achieve complete automation of care, our descendants may romanticize human-based healthcare as a charming and quaint period in human history. Just as we think of artisans in the Middle Ages using wooden looms to make textiles, so might our descendants try to imagine humans measuring vitals and drawing blood. In this framework, Complete Automation requires AI to meet all the technical requirements of the high-automation state, and to do so consistently and in every complex situation.
I’m used to being shocked by the rate of innovation. Nonetheless, I believe fully autonomous health, if it ever comes, is several decades away. Even if the underlying sensing and algorithmic decision-making abilities were impeccable compared to our human healthcare workforce, our society is not ready for it. It would take at least a generation of humanity, maybe more, to accept healthcare without human providers.
How We Move Forward
Computer scientists and technologists, including myself, tend to underestimate the clinical and social complexity of healthcare problems. There is a good reason for this human error. Historically, when a machine learning system excels at real-world problems, it often ends up solving the problem in a completely different way from the manual human process. The technical way a smartphone understands and summarizes human voice calls has little in common with human transcribers.
Healthcare is different. Deploying increasing autonomy in an existing healthcare economy is like replacing the jet turbine engine of the airplane with a portable nuclear reactor while it’s still airborne. It’s impossible to recreate an AI healthcare experience and economy in complete isolation from today’s human-centered healthcare experience. By definition, this is a hybrid transition. As I discuss in my new book, The Machines Will See You Now, emerging capabilities need to augment, empower and partner, not replace the role of human beings.
Doctors, nurses, and other healthcare providers will play a pivotal role in accelerating the paradigm shift to high automation and complete automation. They will help create and train these models, and they will critically evaluate their accuracy and safety. Most doctors, nurses, and administrators I meet are accelerants of this change, but want to ensure these technologies are safe, tested and validated.
The business of healthcare will look different too. How venture capitalists and private equity firms invest in hospital systems, large group practices, and innovative companies will continue to evolve. The market valuations and types of capital available are dramatically greater for AI and technology companies than they are for hospitals and medical practices, which today are lower-margin, lower-growth, service-oriented businesses.
Meanwhile, regulators, state, and federal agencies will need to find a balance between protecting consumers from risks and encouraging innovation. AI researchers and large tech companies could spend billions on LLMs fine-tuned for clinical diagnosis and treatment tasks, but without the requisite policy and regulatory changes, regulators could unintentionally handicap the potential impact on better care and lower costs. Or, in the quest for the best model, companies could lose sight of the importance of protecting consumer data and privacy.
As we automate the lifesaving work of doctors and nurses and other practitioners, these are the kinds of questions we must be asking. Automation promises a world with less burden on providers and clinicians, where costs are reined in, and care is more personalized, more effective, and even delightful. But without a road map, and without being clear about what we expect from that transition, we could be writing ourselves into our own sci-fi dystopia.
This essay is adapted from Anmol Madan’s book The Machines Will See You Now: A Human Roadmap to Autonomous Health Care. Copyright © 2026 by Johns Hopkins University Press. Reprinted by permission of Johns Hopkins University Press.
