The AI Cancer Cure Debate: Promise, Skepticism, and the Broader Stakes of Superhuman Intelligence
核心洞察
Anthropic (搜索) CEO Dario Amodei predicts superhuman AI could compress a century of scientific progress into a decade and reduce cancer mortality by 95%, though most AI experts surveyed expect slower progress.
Cancer biology presents unique barriers to AI-driven breakthroughs, including finite data from clinical trials that cannot run at silicon speeds and ethical constraints on patient experimentation.
Current AI tools like AlphaFold (搜索) and medical imaging algorithms remain underutilized despite Nobel Prize-winning accuracy, suggesting formidable intelligence already exists but has not yet revolutionized drug development.
The question of whether artificial intelligence can cure cancer has moved from speculative hope to urgent public debate, fueled by bold predictions from AI industry leaders and tempered by skepticism from researchers who understand both the technology's potential and its limitations.
In a 2024 essay titled "Machines of Loving Grace," Anthropic (搜索) CEO Dario Amodei predicted that superhuman AI—smarter than Nobel Prize winners, freely using computers, and collaborating with millions of copies of itself—could compress a century of scientific progress into a single decade and potentially reduce cancer mortality by 95 percent. For someone like Dr. Jerusalem, an AI professor who carries a genetic mutation conferring very high risk of breast, ovarian, and other cancers, such a prediction should sound like salvation. By age 40, the breast cancer (搜索) risk for carriers of her mutation rises to one in four, double the lifetime risk for the average woman. Her mother, who also carries the mutation, was diagnosed with breast cancer at 45.
Yet Dr. Jerusalem finds herself "rooting for delays in the creation of this AI—hoping, in my heart of hearts, that GPT-6 will be a disappointment."
The Data Problem in Cancer Biology
The skepticism is grounded in the fundamental differences between domains where AI has excelled and the realities of cancer research. AI systems are strongest in settings such as chess, where they can generate infinite data by playing repeatedly, experiment freely, and observe exactly what happens. Mathematics and coding share these properties, and AI has yielded remarkable progress there.
"But cancer is different," Dr. Jerusalem writes. "Cancer data are finite and come from biological experiments and clinical trials that cannot run at silicon speeds. Experimenting freely on cancer patients would be unethical. And cancer data only imperfectly illuminate the complex processes by which our own cells betray us."
This view is shared by most AI experts in a survey Dr. Jerusalem recently advised, who generally expect slower progress than the leaders of AI labs. Kelsey Piper, a staff writer who participated in a live debate on the topic, articulated a similar position: LLMs are good at verifiable tasks, and cancer biology is full of unverifiable ones. She expects continued progress in curing cancer at roughly the rate of the last 40 years.
Underutilized Intelligence
The intelligence provided by existing AI systems is already formidable and, by many accounts, underused. AlphaFold (搜索), which won a Nobel Prize for predicting protein structures with stunning accuracy, has not yet yielded revolutions in drug development. AI algorithms that match or beat radiologists at many types of image analysis remain incompletely integrated into clinical practice. Chatbots now aid scientists with research, but their full potential is unrealized.
Dr. Jerusalem describes how her Ph.D. students used to write code to analyze medical data; now they express their ideas in plain English and let AI do the rest. "They operate essentially as professors, constrained only by their own imagination," she notes. One of her students recently came to her "giddy with excitement over an AI-aided medical discovery."
The Fable 5 Episode and Institutional Unreadiness
The recent chaotic release of Anthropic (搜索)'s latest model, Fable 5, illustrates how unprepared institutions are to handle the broader repercussions of increasingly powerful AI systems. Anthropic, fearing the model might be misused to develop bioweapons, initially kneecapped its ability to answer most basic biological questions—a measure the company described as temporary. This made the model, ironically, far less useful for cancer research than its less powerful predecessors.
Days later, the U.S. government issued a national-security directive prohibiting foreign nationals from using the model, likely due to concerns about cyberattacks. In response, Anthropic (搜索) shut the model down entirely. Anthropic did not respond to a request for comment about Fable 5's rollout.
"Reasonable people disagree about how risky this model is and whether Anthropic (搜索) or the government is overreacting," Dr. Jerusalem writes. "But clearly, our institutions aren't remotely ready to respond to these rapid deployments."
Speed Versus Caution
Many developers of these models, including Amodei, agree that AI is progressing more quickly than society is adapting. Their proposed solution is for society to speed up, not for AI to slow down—a position Amodei frames in his essay "Policy on the AI Exponential," which treats AI progress as "an iron arc to which society must bend."
But accelerating ahead will inevitably mean more chaos of the type that surrounded Fable 5's release. More fundamentally, it will shorten the time available to respond to societal challenges that powerful AI may raise: mass unemployment, skyrocketing inequality, repressive surveillance, and autonomous warfare. Each of these, Dr. Jerusalem argues, is "an enormous problem, no less obviously important than curing cancer, for which we lack good solutions."
The debate over AI and cancer also reveals divergent definitions of what "curing cancer with AI" even means. When Google's Demis Hassabis says AlphaFold (搜索) will one day cure cancer, he refers to AI tools that can speed up protein prediction and drug discovery—gradual, tool-assisted progress rather than a magic pill. OpenAI's Greg Brockman shared a story about a man who treated his dog's cancer with a personalized drug developed with assistance from ChatGPT, where the AI helped the owner find the right people, machines, and techniques. And when Amodei talks about curing cancer, he relies on recursive self-improvement: a model that trains a better model until something so capable emerges that it solves everything.
The Question of Meaning
Beyond the technical and institutional challenges lies a deeper concern about what is lost when human intelligence is obviated. Amodei grapples with this question in his essays, calling it "more difficult than the others." He suggests humans will still find meaning in deep intellectual pursuits, even if AI can do them better, and points to activities like playing video games, swimming, and talking to friends.
Dr. Jerusalem finds this answer unconvincing. "I would neither spend months struggling with a research problem I knew AI could solve instantly, nor find as much pleasure in the answers it provided," she writes. "I do not want to be merely a spectator to the universe, whatever wonders AI may reveal."
For her, writing itself is a process bound up in self-discovery and human connection—the idea suggested by her sister, the comfort offered by her wife's touch, the late-night writing at a handmade dining-room table inherited from grandparents. These are dimensions of meaning that no AI, however intelligent, can replicate.
"I will wait a little longer for a cure—even if it means losing my fertility and living under the shadow of risk—if it lets us approach this new world more carefully," Dr. Jerusalem concludes, "and ensure that, in curing cancer, we do not lose the things that make cancer worth curing."