Anthropic CEO Dario Amodei recently posted on X, “[S]aying that AI will cure cancer is more a cliché than it is inspiring, and most people think it is deceptive.” He conceded that “we haven’t yet delivered on our big promises to benefit the world.” He’s right.
While the invention of useful AI tools has dramatically elevated scientists’ confidence that they’ll eventually find cures, the task is daunting. The number of plausible new biological molecules is greater than the quantity of water molecules in the world’s oceans. As large and powerful as current gigawatt-scale AI data centers are, the challenge of speeding up discovery is so monumental that it invokes the classic movie line, “You’re going to need a bigger boat.”
Finally, a reason to check your email.
Sign up for our free newsletter today.
This summer, Science assessed the state of AI-centric drug discovery, concluding: “Although a wide variety of AI methods have been developed, applied and benchmarked, evidence of their clinically relevant impact is, so far, disappointingly limited.”
And even after we find a promising new drug, real-world clinical trials take time. Nonetheless, AI’s potential impact on health care is one of its most consequential promises. To see where, one only needs to follow the money.
Cost is a key reason that health care has become a political lightning rod. Health-care spending now accounts for nearly 20 percent of the U.S. economy, up from just 5 percent in 1960, when hospitals first began to use computers. Moreover, some 15 percent of the U.S. labor force works in health care, exceeding employment in retail and dwarfing manufacturing jobs.
How could AI tools make a difference with these burdensome costs? Divide health care into three general categories: 1) discovery—the search for drugs and the scientific study of how and why biological things happen; 2) decision-making by doctors, aided by diagnostic tools; and 3) the deployment of solutions through the ecosystem of hospitals, clinics, and other providers.
The combined federal and private spending on discovery is less than 5 percent of total national health-care outlays. Front-line physicians and diagnostic machines and services all combined consume under 20 percent of total outlays. The rest is accounted for by the vast infrastructures of hospitals and clinics—the massive bureaucratic supply chain that delivers the care, services, and cures. Meanwhile, according to the American Hospital Association, administrative overhead alone consumes some 25 to 35 percent across overall health-care spending.
It was administrative bloat, not more money going to research scientists and doctors, that led to the inflated health-care costs of recent decades. While the number of doctors has roughly doubled since 1980, administrative employment jumped over sixfold. No surprise that health care became the Number One category for job creation.
Improving the productivity of a massive bureaucratic administrative system hardly arouses excitement. But making that sector more efficient offers the best opportunities for containing costs and freeing up money. As it happens, AI can far more easily discover ways to perform administrative tasks than discover new cures.
AI is already reducing some of the administrative overhead. On the front lines at doctors’ offices, Stanford’s AI Index project reports that the most widespread adoption of AI clinical tools so far is the “AI scribe.” The AI scribe listens to doctor-patient conversations and automatically generates the necessary clinical documentation. It has brought meaningful and documented reductions in unproductive note-writing time. It arguably also improves the opportunity for sociability since doctors aren’t preoccupied with typing on a computer.
But for both patients and physicians, AI’s more exciting promise is radically improving diagnostics, one of the oldest and biggest challenges in medicine. For one example of what’s coming, check out this tantalizing, AI-generated video demonstration of a new full-body imaging system. The Midjourney Medical ultrasonic imager produces a whole-body image, the company says, 100 times faster than an MRI with competitive resolution. Hyperbole aside, there is every reason to believe that this is the future for medical imaging.
Notably, AI is also improving the operation and diagnostics of all the iconic advances of the twentieth century in detection and diagnosis equipment, physics-based machines (x-rays, MRIs, ultrasound), and chemical/biological analyzers.
One of the first AI applications, in pathology, is now well-established as a valuable clinical amplifier, or assistant. And the early trope that AI would eliminate the need for human pathologists has been firmly debunked. Instead, it improves the efficiency and accuracy of pathologists. AI can and will perform plenty of tasks better than humans, but many of the most complex ones require the magic of collaboration. Johns Hopkins, for example, developed an AI-powered sepsis-prediction system and deployed it across a dozen-plus hospitals, achieving significant reductions in sepsis mortality.
When it comes to diagnostics, the Holy Grail is a kind of Swiss Army knife of imagers and sensors all in one machine/system—a Star Trek kind of tricorder—to yield a rapid and specific diagnosis. Maybe someday.
Meanwhile, AI tools are particularly adept at resolving and rationalizing volumes of messy data typically inherent to natural systems. A review of the technical literature and the venture capital landscape reveals an explosion of activity chasing specific AI medical applications. This is precisely the indicator that rational forecasts are based on, pointing to the inevitability of meaningful advances in health-care productivity.
But the immediate impacts of AI will be in workflows—the boring, critical, administrative stuff. After diagnosis comes the heavy lift of coordinating care across organizations and professionals, along with the unavoidable administrative work of oversight, compliance, tracking, and insurance. It entails dozens of discrete steps, actions, records, and requirements. And every workflow is different and personal. Making workflows easier, transparent, faster, and cheaper, is home-turf territory for AI.
If touting the cancer cure isn’t a get out-of-jail card for today’s beleaguered data centers, just try selling the idea of improving workflows. It’s too late to rename data centers, say, “Medical AI Diagnostic & Imaging Centers,” or “Drug Discovery Centers.” Half-joking aside, this is serious stuff. AI data centers are now a third rail in American politics. It’s no small irony that, just as a huge swath of the population enters old age, we see bipartisan antipathy toward the very technology that, finally, holds potential for bringing health-care costs under control and improving outcomes.
Returning to where we began with promise of “the cure” and the intersection with the politics of AI, and to continue an analogy: Big Tech is clearly finding ways to build “bigger boats,” but they’re still at sea when it comes to navigating the politics.