To celebrate AI right now almost feels dangerous, because of the widespread and increasingly intense backlash against it. I myself have written about the dangers of AI, not least in these pages, so I understand the anti-AI sentiment. But we must be careful not to throw the baby out with the bathwater.
When my team and I wrote Transcend in 2024, we talked of two sides to AI—the dangerous side, yes, but also the side that had the potential to uplift humanity and create things of unimaginable benefit to all of us. And I have a deeply personal connection to the positive side, a connection that came about through my son, and one of the hardest things a parent could ever have to hear. In the midst of doom and gloom, it may provide a sliver of optimism and remind us that AI is truly a philosopher’s stone—capable of great damage in the wrong hands, and equally great benefit when used well.
Your son has cancer
In the fall of 2021, my son, then 18, started college—normally a fun and exciting time. But before the semester was over, he had been diagnosed with multiple myeloma, which is a cancer of the plasma cells in the bone marrow.
Along with many other things I would much rather never have had to learn, I discovered that multiple myeloma is a disease that typically affects older people. According to the American Cancer Society, the average age of people with the diagnosis is 69, and fewer than 1% of patients are younger than 35.
Because of the profile of the disease, much of what doctors know about multiple myeloma comes from older patients. Young myeloma patients are underrepresented in the research, often appearing only in small numbers within studies dominated by older patients. And some basic questions about very young patients—how their disease differs, how they respond to treatment, and what their long-term outcomes are—remain surprisingly poorly answered.
In other words, much of what anyone knew about my son’s disease had been learned from people decades older. The treatment protocols and survival estimates he was given were therefore drawn largely from populations he did not belong to.
Now, his care was excellent and I remain deeply grateful to the expert professionals who saved his life. But because he was such an edge case, his doctors had to make decisions about him using evidence drawn largely from other kinds of patients. There was no other way to treat him. Medicine has always had to make decisions about individuals using knowledge generated from groups.
But that is changing. AI is beginning to make possible something we have wanted for a very long time: treating the specific patient in front of us, rather than the average patient.
The limitation of standard medical treatment: the problem with averages
The way we find out whether a treatment works is to give it to one group of people and compare them with a control group. The trick that makes the comparison trustworthy is randomization: Assign patients to the two groups by chance, and every difference between individuals—in age, fitness, biology, the things you can measure and the things you can’t—is as likely to fall on one side as the other. Across enough patients, those differences cancel out, and the result the trial reports is the average outcome of one group against the average outcome of the other.
But as researchers have pointed out, the average effect reported by trial can conceal the distribution of effects within the trial (a given average can be a mix of “substantial benefits for some, little benefit for many, and harm for a few”). When you combine this with the fact that most medical treatments are designed for the average patient, it leads to a fact that no patient or medical professional is surprised by—even in the best case, treatments rarely work for everyone, and they work better for some patients than for others.
The cost of treating an individual as an average is not just that the average treatment may not help. It is also the unnecessary suffering of spending months on a treatment that the patient’s biology was never going to respond to, and the opportunity cost of not finding a treatment more suited to them. And all while their illness advances.
Treating the individual: AI and personalized healthcare
To treat someone as an individual rather than an average, you need three things. You need to know who they are biologically; you need a way to choose a treatment based on that knowledge rather than simply on what worked best for the general population; and if the perfect treatment for them does not currently exist, you need to be able to build it. AI can help with all three.
1. Understanding the individual patient
The first step toward more individualized medicine is getting more information from the patient in front of you. AI can help here, most obviously by finding patterns in scans and tissue that are difficult for humans to see, sometimes extracting information that was there all along.
This is already starting to happen. A Mayo Clinic model found signs of pancreatic cancer on scans up to three years before it was diagnosed, when curative treatment may still be possible. A pathology system published in Nature Cancer can recognize cancers across different organs and hospitals from only a handful of example slides, without being retrained for each task.
These results are not confined to research centers. The first randomized trial of AI-supported mammography, involving more than 100,000 women in Sweden, detected more cancers while cutting radiologists’ screen-reading workload by 44%. It also had 12% fewer cancers diagnosed between screening rounds, with a radiologist still reading every mammogram.
Of course, none of this is personalized treatment yet. But it is a necessary first step: seeing more clearly what is happening in this particular patient before deciding what to do about it.
2. Choosing a personalized treatment
In order to decide on treatment for a patient, doctors have to infer from the general to the specific—from the population-level insight, they have to figure out what might best suit the individual patient in front of them. And AI may allow us to make that inference much finer; ultimately, it opens up the tantalizing possibility of using a patient’s own data to decide which treatment is right for them.
In one myeloma study, researchers used machine learning to simulate alternative first-line treatments for 715 newly diagnosed patients and estimated that roughly one in six might have responded better to a different treatment. This was a simulation, not a clinical trial, but it shows the possibility: Use what we know about this patient to ask not simply what works best on average, but what is most likely to work for them.
Relatedly, in a recent conference study, researchers used AI to extract information about a patient’s immune system from an ordinary bone-marrow biopsy slide. They then asked whether those hidden biological differences could help predict which treatment would work best.
The differences were striking. In one group identified by the AI, 87% of patients given the more intensive treatment were still free of a major disease event after 18 months, compared with just 29% given the standard treatment. In another group, patients did about as well whether or not they received a stem cell transplant.
This is still early research, and the findings need to be tested prospectively. But the direction is important. Two patients can have the same diagnosis, even the same clinical stage and genetic risk, and still have very different immune biology. And rather than asking which strategy is best on average, AI lets us start asking which strategy fits the biology of the individual patient.
3. Bespoke medicines
The extreme end of personalization is a treatment that is created in response to a specific patient’s need.
This sounds like science fiction, but on August 19 Merck and Moderna announced that a Phase 3 trial of a vaccine called intismeran autogene had met its main endpoints in patients with melanoma. Moderna’s shares rose 177% in a day, which is one way of measuring the importance of the announcement.
Why is this so exciting? Well, it is a crucial milestone on the road to fully bespoke treatment. Merck describes the vaccine as designed for each patient from the unique set of mutations in their tumor. The tumor is sequenced, a machine-learning-based algorithm selects up to 34 targets, and those are encoded in mRNA, the technology behind the COVID-19 vaccines, which is essentially a written instruction that tells your own cells what protein to make.
Here, AI is not simply helping doctors see the patient more clearly or choose among existing treatments. It is helping identify what should go into a treatment built for that particular person. That is personalization in a much more literal sense.
AI does not mean the end of averages
Now, there is an important qualification to all of this: AI does not mean escaping population-level knowledge altogether. These systems still learn from large groups of patients and use what they learn to make better judgments about an individual.
Much of what we call personalization today is therefore better stratification: finding differences between patients that conventional averages hide, rather than being able to predict exactly what will happen to one person. And bespoke medicines take us a step further, because the treatment itself can be built for a single patient. But even there, the knowledge used to build it comes from what we have learned from many others.
So this is not the end of averages. Indeed, in a sense, AI is still averaging—machine learning is being used to find the subgroups the average hides. What these tools deliver today is therefore not 100% personalized care, but a better estimate of who is likely to respond rather than a precise forecast for one person. That is an enormous improvement on the blunt averages that modern medicine normally uses—and this is just the beginning of the journey.
The future
We got lucky. My son was the patient the average was not built for, but he was treated well because skilled people made the extrapolation work. He is now in remission and back at school.
He is a remarkable and inspiring human being and I am fortunate to have him. And I hope that at some point soon, when an 18-year-old walks into a hospital with a disease whose average patient is 69, the doctors are able to treat him rather than having to extrapolate from the population he is not a part of.
Actually, this is more than a hope. In all honesty, it is the foundational purpose of my life. That is why the proceeds from all my books are pledged to cancer research.
I want to help build a world where each child gets the care their particular biology calls for. And the exciting thing is that the tools and technologies needed to make this dream come true already exist. Let us do our best to use them wisely.
