Her name was June. June would be the last month she ever knew.

I remember the way her bed made her so small, swallowed her up in the middle of the room, turned as it was to face the window. With her liver failing, her skin was a ghostly apricot. June – I called her Nan – had been stalked by the king of terrors, the emperor of all maladies: cancer. That son of a bitch.

The cancer was in her pancreas. Odds of survival? “Not good” – putting it lightly. The maths, putting it darkly: it’s the 12th most common cancer worldwide, the sixth most common cause of cancer death. The World Health Organization says it is “one of the cancer types with the least favourable prognosis”. It was not favourable for my grandmother.

June died nine years ago. Lately I’ve been thinking about alternate futures – the ones in which my grandmother, and the many millions diagnosed with cancer each year, are still here. The thoughts creep in as we are bombarded by artificial intelligence chief executives making bold proclamations such as “AI will cure cancer in five to 10 years” or “AI will cure cancer in our lifetimes”. Those claims stand in contrast to apocalyptic scenarios, including AI being responsible for human extinction before 2030. AI, we hear, is both saviour and annihilator.

Park the world-ending scenarios and you’ll find there has been significant progress in using AI as a tool in the detection, prevention and treatment of cancer. If there is a future where a supposed AI apocalypse is averted, then it may be one in which cancer therapy has been completely redefined. But just how far can AI take us towards a cure?

Today AI is being turned toward drug development, treatment plans and, in some places, it’s being used clinically, studying imagery of lungs for early signs of malignancy and helping to remove brain tumours. Words like “efficiency” and “personalised care” are everywhere; the notion AI can free up time for doctors and medical professionals to get away from the busywork abound.

Cancer is not a single disease

Even with such positive advances, speak to clinicians, oncologists and researchers working at the coalface and you quickly find there is tension between the more grandiose claims of big tech and its race to “solve” disease, and the deliberate, considered progress of science.

“Until AI can actually think and do experiments itself at scale, I’m not sure that we’re going to find a cure for all cancer or medical illness in the next 10 years,” says Sherene Loi, a medical oncologist who specialises in breast cancer treatment at the Peter MacCallum Cancer Centre in Melbourne.

But Loi and other cancer experts the Guardian has spoken to believe there will be efficiencies for treatment, diagnostics and pathology services that will bear fruit, heralding more personalised and more equitable cancer care.

Cancer is not a single disease. It’s a catch-all term that describes the uncontrollable growth of abnormal cells. Typically, we describe cancers based on the location within the body that they are found: breast, prostate, lung, brain, blood, pancreas. These are complex, highly variable diseases, affecting a variety of different cells and underpinned by a host of genetic changes. In addition, cancer outcomes depend on many factors, including the biology of the disease and the treatments available. A single cure for cancer is not only unlikely but nigh on impossible.

Just take the brain. Though not all are cancerous, “there are more than 100 types of tumours in the brain, much more than in any other single organ”, says Felix Sahm, a neuropathologist at the University Hospital Heidelberg. “There are so many unresolved questions in brain tumour research and brain tumour care.”

We’ve made great leaps in cancer survival rates over the past century; chemotherapy and radiation have enabled solid tumours to be cured, monoclonal antibodies ushered in more direct targeting of abnormal cells, and emerging treatments are rewriting how we combat the disease. But, as Loi notes, “Human disease is very complex.” So where does AI fit in?

A shadow on a scan

“AI can improve several parts of that process by helping us identify risk, detect disease, characterise tumours, select treatments and monitor,” says Ajit Goenka, a radiologist and nuclear medicine specialist at the Mayo Clinic in Minnesota.

Medical imaging, in particular, stands to benefit from advances in AI and machine-learning models. Training models on huge amounts of patient data, such as CT scans, X-rays or tissue samples, enables those models to detect changes or patterns that may be imperceptible to the human eye.

Goenka has been exploring the early detection of pancreatic cancer with machine-learning tools since 2021, before the release of ChatGPT fully turned the public’s head toward generative AI. Since then, OpenAI and its rival Anthropic have pushed into new realms beyond chatbots: partnering with companies to accelerate access to treatment; striking deals with research centres; and working with pharmaceutical companies to incorporate large language models and AI into drug discovery.

Goenka noticed that some patients who developed pancreatic cancer had undergone CT scans months or years before diagnosis. Those scans were interpreted as normal but Goenka reasoned there could be microscopic, measurable changes invisible to a radiologist. “We began studying whether quantitative imaging and machine-learning methods could detect that signal,” he says.

In a proof-of-concept paper published in 2022, he confirmed that hunch. Machine-learning tools can detect pancreatic tumours, including in pancreases that appeared normal, before clinical diagnosis. Further studies by Goenka’s group have explored how robust and reproducible the method is and whether it is useful across institutions. The opportunity, he says, is to identify evidence of cancer earlier. My grandmother’s pancreatic cancer appeared as if out of nowhere, a shadow on a scan. Goenka’s models may have been able to see it sooner, which could have improved her outcome.

Reducing the duration of uncertainty

Sahm leads a European multi-disciplinary collaboration called EUcanAI, which is using agentic AI to improve the way brain tumours and central nervous system cancers are treated. The idea is that “every single chunk of this journey gets informed by the previous one and informs the next one in the sequence, so that it gets more efficient, quicker and less cumbersome for a patient”, he says.

He provides one example of a teenager who presented with a tumour in his brain stem. The expectation was that it was fatal, based on imaging and the patient’s age. The patient went in for surgery, his tumour was biopsied and rapidly sequenced to determine the mutations driving his disease.

Traditionally, these results go to a board of specialists to discuss the next steps in surgery and treatment. But today, these results can be delivered during a surgery, saving patients time and money. Sahm imagines a world where agentic AI accelerates this process even further – a surgeon could biopsy and image a tumour, then have an AI provide answers about the kind of mutation and potential drugs that may treat it. Sahm says this work is already happening in trials but stresses it is not yet routine.

The key for Sahm is a more tailored journey for each patient that “reduces the duration of uncertainty”. It’s about providing a wholly optimised approach to managing a patient’s cancer journey. He also notes it could provide more patients with better care.

This was a common thread in discussing AI tools for cancer diagnostics. “I think it will help with regards to equity of care, and access to care,” Loi says.

For instance, she says, in rural communities or locations that do not have cancer facilities or access to specialists or specialist diagnostic tools, access to machine-learning models could help medical professionals provide faster, more complete answers.

Can AI ‘solve’ disease?

The dream of curing cancer is one that big AI firms constantly seed. Dario Amodei, the head of Anthropic, has suggested that within a decade, cancer – and many other diseases – could be cured. The Google DeepMind spin-off, IsoLabs, has a bold mission statement to “solve” all disease.

Sam Altman, the head of OpenAI, has similar lofty goals, suggesting the ability to solve disease might just be a problem of how much computing power backs our AI systems. “Maybe with 10 gigawatts of compute, AI can figure out how to cure cancer,” he wrote. As AI firms experience growing backlash to the environmental cost of datacentres, the argument that just a little more compute could offer life-saving opportunities could be hard to sell.

But Emilia Javorsky, a physician and scientist at the Future of Life Institute, argues the opposite: “We cannot compute the fundamental truth of biology, we can only measure it.”

In her March essay on AI and cancer, she outlines the complicated process of developing and testing any new medication. “Thirteen years into the AI drug discovery movement, we still lack a single [Federal Drug Authority]-approved drug that cleared the full bar of regulatory approval, reimbursement, and clinical adoption,” she says.

Javorsky argues that curing cancer is a noble goal but the research community must be smart about where investment goes. It’s not intelligence that limits progress but the data, regulation and incentives that surround cancer research.

Even if all the data from research across the world was placed into these models, would that be enough to solve these diseases today? What about in five or 10 years? For Sahm, there are biological truths we are yet to uncover. In his work, he still encounters rare, unknown pathology that doesn’t “fit into a box”. While AI may be able to devise technology or tools to understand rare subtypes, it’s only as good as its training data.

The complexity of human physiology and biological processes comes up a lot in discussions with those working in cancer. Although AI may speed up the process of, for instance, hypothesising and creating new drugs that can treat specific forms of cancer, it is fundamentally limited in applying those learnings to the real world. “Someone actually has to go back and do the experiment, right?” Loi says.

An article published in The Australian newspaper in March revealed that a Sydney data scientist had been able to synthesise a bespoke cancer vaccine for his dog, Rosie, by uploading her genetic information to ChatGPT. And while some were astounded by the development, others noted that the hard work had still been done by humans in a lab.

Even if AI can determine the most effective targets for new drugs, it cannot cut through red tape. Sahm notes that the majority of AI-assisted or AI-led research being performed is experimental and remains in clinical trials. Even if AI tools were deemed safe for test patients, he says, the EU lacks a clear regulatory path to allow them to reach the market. “That ends up in a lot of AI research going on, a lot of surprising and promising things happening … but also surprisingly few of those really being applied,” he says.

The most vocal proponents of AI – those building the models – may make grand claims about cures but on the ground the reality is different. There’s a lack of understanding of cancer complexity, many researchers working in the space argue.

But it’s clear that AI does provide a sense of hope for clinicians and researchers. Alone it cannot dethrone the king of terrors; it cannot bring down the emperor of all maladies. Curing cancer will require more than extra-efficient computation.

“AI can come up with a million hypotheses,” Loi says – but we will still need humans to help understand which of them are true.

“Technology,” says Goenka, “is only one part of the answer.”