Could AI Cure Every Disease In The Next 10 Years?

Gavin’s Account

What The Claim Is

The claim circulating online comes from Anthropic CEO Dario Amodei, but the social media version makes it sound more definite than his original argument. In his 2024 essay Machines of Loving Grace, Amodei proposed that sufficiently capable AI could increase the rate of biological and medical discovery by roughly tenfold. His estimate was that progress that might normally require 50 to 100 years could potentially occur within 5 to 10 years if AI became capable enough and researchers could use it throughout biology and medicine. This was a prediction about what future AI could enable, not evidence that most diseases will be cured within a specific decade.

His reasoning centers on how much of biomedical research involves processing information, generating hypotheses, designing experiments, analyzing results, and deciding what to test next. Biology now produces genomic data, protein structures, medical images, chemical libraries, clinical records, and more published research than any individual team can realistically process. AI systems can search and combine those information sources at a scale that humans cannot. If those systems become better at biological reasoning, the potential gain comes from helping researchers make better decisions faster across thousands of projects at once.

⸻

There Is Already Real Research Behind The Idea

• The best evidence that this direction is realistic comes from areas where AI has already shortened research problems that previously required much more manual work. Protein structure prediction is the clearest example. A protein’s three-dimensional structure can determine how it functions, what molecules interact with it, and whether it could become a useful drug target. Experimental structure determination remains extremely valuable, but it can require substantial laboratory work.

AlphaFold2 showed that AI could predict many protein structures with accuracy approaching experimental methods. DeepMind then used the system to generate predicted structures across a huge portion of the human proteome, giving researchers structural information for proteins that had never been experimentally characterized in detail.

Jumper et al., Nature, 2021.

PMID: 34265844

Tunyasuvunakool et al., Nature, 2021.

PMID: 34293799

AlphaFold3 extended the concept beyond predicting isolated proteins. The newer model was designed to predict interactions involving proteins, DNA, RNA, small molecules, ions, and other biological components. Drug development depends heavily on these interactions because researchers need to know more than what a target looks like. They need to understand which molecules bind to it and how those interactions might alter its function.

Abramson et al., Nature, 2024.

⸻

AI Is Starting To Design Biology Too

AI research has moved beyond predicting structures that already exist and into designing new biological molecules. RFdiffusion is one example. Researchers used generative modeling to design proteins with specified structures and binding properties, then experimentally produced many of those proteins to determine whether the computer-generated designs behaved as expected. Some of the generated proteins folded correctly and performed their intended binding functions.

Watson et al., Nature, 2023.

PMID: 37433327

That changes the research process considerably. Traditional drug and protein development often involves screening large numbers of existing molecules and modifying the best candidates over repeated experiments. Generative systems can work in the opposite direction by starting with the biological function researchers want and proposing molecules designed around that objective. The laboratory still determines whether the design is useful, but AI can expand the number of realistic candidates researchers can consider before committing resources to synthesis and testing.

⸻

AI-Designed Drugs Are Reaching Human Trials

• Human clinical research is a much higher standard than computer prediction, and this is where claims about AI curing disease need more restraint. One of the more interesting recent examples is rentosertib, previously known as ISM001-055, which was developed for idiopathic pulmonary fibrosis. AI methods were involved in identifying TNIK as a therapeutic target and developing a molecule against it.

A randomized phase 2a trial published in 2025 included 71 participants who received rentosertib or placebo for 12 weeks. The highest-dose group had a mean forced vital capacity change of approximately +98.4 mL compared with -20.3 mL in the placebo group. The study was small and short, so it cannot establish long-term clinical benefit or prove that the drug will succeed in later development. It does show that an AI-assisted target and drug discovery process can progress far enough to produce a molecule tested in a randomized human trial.

Ren et al., Nature Medicine, 2025.

PMID: 40461817

AI has also identified unexpected drug candidates from existing chemical libraries. In 2020, researchers used a deep-learning model to screen molecules for antibacterial activity and identified halicin, a compound that had not originally been developed as an antibiotic. Halicin showed antibacterial activity in laboratory experiments and animal models, illustrating how AI can identify relationships in chemical data that researchers may not have prioritized through traditional screening.

Stokes et al., Cell, 2020.

PMID: 32084340

⸻

Why This Is Promising

• The major opportunity is not that an AI model will suddenly announce the cure for every disease. The opportunity is that AI can potentially speed up many individual parts of biomedical research at the same time. Drug development contains enormous numbers of decisions about which disease target to pursue, which molecule to synthesize, which experiment to run, which candidate to abandon, and which population should enter a clinical trial. Small improvements at each of those stages can compound across an entire development program.

AI systems could help researchers work through areas such as:

1. Reading and comparing enormous amounts of scientific literature and biological data.

2. Predicting protein structures and molecular interactions before expensive laboratory experiments.

3. Screening very large chemical libraries and removing weak candidates earlier.

4. Designing new proteins, antibodies, peptides, and small molecules around specific biological targets.

5. Identifying biomarkers that separate broad diseases into more specific biological subtypes.

6. Choosing experiments that provide the most useful information for the next stage of development.

7. Matching clinical trial participants to molecular features that may predict treatment response.

The speed increase could become larger as AI systems are connected with laboratory automation. A model could propose an experiment, robotic equipment could perform it, the resulting data could be analyzed automatically, and the next experiment could be selected based on those results. Instead of having long delays between hypothesis generation, experimental execution, analysis, and the next round of testing, parts of that cycle could happen much faster.

That is probably the strongest part of Amodei’s argument. Medical progress does not require AI to know every biological answer. A major acceleration could happen if AI allows researchers to test far more good ideas while spending less time and money on weak ones.

⸻

Where The Prediction Runs Into Reality

• Biology cannot be accelerated at the same rate as computer processing because many medical questions require physical experiments. An AI can design a molecule quickly, but researchers still need to manufacture it, confirm its chemistry, test its pharmacology, evaluate toxicity, understand how it is metabolized, and determine whether it reaches the intended tissue. If the candidate survives those stages, human clinical trials still have to establish whether it improves an outcome that matters.

Some medical endpoints cannot be compressed very far because time itself is part of the experiment. Cancer recurrence, cardiovascular events, Alzheimer’s progression, fibrosis, bone remodeling, and longevity can require months or years of observation. A computer model can help predict which treatment is likely to work, but a prediction cannot completely replace waiting to see whether a disease progresses differently in real people.

There is also a large gap between correctly predicting a molecular interaction and producing a successful drug. A molecule can bind its intended target and still fail because of toxicity, poor absorption, metabolism, immune effects, unexpected interactions, or inadequate exposure in the correct tissue. AI can improve candidate selection without eliminating these biological failure points.

As of 2026, AI-assisted and AI-designed drug candidates have entered human clinical development, but the field has not produced evidence that AI has removed the high failure rate that exists throughout pharmaceutical development. Some programs will succeed and others will fail. The meaningful question is whether AI can improve the percentage of good candidates entering expensive human trials and shorten the amount of time required to identify them.

⸻

What This Could Mean For Disease Research

• Cancer provides a good example of why the phrase “AI could cure most diseases” needs context. Cancer is not one biological problem. Different cancers contain different mutations, signaling pathways, immune environments, and treatment responses. AI could still produce a large improvement without finding one universal cancer cure by helping researchers identify better targets, design more selective drugs, predict treatment response, improve imaging, and match therapies to specific tumor characteristics.

Rare genetic diseases may be another area where AI has a strong fit. Once a disease-causing mutation is identified, researchers still need to understand how that mutation changes a protein or cellular pathway and determine how the abnormality might be corrected. Protein prediction, molecular modeling, gene editing, RNA therapeutics, and automated drug design could potentially connect those steps much faster than traditional research programs have been able to.

The same principle extends to Alzheimer’s disease, metabolic disease, autoimmune conditions, infectious disease, and aging research. AI does not remove the need for biology. It increases the number of biological questions researchers can analyze and the speed at which they can decide which answers deserve experimental testing.

⸻

The Simple Explanation

The prediction that AI could compress 50 to 100 years of medical progress into 5 to 10 years is still a prediction, but there is real science behind the idea. AI has already changed protein structure prediction, researchers are using generative models to design new proteins, AI systems have identified unexpected drug candidates, and AI-assisted drug development has reached randomized human trials. The biggest opportunity comes from allowing researchers to analyze more biological information, test more useful ideas, and eliminate weak candidates earlier. The biggest limitation is that biology still has to happen in the real world. Drugs must be manufactured and tested, tissues still take time to change, and clinical outcomes still need to be measured in humans. AI can speed up the research process considerably without removing those physical limits.

Disclaimer: As always nothing in my breakdowns is meant to be medical or legal advice and is purely educational