If we want AI to solve every disease in our lifetime, super-intelligence is not enough. We need an equally ingenious policy to guide the discovery of new drugs.
In 1991, Pfizer tested a new compound called sildenafil as a treatment for heart disease. The drug failed. But the men in the trial reported a weird side effect: erections. Pfizer got the message. Years later, sildenafil was approved as Viagra. Decades later, a diabetes drug called semaglutide showed another curious side effect: patients lost a ton of weight. When molecule became better known as Ozempic or Wegovy, an entire industry was born.
These stories are told as happy accidents. But they shouldn’t be accidents. Thousands of drugs, both failed and approved, sit in trial data that almost nobody reads twice. AI is perfectly suited to reading it a second, third, and billionth time by scanning the trial results for signals that human scientists missed and flagging old compounds that might treat diseases nobody thought to test them against.
Artificial intelligence advocates love to promise that AI will one day cure every disease through sheer genius. But intelligence is not the only bottleneck to medical progress. Curing diseases in our lifetime will require ingenious policy that puts AI to work in the right way. In today’s guest article, Deena Mousa, a program officer at Coefficient Giving, who writes the newsletter Under Development, outlines how government policy could marshal AI for the discovery and development life-saving drugs—including by forcing pharmaceutical companies to quickly make public all clinical trial results, including failed ones, so that we get more discoveries like Viagra and GLP-1s.
Medicine is at the center of almost every optimistic and even utopian vision of how AI could change the world.
Dario Amodei, the CEO of Anthropic, has argued that AI could compress a hundred years of medical progress into a decade, ending most infectious diseases, eliminating most cancers, preventing Alzheimer’s, and doubling the human lifespan. Demis Hassabis, who runs Google DeepMind, told CBS’s 60 Minutes that the decade-long, billion-dollar slog of designing a single drug could soon shrink to months or even weeks. Asked about AI curing all diseases within ten years, he said “I don’t see why not.”
These are exciting ideas. When it comes to exactly how this will happen, however, many forecasts steamroll through the messy details. For example, Amodei’s essay Machines of Loving Grace notes that intelligence is one “factor of production,” and that its complements—like data, or the speed of physical experiments—might constrain even a super-intelligence. However, he argues that a clever enough AI will eventually route around most of these constraints in time by, for example, producing drugs so much more effective than the ones we have today that they will clear trials and win regulatory approval faster.
Maybe.
But the barriers to technology solving all of our problems are all too easy to underestimate. For example, after the model Watson won Jeopardy! in 2011, IBM trained a version of it to recommend cancer treatments and promised that it would bring expert-level care to hospitals everywhere. It didn’t really work. One limitation was that the system had been trained on a small number of hypothetical patient cases devised by a few doctors, rather than on large amounts of real patient data.
Fortunately, we don’t have to wait and hope for a superintelligent system to somehow bypass outdated institutions and fill in missing information. Deliberate choices today, especially by the US government, can make it more likely that AI contributes to a golden age of drug discovery and life extension in the next few years.
Here are three things we should do.
1. Build an “internet of the human body”
Large language models do better with more data. For example, they are proficient writers precisely because they have been trained on the internet, which, as we all know, has quite a bit of writing.
The same principle applies to science. AlphaFold can predict the 3D structure of proteins, in large part because of the existence of the Protein Data Bank, where roughly 200,000 protein structures have been painstakingly organized by researchers over 50 years.
Unfortunately, the Protein Data Bank is a uniquely comprehensive project, and most of human biology is not similarly well-documented. Biological data often has to be generated through physical lab work, for example by culturing cells or dosing an animal and then waiting for a result. The time, labor, equipment, and materials required to do this are much greater than those required to generate text or code. The human body is also made up of many complicated interacting systems, so it’s very hard to isolate causal effects, and many of our measurement tools are imperfect and noisy.
Public data is a public good; no one can be prevented from using it, and its benefits accrue widely. As a result, private actors do not have much of an incentive to produce it. When data is produced, it’s often locked up behind regulations, as in the case of individual health records, or behind commercial secrecy, as in the case of many pharmaceutical trials. So how could the US government stimulate the creation of an “internet” of public data about the human body that we could use to train AI?
We could start by enforcing the law. American regulations already require the results of most clinical trials to be posted publicly within a year. This rule is widely flouted. Only about 41 percent of trials are reported on time, and roughly a third are never reported at all. Many of these trials were funded by the U.S. government itself. The law allows fines of roughly $10,000 a day for noncompliance, but the FDA has issued only a handful of formal notices and, as far as the public record shows, has never actually collected a penalty.
We could extend these disclosure requirements to failed clinical trials and other shelved data. Drug companies sit on decades of results for compounds that failed or were shelved for other reasons. That information would be extremely valuable in the public domain, both because it contains a record of what didn’t work, and because those trials often capture many secondary outcome measures on which the drug might have shown more promise, but which were less of a commercial priority for that particular company.
Many people today are familiar with the story of how GLP-1 drugs were originally intended to treat type 2 diabetes until scientists observed significant weight loss among patients. In another famous case, Pfizer’s trial for sildenafil as a heart drug and treatment for angina failed, but men in the trials reported unexpected erections. The compound was later reborn as Viagra. Pfizer pursued that use because it addressed a large and profitable market. But if the signal had been more obscure or seemed less financially worthwhile, the research and drug could easily have been shelved for years. If all clinical trials were made public, private and philanthropic actors alike could set frontier AI models loose on them to identify similar findings. In short, imagine all of the treatments and cures we might have for any number of conditions if the serendipitous discoveries like Ozempic and Viagra became the norm.
The US government could also expand funding to build new publicly accessible medical datasets, like the Protein Data Bank. The Arc Institute, a nonprofit lab, has assembled an open “Virtual Cell Atlas” of measurements from more than 600 million cells (including a dataset mapping how 50 cancer cell lines respond to more than a thousand different drugs)—and released it for anyone to use in training models. Directly funding the creation of specific datasets like this is one path. Another is to create prizes and advance purchase commitments specifically for datasets, rather than focusing only on finished drugs. A model trained on a map of how drugs affect living cells could help predict how an untested compound will act on cells before researchers run costly and time-consuming experiments. Narrowing a vast search space to the few candidates worth the lab time would significantly accelerate the discovery of new medicines.