In a recent interview, Google DeepMind CEO and Nobel laureate Demis Hassabis repeated a conviction I share: the most important use of AI is to improve human health. From AlphaFold’s breakthrough in protein folding to new tools like AlphaGenome, his team is building systems that may transform how we understand diseases and discover medicines. Yet between an elegant protein structure on a computer screen and a patient actually receiving treatment lies a long corridor ruled more by politics and profit than by science. In my previous writing, I explored how AI can assist medical research and make healthcare more personal and accessible. This time, I want to stay with optimism but also ask a harder question: if AI helps us “cure” more diseases, who will live longer as a result, and who will be left behind? The interview opens with a simple but important correction to popular narratives about AI. The biggest impact on our lives, Hassabis insists, will not come from chatbots or image generators. It will come from invisible tools buried in scientific workflows: models for drug design, climate and weather prediction, fusion, quantum computing, and more. AlphaFold is the emblematic example. As an undergraduate in Cambridge, Hassabis encountered the “protein folding problem” from biologist friends: the challenge of predicting a protein’s 3D structure from its amino acid sequence, a grand challenge in biology that had resisted solution for 50 years. Proteins perform almost every function in our bodies, and their shape largely determines what they can do. Traditional methods to finding those shapes, such as X‑ray crystallography, were slow and expensive, sometimes taking years and costing hundreds of thousands of dollars for a single protein. AlphaFold changed that. By 2020, DeepMind had built a system that could predict protein structures with remarkable accuracy, using AI models trained on known structures and sequences. In 2021, during a now-famous internal meeting (fortunately captured on video), Hassabis did a back‑of‑the‑envelope calculation: if AlphaFold could fold a protein in seconds, and if they knew roughly how many proteins were “known to science”, they could, in principle, fold them all within about a year using Google’s computing resources. Instead of waiting for researchers to submit sequences one by one to a server, they could run the entire known universe of proteins and put the results in a free public database. That is exactly what they did. Today, AlphaFold provides predicted structures for almost all known proteins, and the database is regularly updated as new sequences are discovered. More than three million scientists are using it, from plant biologists trying to make crops more resilient to climate stress, to researchers investigating huge, complex structures like the nuclear pore complex that had eluded structural biology for decades. Crucially, Hassabis highlights two communities that particularly benefit from this openness: plant scientists working on critical crops like wheat, who often lack resources for expensive structural experiments, and non‑profit groups focused on neglected diseases such as malaria, Chagas disease, and leishmaniasis, where big pharmaceutical companies see little profit because most patients are poor. For these researchers, receiving high‑quality protein structure predictions for free is a genuine leap forward. They can “jump straight to the problem they’re interested in” instead of spending years on crystallisation. On this level, AI in health is a humanist success story. It democratises a crucial layer of scientific knowledge, giving underfunded researchers a tool they could never have built on their own. But AlphaFold is only the first step of a much longer process. In the same conversation, Hassabis describes how DeepMind spun out Isomorphic Labs to build on AlphaFold and related models for end‑to‑end drug discovery. Knowing a protein’s structure is one piece; designing a molecule that binds to the right part of that structure, strongly and selectively, without causing toxicity elsewhere in the body, is an enormous search problem. Here, AI again offers an impressive acceleration. Isomorphic Labs’ systems can propose chemical compounds that might bind to a target protein, simulate how strongly they bind, and rapidly check predicted interactions against thousands of other proteins to minimise side effects. Instead of testing a handful of molecules in a wet lab, researchers can explore thousands or even millions virtually and only then validate the most promising candidates experimentally. Given that traditional drug development takes around ten years on average and suffers high failure rates, this could save enormous amounts of time, money, and human suffering. According to feedback from one industry scientist, “almost every drug developed from now on will probably have used AlphaFold” at some stage in its pipeline. Isomorphic Labs itself is already working on dozens of programs across cardiovascular disease, cancer, and immunology. And here is where the optimism meets the wall. AlphaFold’s database is open. Isomorphic Labs’ eventual drugs will not be. They will be governed by patents, trade secrets, pricing strategies, reimbursement negotiations, and national health-system budgets. None of these downstream mechanisms is determined by AI. They are determined by corporate boards, regulators, trade agreements, and political choices. In other words, AI can make it cheaper and faster to discover candidate drugs, but there is no automatic mechanism that converts that efficiency into fairer access. Under current incentives, it is just as likely to be converted into higher margins, faster pipelines, and stronger competitive advantages for existing pharmaceutical players. AI may dramatically reduce the scarcity of scientific knowledge while leaving the scarcity of medical access largely untouched. AlphaFold essentially made structural biology knowledge dramatically less scarce. But the distribution of medicine is constrained by patents, regulations, insurance systems, geopolitics, manufacturing capacity, and pricing models. From a humanist perspective, this is the crucial tension. We have AI systems designed by people like Hassabis with a sincere desire to “be of benefit and service to humanity”. Yet we deploy them into a pharmaceutical and health infrastructure that, especially in many high‑income countries, still places shareholders and cost savings above patients who cannot pay. If we are not careful, the age of AI in medicine will quietly become another chapter in an old story: the rich live longer, the poor are left behind. AI‑accelerated drug discovery, precision oncology, and bespoke gene therapies will almost certainly arrive first as premium services in well‑funded health systems and private clinics, marketed as “personalised” or “longevity” solutions for those who can pay. Meanwhile, in under-resourced hospitals and rural clinics, doctors will still be fighting old battles with overstretched staff, intermittent electricity, and shortages of even basic medicines. The same algorithms that help a multinational design its next blockbuster cancer drug could, in principle, also optimise supply chains for essential generics or support community health workers. But there is no automatic mechanism that steers AI in that direction. Without deliberate choices by regulators, funders, and companies, the benefits of these tools will naturally flow along existing fault lines of wealth and geography, deepening them rather than healing them. The interview even reinforces this duality. On the one hand, Hassabis points to an inspiring use cases: under‑resourced plant labs gaining access to structural data, non‑profits working on neglected diseases using AlphaFold to skip years of preliminary work. On the other hand, he acknowledges that many of Isomorphic Labs’ current programmes target familiar, commercially attractive therapeutic areas where wealthy systems spend the most, like cardiovascular disease, cancer, and autoimmune conditions. None of this is a criticism of DeepMind or Isomorphic Labs. They still need to operate within the existing structures. It is a reminder that the struggle over AI and health extends from the lab bench to the governance of health systems, intellectual property, and global markets. Hassabis spends a significant part of the interview discussing AI risks and governance. He worries about two broad categories: misuse of AI by bad actors (including states) and the possibility of powerful, increasingly “agentic” systems going off the rails as they become more autonomous. He calls for international cooperation, safety research, and institutions (something like a “CERN for AGI”) to navigate the transition to artificial general intelligence responsibly. From a humanist and diplomatic perspective, we should extend that logic into the domain of health equity. If we accept that “almost every drug developed from now on” will integrate AI somewhere in its pipeline, then the governance of AI in health is not a niche issue. It is central to the future of global health justice. The AlphaFold database is a glimpse of what it looks like when frontier AI is treated as a global public good. The question is whether we can design analogous public‑interest mechanisms for later stages of the drug pipeline. Some possible directions which diplomats and policymakers could explore: Conditional public funding Support for open tools and neglected diseases Global health impact assessments for AI These are choices about how technical systems are governed and who they serve. Without them, we risk living in a world where AI‑designed cures exist while millions still die for lack of access to older, simpler treatments. What I appreciate in Hassabis’s public appearances is his willingness to talk about responsibility and humility. He speaks openly about wishing AI had stayed longer in the lab so that teams like his could have focused more on tools like AlphaFold, perhaps even “curing cancer”, before the sudden consumer AI “code red” changed priorities. He is candid about the pressures of a commercial and geopolitical race that pushes labs to move faster than their ideal scientific pace. In this context, expanding access becomes a complementary responsibility. His project is to build systems that expand the frontiers of knowledge and make new kinds of treatment possible. Ours, as citizens, diplomats, policymakers, and advocates, is to ensure that those possibilities are not reserved only for those who can afford them. Hassabis says he hopes people will remember his life as “of benefit and service to humanity”. Achieving that hope will require more than clever algorithms and Nobel‑winning breakthroughs. It will require a parallel commitment from many actors to reshape the economic and political machinery around medicine. Otherwise, AI may help humanity discover cures faster than humanity learns how to distribute them fairly. Author: Slobodan Kovrlija
Demis Hassabis, AlphaFold, and the politics of access
Invisible breakthroughs: AI in the lab, not the app
From open data to closed drugs
The risk of an AI‑powered health divide
Humanist governance: finishing the job that AlphaFold starts
When public or multilateral funds support AI‑enabled drug development, access conditions could be built in: tiered pricing, subsidies for low‑income countries, or even non‑exclusive licensing in certain markets.
Building on the AlphaFold model, international initiatives could fund AI platforms specifically targeted at diseases and health needs that are not commercially attractive and keep those tools openly accessible.
Just as we discuss AI safety and alignment for autonomous systems, we might require large health AI projects to conduct and publish distributional impact assessments: who stands to benefit, who is likely to be excluded, and what mitigation measures are planned.Shared intent, different responsibilities
