“First solve intelligence, then use intelligence to solve everything else.” That was how Demis Hassabis summed up the vision of DeepMind (now Google DeepMind) when he founded the company in 2010. He gave his company 20 years to get there.
This ethos – the belief that intelligence derived from AI (given enough compute and data) will save us all – has been so rarely questioned that by the time Hassabis went on 60 Minutes last year and stated “I think one day we can cure all disease with the help of AI,” he was only repeating claims that other prominent AI CEOs had already made.
Including curing cancer. Especially curing cancer.
Last year, OpenAI CEO Sam Altman testified before the Senate Commerce Committee and said, “ We are working to build tools that one day can help us make new discoveries and address some of humanity’s biggest challenges, like curing cancer.” At Davos, Dario Amodei of Anthropic said a superintelligent AI would grant humans “wonderful things, like the ones I talked about in Machines of Loving Grace. It will help us cure cancer.”
Unfortunately, their vision of a magic genie both fundamentally misinterprets and undermines cancer science, says Dr. Emilia Javorsky, a physician, public health researcher, and director of the Futures program at the Future of Life Institute. She has worked across scientific research, clinical trials, tech startups, and AI policy, and she recently wrote a paper called How AI Can and Can’t Cure Cancer, in which she argues that the promise of superintelligent AI curing cancer falls apart under scrutiny.
Javorsky is not anti-AI. Just the opposite. As she told Tristan Harris in a recent conversation on Your Undivided Attention, there are tons of astonishing ways AI advances the fight against cancer every day – from breakthroughs in breast cancer imaging to reading and analyzing genetic data at scale. Meanwhile, she says we are actually losing ground in the fight against cancer by diverting resources away from research and technology that could actually make a difference.
The point Javorsky kept coming back to in the interview is that there are two types of AI: the first are general-purpose models like Claude, Gemini, and ChatGPT, which the big AI companies are pushing to dominate the market in the hope that they’ll unlock superintelligence. The second is narrow AI: bespoke models created for a specific problem, like detecting cancer from blood tests or helping an agricultural company with weeding. Right now, the lion’s share of resources is flowing into the former, not the latter. Javorsky argues that this is entirely the wrong approach. “It’s absurd to me, the situation that we’re in,” she says. “There’s so much we could be doing that we are not actually doing, and we’re doing all of the wrong things.”
Is this content resonating with you? Every share helps amplify it.
Scientists have made massive strides in medical knowledge over the past 75 years. We aren’t lacking in breakthroughs or brainpower: today, we have an oversupply of human scientists relative to the resources we can actually allocate to experimentation.
But despite that acceleration and knowledge, “we’ve noticed that therapeutics approved to actually help people have remained markedly flat,” says Javorsky. “We actually haven’t made commensurate progress, so the intelligence that we’ve gained hasn’t really been coupled to actually moving the needle on saving people’s lives.”
Meanwhile, cancer rates have only climbed worldwide, with new diagnoses more than doubling since 1990. And the survival rate is almost exactly the same as it was over a decade ago.
So knowledge itself is not the missing piece, and neither is AI. In fact, we now have AI to thank for leaps forward in diagnosing hard-to-spot breast cancers, helping surgeons excise tumors by identifying tumor margins in real time, and predicting whether a novel treatment would be toxic or non-toxic.
But all of these examples rely on human-supplied data – images and extensive libraries of existing compounds – which AI models have been able to learn from. Take AlphaFold, the poster child for success with AI and biology. That breakthrough in solving protein folding would not have been possible without the global Protein Data Bank, an archive of 3D protein structure data to which scientists had been contributing for decades.
Most people aren’t aware that the world has no equivalent data bank for cancer, which is what a superintelligent AI – if that ever exists – could train on to develop cures.
“We don’t even have a national sort of data commons of cancer genetics and imaging data, the things that scientists could learn from, that’s interoperable,” says Javorsky. When doctors collect specimens in a clinic, they are usually not shared beyond their health system.
Cancer is complex, and treating it is highly individualized Compared to things like treating the flu or treating high blood pressure, which are more static biological processes, cancer “is something that is co-evolving with us. It’s dynamic, it’s complex, and it’s highly individualized,” says Javorsky. We’re a long way from our former simplistic understanding of cancer as something that originates in a cell that gets a mutation, goes rogue, and makes a tumor.
Cancer is now understood as a much more complex disease involving the immune system and the blood supply. And each person’s case is unique. Javorsky points out that science has yet to cure any complex, chronic disease, like diabetes and Alzheimer’s. The best modalities for treating cancer center on the individual, and that will still be true in the age of superintelligent AI.
Curing cancer has other inherent bottlenecks If cancers were as easy to treat in people as they are in mice, we would be well on our way to curing cancer. Unfortunately, more than 90% of cancer cures that look promising in mice fail when they are tested on humans. Simulating a human’s unique biology, which has no clean ruleset, won’t be possible with AI (no matter how powerful it is) because biological truth cannot be computed, says Javorsky.
Diseases also require that we take time – human time – to test out whether a cure or treatment is working. And that delay is compounded by the FDA’s clinical trials process, a notorious bottleneck that few companies can even afford to attempt.
Finally, there will always be an issue with scaling human testing at the breakneck pace needed to test all the potential therapies AI can develop. “On the AI side, we’re rapidly flooding the system with new molecules that we want to test. But we can’t scale people. We can’t scale the number of patients in a clinical trial. We can’t scale the number of tumor specimens that come from a patient to test,” says Javorsky.
“On the AI side, we’re rapidly flooding the system with new molecules that we want to test. But we can’t scale people. We can’t scale the number of patients in a clinical trial. We can’t scale the number of tumor specimens that come from a patient to test.” — Dr. Emilia Javorsky
In the race to save lives, we need the right tools, and AI will be one of them. But it needs to be purpose-built AI applied to initiatives that will save lives today, such as reducing the cost of manufacturing a new drug or developing new medicines and tools.
“Fundamentally, I’m super bullish on the promise of AI in oncology and medicine in general,” Javorsky says. “It’s just the right kind of AI development that’s targeted to actually solving the problems and unblocking the things that are holding up our ability to move science forward.”
See the full conversation with Dr. Javorsky here:
We’re working to empower policymakers, technologists, and everyday people to guide technology toward the public good. If you value our work and want to support it, consider donating.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.