How Superintelligence Will Eradicate Disease: AI's Boldest Promise

In December 2020, DeepMind’s AlphaFold predicted how a protein folds from its amino acid sequence, a problem biologists had worked on for half a century. Within two years it had mapped the structures of virtually every known protein, over 200 million of them. Experimental crystallographers would have needed centuries.

AlphaFold is narrow AI. The question researchers, investors and policymakers are now arguing about is what happens if artificial intelligence becomes smarter than all humans combined, and is pointed at disease.

Deployed today

Traditional drug development takes 10 to 15 years at an average cost of $2.6 billion, with a 90% failure rate. Insilico Medicine took a pulmonary fibrosis candidate into Phase 2 trials in under 30 months. Recursion Pharmaceuticals runs millions of experiments a week and uses machine learning to find therapeutic relationships no researcher would have spotted. Isomorphic Labs, the DeepMind spinout, uses AlphaFold’s output to design molecules.

In radiology, deep learning systems detect lung nodules and brain haemorrhages with sensitivity comparable to board-certified radiologists, and they do not get tired at the end of a shift. In pathology, AI reaches diagnostic accuracy above 95% for certain cancers. In dermatology, smartphone tools classify skin lesions at specialist level, in places that have never had a specialist.

Genomic sequencing, health records, wearable data and AI analysis now combine to match therapy to an individual’s biology. Tempus matches cancer patients with the therapies most likely to work on their specific tumour profile. Foundation Medicine identifies actionable mutations across hundreds of cancer genes. These are commercial services with patients on them, not research programmes.

The superintelligence claim

Superintelligence, meaning AI that exceeds the cognitive capability of all humans combined, is still argued over. Demis Hassabis of DeepMind has said it could arrive within a decade. Others put it centuries out.

Applied to medicine, such a system could in principle model human biology as one integrated system, every gene, protein interaction, metabolic pathway and environmental variable, and identify the interventions that prevent or cure a given disease. Cancer is thousands of diseases rather than one; a system working at that scale could map and counter them faster than multi-year clinical trials allow. Alzheimer’s has resisted four decades of concentrated research; it might give way to a system that simulates neurological processes at a fidelity no lab can approach.

The most provocative version of the claim is that superintelligence could treat ageing itself. Ageing is a set of identifiable molecular processes: telomere shortening, cellular senescence, mitochondrial dysfunction, epigenetic drift, accumulated DNA damage. A system able to model all of them at once could, in theory, turn ageing into a treatable condition.

Limits

Predicting protein structure is not predicting protein function, and predicting function is not predicting what a molecule does inside a living body. Drug candidates that look excellent in silico fail in human trials all the time. AI speeds up hypothesis generation. The biology stays unpredictable.

The data is bad in specific ways. Medical datasets have gaps, biases and inconsistencies. Clinical trial populations have historically under-represented women, racial minorities and the elderly. A model trained on that data reproduces the bias at scale.

Regulators are behind. The FDA has approved over 800 AI-enabled medical devices, under a model designed for technologies that do not change after approval. Modern AI systems keep learning after deployment, and no one has settled how to certify that. Japan’s PMDA faces the same problem, plus the need to line up with international standards.

And the timeline is genuinely unknown. Current systems work by pattern recognition, which may not be the causal reasoning that curing cancer or Alzheimer’s requires. How far “very good AI” is from “AI that understands biology better than every scientist who has ever lived” is an open question, and the serious researchers say so.

Japan

Japan has 29% of its population over 65 and a life expectancy of 84.6 years. Healthcare expenditure passed JPY 46 trillion ($310 billion) in 2025, with age-related conditions driving most of the growth. No other major economy has this combination yet.

The response is already under way. The Fugaku supercomputer at RIKEN’s Center for Computational Science in Kobe has been used for drug simulation, modelling how candidate therapeutics interact with target proteins at atomic resolution. Takeda and Daiichi Sankyo have AI drug discovery partnerships. The University of Tokyo is building clinical decision support systems for Japan’s specific disease burden.

The PMDA has been developing frameworks for AI-based medical devices, drawing on the precedent of the 2014 Act on the Safety of Regenerative Medicine. Willingness to adopt is not the constraint in Japan. The constraint is whether adoption moves as fast as the demographics.

Access, bias, liability, control

If AI-driven therapies reach only wealthy countries or premium patients, the best care gets dramatically better while most of the world is left where it was. Healthcare is a human right, and a technology distributed that way cuts against it.

Dermatology AI trained mostly on lighter skin tones has shown lower accuracy for darker-skinned patients. A more capable system trained on the same skewed data would apply the same error to more patients.

When an AI recommends a treatment or marks a scan benign and is wrong, who is responsible? The frameworks that govern human physicians do not map onto algorithmic decisions, and in healthcare a black box can kill someone.

A handful of companies control the most advanced AI research on earth. If superintelligent medicine comes out of those labs, a large part of global health governance would sit with organisations that have no democratic mandate.

Rules written before the technology

Nobody is building superintelligence in one step. What exists is a stack of smaller advances in models, datasets, compute and molecular design, each one useful on its own.

BenevolentAI identified baricitinib as a potential COVID-19 treatment early in the pandemic; the finding was later confirmed in clinical trials. AlphaFold’s database now serves over two million researchers. AI-designed antibodies are in clinical trials.

The FDA’s device rules, the PMDA’s frameworks and the data-sharing terms agreed between hospitals and the companies named above are being written this decade. Whether superintelligence arrives in ten years or fifty, it will arrive into whatever those documents say.

Join the Conversation

On April 26, 2026, the Tech for Impact Summit convenes senior executives, policymakers and technologists at Tokyo Garden Terrace Kioi Conference. The theme is “Beyond Boundaries: Building 2050 Together”. The boundary between human and artificial intelligence is on the agenda.

Confirmed speakers include Taro Kono (former Minister of Digital Affairs), Charles Hoskinson (Cardano), Yoshito Hori (GLOBIS), Kathy Matsui (MPower Partners), Ken Suzuki (SmartNews), Jesper Koll (Monex Group), Sota Watanabe (Astar/Startale), and Hiroshi Aoi (Marui Group).

If you run a healthcare enterprise, a pharmaceutical company or a policy institution dealing with AI in medicine, this is the room for that conversation.

Explore partnership and membership opportunities →

Watch highlights from previous summits: youtu.be/ujy7ZXflrt4

The Tech for Impact Summit is an invitation-only executive gathering taking place April 26, 2026, in Tokyo as a partner event of SusHi Tech Tokyo. Learn more at tech4impactsummit.com.