Much of the public conversation about artificial intelligence has settled into anxiety, whether job losses, deepfakes, or a vague sense that something is being taken from us. Those concerns deserve serious attention. But they describe only one edge of a much larger picture. Step into the laboratories, clinics, and research institutes where AI is actually being put to work, and a different story comes into focus: AI is emerging as the most powerful catalyst we have ever built for solving humanity's hardest problems. And increasingly, AI systems are compounding on one another, each advance becoming the tool that unlocks the next.
The problems in question are not abstract. They are the diseases that have inflicted generational injustice on communities without the resources to fight back. They are the energy and climate challenges that will define whether the next century is livable. They are the millions of people who fall ill in places with no specialist within a hundred miles. These are problems humanity has thrown its best minds at for decades, and where progress has been throttled less by a lack of will than by a lack of reach. What has changed is that computation has grown capable enough to explore possibilities at a scale no human team ever could, and to hand its results to the next system in line.
This paper makes that case in two movements. First, a brief survey of where AI is already bending the curve on humanity's critical problems, enough to show the pattern is real and broad. Then a closer look at two results from the past year in human health, chosen because they demonstrate something biology long assumed impossible: that damage once considered permanent can, in fact, be reversed.
None of the examples below is science fiction, and none is finished. Each is a real, published, or deployed system doing work that was out of reach a few years ago. Read together, they reveal a single mechanism at play: AI lets us search spaces of possibility (molecules, materials, forecasts, diagnoses) that are simply too vast for human effort alone.
Malaria, tuberculosis and other "neglected" diseases kill overwhelmingly in the world's poorest communities, precisely because there is little commercial incentive to develop drugs for them. AI is rewriting that arithmetic: models now screen millions of candidate molecules for a fraction of the traditional cost, and free, open platforms put that power in the hands of researchers who could never afford a corporate pipeline.
e.g. Medicines for Malaria Venture and deepmirror's free AI drug-discovery platform; AlphaFold's open database of 200M+ protein structuresFusion promises near-limitless clean power, but controlling a plasma hotter than the sun's core is a fiendish real-time problem. AI has learned to hold that plasma stable inside an experimental reactor. In parallel, AI has proposed hundreds of thousands of new stable materials (candidates for better batteries, solar cells and superconductors), accelerating a materials search that used to take decades.
e.g. DeepMind and the Swiss Plasma Center's learned tokamak control; DeepMind's GNoME materials discoveryAccurate warning saves lives. AI weather models now produce ten-day global forecasts in under a minute, matching or beating the world's leading physics-based systems that need supercomputers and hours to run. Cheaper, faster, more accessible forecasting means earlier warnings of storms, floods and heatwaves for the communities most exposed to them.
e.g. DeepMind's GraphCast and GenCast forecasting modelsMuch of the world has no radiologist, no ophthalmologist, no pathologist within reach. AI is beginning to close that gap: screening chest X-rays for tuberculosis, reading retinal scans for diabetic blindness, and triaging patients from a smartphone. The specialist's judgment, encoded and made portable, can now travel to a rural clinic that has never had one.
e.g. WHO-endorsed AI TB screening from chest X-rays; automated diabetic-retinopathy screening deployed in India and ThailandThe through-line is worth naming, because it also points to where this is heading. In each case AI is not replacing the scientist, the engineer, or the doctor. It is extending their reach into territory that was previously unsearchable. And the systems are beginning to build on each other: a model that predicts protein structures feeds a model that designs drugs; a model that discovers materials feeds the engineering of a cleaner grid. AI working alongside AI is becoming a compounding engine. And nowhere is that compounding more vivid, or more human, than in the effort to repair the body itself. The rest of this paper looks closely at two examples from that frontier.
In 2006, the Japanese scientist Shinya Yamanaka made one of the most consequential discoveries in modern biology. He showed that an ordinary adult cell could be coaxed back to a youthful, embryonic-like state using just four proteins: Oct4, Sox2, Klf4, and c-Myc, now known collectively as the Yamanaka factors. The work won the Nobel Prize and opened a tantalising question. If a cell can be reset, can aging itself be partially rewound? The catch is that the natural factors are blunt instruments. They are slow, inefficient, and, pushed too far, can erase a cell's identity entirely or tip it toward cancer. For nearly two decades, improving them was painstaking, one hand-designed variant at a time, with success rates that rarely rose above single digits.
In August 2025, a collaboration between OpenAI and Retro Biosciences reported a different approach. Rather than relying on human intuition to guess which mutations might help, they trained a specialised model, GPT-4b micro, on the language of proteins, the way a text model learns the patterns of human language. They then asked it to redesign the Yamanaka factors from the ground up. The model proposed sequences no biologist would have written: its versions of SOX2 and KLF4, nicknamed RetroSOX and RetroKLF, differed from the natural proteins by more than a hundred amino acids on average, yet folded and functioned better.
The numbers matter because they describe a change in kind, not just degree. Traditional laboratory screens tend to surface a superior variant less than one time in ten. The model's designs beat the natural SOX2 protein in more than 30% of attempts, a hit rate that compresses what might have been years of trial and error into a handful of experiments. When the engineered factors were applied to cells taken from donors over the age of fifty, more than 85% activated critical pluripotency markers within twelve days, and crucially, more than 85% of those cells retained healthy, stable chromosomes. Efficiency without genomic chaos is precisely the balance that had eluded the field.
"When researchers bring deep domain insight to our language-model tooling, problems that once took years can shift in days." — Boris Power, Research Partnerships, OpenAI
It is worth being precise about what this is and is not. This is laboratory science, performed on cells in dishes, not a therapy available to patients. No one has been rejuvenated. The path from a better reprogramming factor to a safe medicine remains long and uncertain. But the significance is real. An AI model, given the accumulated grammar of biology, proposed protein designs that outperformed the products of billions of years of evolution and two decades of expert refinement, and it did so fast enough to change how the work feels. The bottleneck in this kind of research has always been the size of the search space. A protein a few hundred units long has more possible sequences than there are atoms in the universe; human intuition can sample only a vanishing fraction. What AI supplies is not magic but reach.
If the first story is about a cell's software, the second is about its physical architecture. Between and around our cells lies the extracellular matrix: the mesh of long-lived proteins, principally collagen and elastin, that gives skin its elasticity, arteries their flexibility, and organs their shape. These proteins are laid down early and rarely replaced, which makes them extraordinarily durable and, unfortunately, extraordinarily vulnerable. Over decades they accumulate chemical scars called advanced glycation end products, or AGEs: the same class of browning reaction that turns toast golden, running slowly inside living tissue. As AGEs build up, blood vessels stiffen, skin loses its resilience, and inflammation rises. Since the 1980s, this kind of damage was regarded as effectively permanent, a fixed cost of having lived.
On 14 July 2026, Revel Pharmaceuticals, working with Calico Life Sciences (Alphabet's longevity company) and the University of Colorado Anschutz Medical Campus, published a paper in Nature Communications that punctures that assumption. The team engineered a novel enzyme, CMLase, designed to seek out one of the most abundant of these scars, Nε-carboxymethyl-lysine (CML), and chemically undo it, restoring the original, undamaged protein without harming the surrounding tissue.
Here the role of computation is different from Part One, and it is worth being exact rather than sweeping. CMLase was not written by a language model. It was built through directed evolution (an iterative process of mutation and selection) but at a scale only modern computing makes possible. The team began by computationally screening roughly 45,000 protein structures to find a promising starting scaffold, then ran five rounds of directed evolution across more than 500 million enzyme variants. This is high-throughput, algorithm-guided protein engineering: the search is steered and evaluated by software operating far beyond human scale, even where the final selection is grounded in wet-lab measurement rather than pure prediction. It sits squarely on the trajectory of the survey above: the same path along which AI protein-design tools like AlphaFold and a new generation of generative enzyme models are rapidly moving the entire field, each advance becoming the foundation for the next.
"This class of damage has been seen as a fixed part of aging since the 1980s. What we've shown is that CML damage in human tissue can, in fact, be reversed under laboratory conditions." — Aaron Cravens, Chief Executive Officer, Revel Pharmaceuticals
The researchers are candid about the distance still to travel. CMLase is a proof of concept, and the team frames it as the first entry in a hoped-for systematic toolkit: "one enzyme per type of damage." Harder targets loom. Glucosepane and related crosslinks actually weld proteins together rather than simply modifying a single site, and undoing them is a steeper engineering problem. And moving from tissue in a dish to a treatment in a body raises real questions about delivery, immune response, and whether an enzyme can even reach damage buried deep in intact matrix. None of that is solved. What has changed is that the reversal is now demonstrably possible: a door long assumed to be walled shut turns out merely to have been locked.
Set everything in this paper side by side: the drug hunt for neglected diseases, the plasma held steady, the forecast that arrives in time, the two aging processes reversed in a dish. A single pattern emerges that is larger than any one result. In each case, a problem long considered fixed yielded not to a lone breakthrough but to a change in how the search was run. Humanity's hardest problems have almost always been combinatorial: the number of possible drugs, materials, and interventions is astronomically larger than any team could test in a lifetime. For the whole history of science, we have navigated that vastness with intuition, luck, and patience. What is genuinely new is that computation can now explore it at a scale that changes outcomes, and that each system's output becomes the next one's starting point.
This is the part of the AI story that the anxious narrative misses. Public debate tends to treat AI as something that acts on people: sorting them, replacing them, watching them. But the most consequential applications may be the ones that act for them, in domains where human capability simply runs out of reach. No researcher can personally evaluate 500 million enzymes, reason through a hundred simultaneous mutations, or screen millions of molecules for a disease the market has ignored. These are not failures of effort; they are limits of a single mind and a single lifetime. AI, used well, lifts those limits, and the early returns are arriving in the arenas that matter most: the health of the poor, the stability of the climate, the reach of a doctor.
Emerge exists because the people who stand to benefit most from this technology deserve to understand it clearly, without the breathless promises of the hype cycle or the reflexive dread of the backlash. The truth is more interesting than both. In quiet laboratories and under-resourced clinics, using tools that barely existed three years ago, people are turning "impossible" into "done," and "irreversible" into "reversed." That is not a threat to be managed. It is a frontier to be understood, and, we would argue, a reason for genuine, evidence-based hope.
This whitepaper summarises publicly reported research and deployments for general education. It describes laboratory findings and early-stage systems and does not constitute medical advice or a claim about available treatments. © 2026 Emerge.
Tell us what you're working towards, and we'll help you build a hyper-efficient, connected organisation — supported end to end, wherever your people are.
Get in touch · 1300 EIT AUS