What "Curing" Cancer Could Mean In The Age Of AI

Disclosure: The author holds no position, long or short, in any company mentioned. This article is for informational and educational purposes only and is not financial advice. Although the author is a physician, this content is commentary on the underlying science and does not constitute medical advice, diagnosis, or treatment recommendation. The author receives no compensation for this article and has no business relationship with the companies mentioned. See the full Legal Information and Disclosures section below.

Anthropic CEO Dario Amodei expects AI to cure most human disease, cancer included, within five to ten years. In his 2024 essay Machines of Loving Grace the claim was conditional on the arrival of what he calls "powerful AI." He repeated this in an X post in August 2026 and a September 2026 essay without that condition. That essay also says that AI has been advancing "drastically faster" since this summer. Together this reads as if he now thinks powerful AI has arrived or is close.

The 2024 essay lists the "speed of the outside world" among the limits on intelligence and grants that developing a cancer cure may have an irreducible minimum time. Even with powerful AI, "curing cancer" within a decade will be extremely hard. At the same time, cancers are already cured every day by surgery, systemic therapy and radiotherapy or combinations thereof.

Amodei calls reductions of 95 percent or more in both mortality and incidence possible, counts prevention and the interception of early cancer alongside treatment, and concedes that cancer is "extremely varied and adaptive."

Leukemias, lymphomas and myeloma arise from blood-forming and immune cells in the bone marrow and lymphatic system, and make up about nine percent of new cancers in the United States. Surgery is mostly limited to biopsies (e.g., lymph node excision in lymphoma), while treatment is based mainly on drugs: conventional chemotherapy, targeted therapy and, in some cases, a stem cell transplant. Childhood acute lymphoblastic leukemia reaches five-year survival of about 90 percent in children younger than 15 on conventional chemotherapy-based regimens. Machines of Loving Grace names some leukemias treated with CAR-T as subtypes already largely cured. In children, CAR-T is approved only when the disease is refractory or in second or later relapse, and in its registration trial it reached event-free survival of 44 percent at three years.

The other nine in ten cancers are solid tumors. A solid tumor that is still localized can be resected or irradiated, and this is how most solid tumors are cured. Chemotherapy or another systemic therapy is often added before or after surgery to kill the cells that have already left the tumor.

Among more than a million patients with a first cancer diagnosis in England, 52 percent were alive five years later. Of those survivors, 80 percent had undergone surgery, 39 percent had received radiotherapy and 29 percent chemotherapy. The Lancet Oncology Commission on global cancer surgery expected over 80 percent of newly diagnosed patients to need an operation.

Once a solid tumor has spread, systemic therapy on its own can cure only a few exceptionally responsive tumors. Testicular germ cell tumors are one example, with a five-year relative survival of 95 percent across all stages and 72 percent even after distant spread.

In most solid tumors success is measured in months of median survival and in the share of patients alive at a fixed point, usually five years, and a drug that turned a lethal cancer into a controlled one would count as a major success while still falling short of what a general reader hears in the word "cure."

In cancer registries a patient counts as cured once the excess risk of dying from the cancer has faded. A study of 818,902 Italian patients put that point at under ten years after diagnosis for stomach, colorectal and pancreatic cancer and at 19 years for breast cancer. So the time it takes to count as cured varies from cancer to cancer. Five-year relative survival for all cancers combined in the United States reached 70 percent for patients diagnosed between 2015 and 2021, up from about half in the mid-1970s. Surviving cancer is no longer the exception.

The word cancer covers more than 100 distinct diseases, an umbrella term for the many forms of neoplasia with vastly different behaviors that pathologists have been trying to classify for decades. For instance, basal cell carcinoma invades locally and metastasizes in less than one percent of cases, whereas pancreatic cancer still has a poor prognosis with a five-year relative survival rate of about 13 percent.

Pancreatic cancer is a good example of the pitfalls of cancer terminology. In clinical and scientific usage the term pancreatic cancer almost always refers to pancreatic ductal adenocarcinoma (PDAC), which accounts for roughly 90 percent of pancreatic malignancies, whereas neuroendocrine neoplasms of the pancreas follow a separate classification with their own prognosis and treatment. The often cited 13 percent five-year survival rate blends the two: in the same registry data PDAC reached 8 percent and neuroendocrine tumors 72 percent, and part of the rise in survival since the mid-1990s, from 4 to 13 percent, reflects incidentally detected well-differentiated neuroendocrine tumors.

Even inside PDAC the rare adenosquamous variant showed shorter survival after resection than conventional ductal adenocarcinoma. A single label therefore covers entities that differ in cell of origin, biology and outcome, and any statement about "curing pancreatic cancer" applies to at most one of them at a time.

Molecular pathology keeps dividing histopathological categories further, so that a tumor type once defined by its histology may now split into several molecular subtypes, some with their own standard of care. Sequencing every tumor genome to match it with a therapy might look like the natural next step. Yet an estimate published in 2021 put the share of US patients with advanced or metastatic cancer who responded to genome-targeted therapy at around seven percent in 2020, up from under three percent in 2006, with only about half of those eligible for a targeted drug responding to it.

The 2022 version of Douglas Hanahan's landmark paper "Hallmarks of Cancer" adds phenotypic plasticity, non-mutational epigenetic reprogramming, polymorphic microbiomes and senescent cells to a list that already covered cancer metabolism and the tumor microenvironment, and the 2026 update adds the normal cells recruited into a tumor and systemic factors such as aging and obesity as dimensions of their own.

The cancer genome, which can be sequenced and handed to AI, is only one layer of the disease. In PDAC, cancer cells often make up only 5 to 20 percent of a primary tumor, the rest being stroma, mostly cancer-associated fibroblasts and immune cells. Viewing the cancer cells and their genome in isolation would significantly understate the complexity of this disease.

"Solving" the cancer genome will therefore probably not mean solving cancer. I often think of the cancer genome as a book left on the surface of Mars. Without somebody able to read it, the book is no more than a lump of carbon (and if a reader arrives, its meaning might depend on that reader's interpretation). Thus we also have to "solve" the reader, by which I mean the cancer cell (its epigenetics, proteome, metabolism etc.), the network of different cancer cells and how they interact (e.g. cancer stem cells), and the tumor microenvironment as well as other factors, for example the tumor microbiome or even factors we do not know yet. A mutated genome might initiate the disease but is only one piece of the puzzle in a clinically apparent tumor.

A tumor is also a population of cells under evolutionary selection, and any new treatment adds a new pressure. In the TRACERx study of non-small cell lung cancer, whole-exome sequencing of 1,644 tumor regions from 421 patients found mutations in 22 of 40 common cancer genes under significant subclonal selection in treatment-naive lung adenocarcinoma, so the drivers may already differ between regions of the same tumor at diagnosis. A drug that eliminates the dominant clone could then leave the others with a selective advantage, and the recurrent tumor may carry different drivers than the one the drug was chosen against.

Cell lines grow in petri dishes without stroma, immune cells or vasculature, xenografts grow in immunodeficient mice whose drug metabolism is not human, and a computational simulation is built on the parameters someone has measured and chosen to include, so the tumor in a real patient differs from every model of it. Each model can be used to analyze or simulate a mechanism but cannot verify whether a treatment works in a human. We still need clinical trials to determine efficacy. In a database of trials run between 2000 and 2015, oncology programs went from phase 1 to approval about 3.4 percent of the time, while one can assume that every candidate that reached phase 1 looked very promising in the lab.

In late August 2026 the FDA approved daraxonrasib, an oral RAS inhibitor, after it nearly doubled median overall survival in previously treated metastatic PDAC: 13.2 months against 6.7 on physician's choice chemotherapy, with a response rate of 30 percent against 11. That marks an inflection point for the disease. Yet seven in ten patients still have no objective response, and median progression-free survival is about seven months, even though oncogenic RAS drives more than 90 percent of PDAC and the drug inhibits both mutant and wild-type RAS. Even this historic breakthrough is still far from a "cure."

Among 44 patients who had benefited from daraxonrasib for more than three months in the earlier phase 1/2 trial and then progressed, circulating tumor DNA showed acquired oncogenic alterations in 59 percent, most often extra copies of the mutant KRAS allele. To get past that resistance, the authors combined daraxonrasib with zoldonrasib, a KRAS G12D inhibitor, or with antibodies against growth factor receptors, and tested the combinations in cell lines and mice. The daraxonrasib plus zoldonrasib doublet has so far produced phase 1 results in 60 previously treated patients, and a phase 3 trial against chemotherapy in first-line KRAS G12D disease has started, so randomized evidence is still years away. Every further combination would need trials of its own, with years of follow-up.

With mutation-specific drugs such as zoldonrasib, molecular subtyping could change how PDAC is treated, and in a few years oncologists might speak of G12D or G12V PDAC the way they now speak of HER2-positive breast cancer.

Pancreatic cancer will be diagnosed in an estimated 67,530 people in the United States in 2026, roughly 90 percent of them, about 60,000, with PDAC. According to a real-world sequencing database more than 90 percent of them, or roughly 56,000 a year, carry an oncogenic RAS mutation, which further divides into KRAS G12D in about 40 percent of PDAC, G12V in 29 and G12R in around 15 percent, about 24,000, 17,000 and 9,000 patients a year. Daraxonrasib is designed to work across RAS variants, so its trials could recruit from all of these patients. Zoldonrasib can draw only on the G12D share, and a drug candidate targeting G12V or G12R on an even smaller pool. Such trials would probably take longer to enroll and yield fewer outcomes. The more specific the target and the more personalized the therapy, the harder it becomes to generate randomized evidence.

The supply of public, human-written text available to train the largest language models runs to hundreds of trillions of tokens. Patient outcomes, the training data for a model of what a drug does in patients, cannot be stockpiled in the same way, because a clinical endpoint exists only after a patient has been treated and followed over time.

There may simply not be enough data for AI to learn from: the registrational trial of daraxonrasib in metastatic PDAC randomized 500 patients worldwide, and the resistance mechanisms were analyzed in 44. That is enough to show that the drug extends survival in the trial population as a whole. It is far too little to learn what it does in each subgroup or even individual patients, which is what an AI matching patients to therapies would need.

Among patients relapsing on daraxonrasib, acquired KRAS amplification appeared in 15 of 32 with a pretreatment TP53 mutation and in 1 of 12 without, which leaves an even smaller group in which any one effect can be measured.

Dario Amodei might answer that AI will generate its own data by directing experiments, which it might well do in the lab. However, patient outcomes in clinical trials decide whether a drug works, and the time needed to follow those patients is hard to compress. The question whether one inhibitor extends survival in patients with one specific KRAS mutation can only be answered reliably with hundreds of outcomes, one per patient.

A novel cancer drug candidate is usually first administered to patients with advanced or metastatic disease in the second line or later who have run out of standard options. This kind of research on patients is only allowed and ethically justifiable when the importance of its objective outweighs the risks and burdens and the patient wants to take part and gives informed consent. Since the toxicity and the needed dose are not clear yet, the drug is first tested in a few patients.

Besides these clear ethical limitations there is also an economic one: pivotal trials cost tens to hundreds of thousands of dollars per patient, so larger trials can usually only be funded after smaller ones have shown an effect, and a drug usually moves into earlier lines only after it has beaten the standard of care.

Daraxonrasib went this way: 38 second-line patients in the phase 1/2, 500 in the phase 3, approval for previously treated patients and for those who are not candidates for multiagent systemic therapy, and dosing in a first-line phase 3 trial only since April 2026.

Amodei's essay links an analysis of AI-discovered molecules as a sign that AI may reduce the need for iteration in trials. In that analysis, phase 1 success ran at 80 to 90 percent, well above historical averages, while phase 2, the first step that tests efficacy in patients, succeeded around 40 percent of the time on a small sample, almost in line with the historical, pre-AI rate.

The adjuvant trial of daraxonrasib, RASolute 304, has disease-free survival as its primary endpoint, which takes years of follow-up to measure. Amodei's essay concedes that clinical trials often take years but argues that a drug which works really well could move much faster. Daraxonrasib tested that hypothesis under favorable conditions, with an "unprecedented" hazard ratio of 0.40 for death, breakthrough designation and an approval six and a half months ahead of the goal date. It still took about four years from the first patient dosed in 2022 to approval in August 2026, for one indication in one type of cancer. AI might trim months from enrollment and analysis, but cannot shorten the time a tumor takes to recur (or the needed time to observe non-recurrence) or the follow-up needed before a survival curve separates and stays separated.

One could argue that regulators might accept earlier surrogates for cancer recurrence, which might shorten the wait further, but the candidates are weaker than they look. The standard PDAC marker CA19-9 is not a validated surrogate endpoint and is scarce or absent in the 5 to 10 percent of patients who are Lewis antigen negative. In a 2026 series of resected PDAC, postoperative circulating tumor DNA flagged later relapse with a sensitivity of 36 percent. The FDA's 2024 guidance calls ctDNA not validated as an early endpoint and still recommends disease-free, event-free or overall survival as the primary endpoint.

So "curing" cancer in the age of AI could mean two different things. One is the sense a general reader hears, a drug that clears (metastatic) disease. The other is more patients reaching the cures that already exist: a tumor found while it can still be removed, and treatment at the standard of the best centers, from the operation to the drugs around it.

If most solid tumors are cured by removing or irradiating them before they spread, the nearest lever for AI may be to find them in time. A deep learning model trained on non-contrast CT scans, the kind ordered every day for other reasons, detected pancreatic lesions with a sensitivity of 92.9 percent and a specificity of 99.9 percent in 20,530 consecutive patients. Whether finding tumors this way saves lives is another question. The US Preventive Services Task Force still recommends against screening asymptomatic adults for pancreatic cancer: no trial has shown a benefit, and the harms of false positives and of treating screen-detected disease are rated at least moderate.

Significantly reducing mortality through screening would be a monumental achievement even without curing clinically apparent tumors, but proving such interventions work requires even more patience. For example, the NELSON lung screening trial randomized participants between 2003 and 2006, yet only reported its ten-year mortality results in 2020.

Where a screening program already exists and its mortality benefit is established, the question shrinks to whether the AI finds more of the cancers that matter, and that can be answered faster. The MASAI trial randomized 105,934 women in Sweden's mammography program between 2021 and 2022. AI-supported reading found 29 percent more cancers at a 44 percent lower reading workload. In January 2026 the trial reported its primary endpoint, a rate of interval cancers that was non-inferior, though not shown to be lower. All of this took under five years, against 16 for NELSON.

A tumor found in time still has to be resected to the right oncological extent, and that differs by surgeon and by hospital. In rectal cancer, pathologists grade each specimen by the plane in which the surgeon dissected. In the CR07 trial, only 52 percent of 1,156 specimens had been removed in the intended mesorectal plane, and three-year local recurrence was 4 percent after those operations against 13 percent when the dissection had cut down to the muscle wall of the rectum. In the German CAO/ARO/AIO-04 trial a few years later, 81 percent of specimens were in the mesorectal plane.

In complex operations like pancreatic resections, outcomes differ between hospitals with the number of such operations they perform. Across 60,858 major pancreatic resections in Germany from 2009 to 2014, risk-adjusted in-hospital mortality was 6.5 percent in the highest-volume hospitals and 11.5 percent in the lowest. The German minimum for offering these operations was raised to 20 a year only in 2025.

In the operating room itself, a neural network reading stimulated Raman histology of fresh brain tumor tissue returned a diagnosis in under 150 seconds. Its accuracy was 94.6 percent, against 93.9 percent for pathologists reading conventional slides, in a prospective trial of 278 patients. Models trained on annotated videos of gallbladder removal can mark safe and dangerous zones of dissection, and a cluster-randomized trial of that guidance on the live image is recruiting. A robot sutured two ends of intestine in pigs laparoscopically with minimal human intervention, and AI might one day conduct operations itself.

AI is entering radiotherapy too, the treatment that 39 percent of the five-year survivors in the England data had received. Automated contouring and planning has been tested in more than 1,000 patients with cervical, head and neck and prostate cancer at six public cancer centers in low- and middle-income countries, where planning can take weeks and the software just over an hour. Results reported in 2026 found that the software planned radiotherapy to a high standard in more than 95 percent of cervical and 85 percent of prostate cancer cases.

Within a decade, tools like these could democratize the experience and technical know-how of high-volume centers. Each would still have to show in a trial that it changes complications or survival.

Systemic therapy around the surgical resection of a tumor can also improve outcome. Adjuvant therapy follows the resection and targets the cancer cells that have already left the tumor. Neoadjuvant therapy is administered before surgery, to shrink the tumor, reach those cells earlier and show whether the tumor responds. Perioperative therapy is intended to do both. In PDAC, for example, adjuvant mFOLFIRINOX raised five-year survival from 31 to 43 percent in PRODIGE 24.

RASolute 304 follows the same route for RAS inhibition in PDAC. If the drug that nearly doubled median survival in metastatic disease prevents recurrence after resection, the cure rate of pancreatic cancer could rise significantly. AI could help choose the drug and the patient for such trials and read their data faster, but it cannot shorten the years of follow-up they need.

AI may well advance oncology through better target selection, better molecule design, better patient stratification and faster analysis of the data that trials generate. Daraxonrasib itself is a product of the structure-guided medicinal chemistry that AI could speed up. But "curing cancer" in a decade, in the sense a general reader hears, requires fitting discovery, sequential clinical trials, regulatory review and long-term durability data into a single development cycle. It requires doing this simultaneously for over a hundred distinct diseases against tumors that evolve under every effective treatment. In the sense in which solid tumors are cured today, AI could help find more tumors while they can still be resected or irradiated and treat them as well as the best centers do, from the operation to choosing the right drugs based on molecular diagnostics.

Follow me on X for frequent updates (@chaotropy).

Legal Information and Disclosures

General Disclaimer & No Financial Advice: The author is not a registered investment advisor or financial analyst, and nothing here is financial, legal, tax, or investment advice, a personal recommendation, an assessment of suitability for any investor, or an offer, solicitation, or recommendation regarding any security. This is not a securities research report and was not prepared under the requirements designed to promote the independence of investment research. It has not been peer-reviewed or audited, represents the author's opinion as of the publication date, and may change without notice. Readers should conduct their own due diligence and consult a certified financial professional before making any investment decision. Readers act on this content at their own risk, and to the fullest extent permitted by applicable law the author accepts no liability for any loss or damage arising from reliance on it.

Medical Disclaimer: The author is a physician, but this article discusses clinical trials, drug mechanisms, surgery, radiotherapy and cancer screening as general scientific and educational commentary. It is not medical advice, diagnosis, or a treatment recommendation, and it establishes no physician-patient relationship. Readers should never disregard or delay professional medical advice because of anything written here, and should consult a qualified healthcare provider about any medical condition.

Accuracy and Third-Party Data: Data here come from government and WHO sources, clinical trial registries, peer-reviewed journals, conference abstracts, company communications, and other public sources, believed reliable but not guaranteed complete, current, or correct. Some figures come from analyses by the companies developing the drugs discussed or from company announcements rather than peer-reviewed publications and have not been independently verified. Despite the author's efforts to verify the sources cited, this article may contain errors of fact or interpretation.

Disclosure of Interest: The author holds no position, long or short, in any company mentioned. The author reserves the right to buy or sell securities of any publicly traded company mentioned at any time without further notice. He received no compensation for this article, maintains no business relationship with any company mentioned, and has no access to non-public trial data, unblinded interim results, or internal company information.

Forward-Looking Statements & Risk: Statements here about clinical results, regulatory outcomes, and the future of AI in drug development, cancer screening, surgery and radiotherapy are predictions based on current expectations, subject to significant risks and uncertainties, and actual results may differ materially. Biotechnology and pharmaceutical investing carries a high degree of risk, including total loss of principal, and past performance does not indicate future results.

Copyright: Original text and images are the property of the author and may not be copied, reproduced, or published in whole or in part without prior written consent, except as applicable law or the publishing platform permits. Use of this content to train machine learning or AI models requires explicit authorization. Third-party and public domain images remain subject to their respective rights and are not claimed as the author's property.