SWITCHED ON
The daily technology series nobody asked for but everyone needed
The Robot Doctor: AI in Medicine and the Future of Healthcare
AI can already detect certain cancers more accurately than radiologists, design drug candidates in days rather than years, and predict patient deterioration before clinicians notice the signs. The question is not whether it works. It is who it works for.
Medicine is the domain where AI's potential to save lives is most concrete and most immediate. It is also the domain where the gap between what AI can do in a well-resourced research hospital and what it can do in the healthcare systems where most of the world's sick people are actually treated is largest. Both things are true, and only one of them gets most of the coverage.
The Season Two finale asked what we keep when the machines take everything else — embodiment, genuine relationship, the process of making rather than just the product, and the irreducibly political choices about who decides how technology develops. Having answered that, or at least having asked it seriously, we are back. Season Two continues. There are fifteen more episodes to run and the list of consequential technology topics has not got any shorter while we were being philosophical. Today we are in the hospital. Specifically, we are examining what artificial intelligence is doing to medicine — the diagnostic tools that are outperforming human specialists in certain tasks, the drug discovery platforms compressing development timelines from decades to years, the robotic surgical systems operating with precision no human hand can match, and the healthcare equity question that all of this raises and that the technology press almost never asks. Welcome back. The stakes are, as usual, considerable.
01 — What AI Diagnostics Can Actually Do
The evidence for AI's diagnostic capability in specific medical imaging tasks is now sufficiently robust to move beyond the question of whether it works and into the harder questions of how it is deployed, for whom, and under whose oversight. The headline results have been striking: a 2020 Nature Medicine study found that a Google-developed AI detected breast cancer in mammograms with fewer false positives and fewer false negatives than a panel of six radiologists. A DeepMind system demonstrated performance equivalent to or exceeding expert ophthalmologists in detecting over fifty eye diseases from optical coherence tomography scans. AI systems for detecting diabetic retinopathy — a leading cause of blindness that is entirely preventable with early detection and treatment — have received regulatory approval in multiple jurisdictions and are being deployed in screening programmes in countries including the US, UK, and Thailand.
The mechanistic explanation for these results is straightforward: AI systems trained on large, well-labelled datasets can identify subtle patterns in images that are too faint or too complex for reliable detection by human visual processing. They do not get tired, do not have bad days, and do not anchor to initial impressions in the way that human cognition tends to. For tasks involving pattern recognition in large volumes of standardised data — reading scans, analysing pathology slides, interpreting ECGs — these are genuine performance advantages with genuine clinical implications.
The AI diagnostic results should be read as establishing that AI is a powerful tool for specific image analysis tasks, not that it is a replacement for clinical judgment across the full complexity of patient assessment. The radiologist reading a scan is also taking a history, integrating prior knowledge, communicating with a frightened patient, and exercising the kind of contextual clinical judgment that performance on isolated imaging benchmarks does not assess.
The deployment gap is significant. The research demonstrating AI diagnostic performance has been conducted overwhelmingly in high-income country hospitals with large, well-curated datasets from patients whose demographics reflect those hospitals' populations. When AI diagnostic tools trained on these datasets are deployed in different clinical environments — different imaging equipment, different patient demographics, different disease prevalence — performance can degrade substantially. The diabetic retinopathy AI trained primarily on data from one population has shown reduced accuracy when deployed in populations with different retinal characteristics. This is a known challenge that the field is actively addressing and has not yet fully resolved.
02 — Drug Discovery: The AlphaFold Revolution
DeepMind's AlphaFold protein structure prediction system, which solved a fifty-year-old grand challenge in biology by accurately predicting how proteins fold from their amino acid sequences, has had a more immediate and arguably more significant impact on medicine than any of the diagnostic AI systems. Understanding protein structure is foundational to drug development — drugs work by binding to specific proteins, and knowing the three-dimensional structure of a target protein is essential for designing a molecule that fits it precisely. Before AlphaFold, determining protein structure experimentally took months to years and significant resources. AlphaFold can predict structures in hours, and its predictions have been made freely available to researchers globally.
The downstream effects on drug discovery are already visible. Isomorphic Labs, the DeepMind spin-off applying AlphaFold to drug development, has signed multi-billion-dollar partnerships with Eli Lilly and Novartis to develop AI-designed drug candidates. Recursion Pharmaceuticals, Exscientia, and a growing ecosystem of AI-first drug discovery companies are running molecular design processes that were previously impossible at the speed and scale now achievable. The traditional drug discovery process — from target identification to clinical candidate — has historically taken five to ten years and cost hundreds of millions of dollars before a molecule reaches human trials. AI-assisted discovery is compressing the pre-clinical stages significantly, though the clinical trial process itself, which is where most drugs fail, has not been similarly accelerated.
03 — Robotic Surgery and Precision Intervention
Robotic surgical systems — led by Intuitive Surgical's da Vinci platform, which has performed millions of procedures across multiple surgical specialties — are not new. What is new is the integration of AI into surgical guidance, enabling systems that can provide real-time anatomical mapping, identify structures at risk during surgery, and in some cases perform specific sub-tasks autonomously under surgeon supervision.
The Smart Tissue Autonomous Robot, developed at Johns Hopkins, performed supervised autonomous laparoscopic bowel surgery on pig models with results that were statistically superior to both human surgeons and robot-assisted human surgeons on specific outcome measures — wound consistency, leak rates, and suture spacing. This is a research result on animal models, not a deployed clinical system, and the regulatory and liability pathway for genuinely autonomous surgical systems in human patients is long and appropriately demanding. What it demonstrates is the trajectory: from tools that enhance and extend surgeon capability toward systems that perform specific well-defined tasks with precision that human motor control cannot reliably match.
The equity implications of robotic surgery are straightforward and largely unaddressed: the da Vinci system costs approximately $1.5 million to purchase and significant ongoing maintenance costs. It is concentrated in well-resourced hospitals in wealthy countries. The surgical precision advantages it confers — reduced blood loss, smaller incisions, shorter recovery times — accrue primarily to patients who can access those hospitals. The distribution of surgical innovation tracks the distribution of wealth, not the distribution of surgical need.
04 — AI in Clinical Decision Support
Beyond diagnostics and drug discovery, AI is being deployed across healthcare systems in clinical decision support roles — tools that alert clinicians to deteriorating patients, flag potential drug interactions, suggest differential diagnoses, and identify patients at high risk of specific adverse outcomes. The Epic and Cerner electronic health record systems, which between them hold records for the majority of hospital patients in the United States, have integrated AI alert systems that are now making clinical recommendations to clinicians on a scale that dwarfs any of the more publicised diagnostic AI deployments.
The evidence for these systems is more mixed than the headline diagnostic AI results. A 2021 study of a sepsis prediction algorithm deployed across hospitals in the University of Michigan system found that the algorithm generated alerts for a large number of patients who did not develop sepsis and missed a significant number who did — a false positive rate and false negative rate that, when examined alongside the alert fatigue it generated in clinical staff, raised serious questions about net clinical benefit. Subsequent studies of other EHR-integrated AI systems have found similar patterns: algorithms that perform well on the retrospective datasets used to train and validate them show reduced and variable performance in prospective deployment, partly because healthcare data is messier and more varied in real-world use than in curated research datasets.
05 — The Health Equity Question
The health equity implications of AI in medicine are the topic that receives the least coverage relative to its importance, and the most important thing this episode can add to the breathless innovation coverage that dominates the field.
Medical AI systems are trained on data from the healthcare systems in which they are developed — overwhelmingly the healthcare systems of wealthy countries, with patient populations that reflect the demographics of those systems. The result is AI tools that perform best for the populations best represented in training data — typically white, typically affluent, typically in high-income countries — and worse for populations that are underrepresented. Pulse oximeters, widely deployed and trusted clinical tools, were recently found to overestimate blood oxygen saturation in patients with darker skin tones at a rate that led to delayed treatment of hypoxia. AI systems built on physiological training data that inherits this bias will propagate it.
The global health equity dimension is starker. The AI drug discovery platforms developing the next generation of treatments are funded by the pharmaceutical industry and oriented toward the conditions — cancer, metabolic disease, neurological conditions — that generate the largest revenues in wealthy-country healthcare markets. Neglected tropical diseases, which affect hundreds of millions of people in low-income countries, attract a small fraction of AI-assisted drug discovery investment proportional to their disease burden, because the commercial model does not produce adequate return. The AI healthcare revolution, as currently constituted, is disproportionately a revolution in healthcare for the already well-served.
Tomorrow we are going somewhere that connects AI's role in medicine to a broader question about trust in institutions — specifically the relationship between technology companies and the scientific publishing and research ecosystem. Who funds AI research, what gets published, what gets suppressed, and whether the replication crisis in science is being made better or worse by the AI tools now being used to conduct and evaluate it. See you then.
Switched On is a daily technology series covering the ideas, systems, and arguments shaping the digital world. Opinionated. Witty. Occasionally wrong. Always worth the argument.
