Table of Contents
When people ask me how to separate AI hype from reality in medicine, I suggest starting with three … [+]
Artificial intelligence has long been heralded as a transformative force in medicine. Yet its potential has remained largely unfulfilled until recently.
Consider the story of MYCINa “rules-based” AI system developed at Stanford University in the 1970s to help diagnose infections and recommend antibiotics. Although MYCIN showed promise early on, it relied on rigid, predetermined rules and lacked the flexibility to address unexpected or complex cases that arise in real medicine. Ultimately, the technology of the time could not match the nuanced judgment of skilled physicians, and MYCIN never saw widespread clinical use.
Fast forward to 2011, then Watson from IBM gained worldwide fame by beating reputation Danger! champions Ken Jennings and Brad Rutter. Soon after, IBM applied Watson’s enormous computing power to healthcare, seeing it as a game changer in oncology. Watson, charged with synthesizing data from the medical literature and patient records at Memorial Sloan Kettering, wanted to recommend customized cancer treatments. However, the AI struggled to make reliable, relevant recommendations – not because of calculation errors, but because of inconsistent, often incomplete data sources. These include inaccurate electronic health records (EHR) and research articles that leaned too heavily toward favorable conclusions and did not hold up in real-world clinical practice. IBM closed the project in 2020.
Today, healthcare and technology leaders are wondering whether the latest wave of AI tools – including the acclaimed generative artificial intelligence (GenAI) models – will deliver on their promise in medicine or become a footnote in history like MYCIN and Watson.
Anthropic CEO Dario Amodei is among the AI optimists. Last month in a extended essay of 15,000 wordshe predicted that AI would soon reshape the future of humanity. He claimed that by 2026, AI tools (probably including Anthropic’s Claude) will be “smarter than a Nobel Prize winner.”
Specific to human health, Amodei praised AI’s ability to eradicate infectious diseases, prevent genetic disorders and double life expectancy to 150 years – all within the next decade.
As author of “ChatGPT, MD”, a non-fiction book about the future of generative AI in medicine, I admire parts of Amodei’s vision. But my scientific and medical background makes me doubt some of his most ambitious predictions.
When people ask me how to separate the AI hype from the reality in medicine, I suggest starting with three critical questions:
Question 1: Will the AI solution speed up a process or task like humans could? possibly complete on your own?
Sometimes scientists have the knowledge and expertise to solve complex medical problems, but are limited by time and costs. In these situations, AI tools can deliver remarkable breakthroughs.
To consider AlphaFold2a system developed by Google DeepMind to predict how proteins fold into their three-dimensional structures. For decades, researchers have struggled to map these large, complicated molecules; the exact shape of each protein took years and millions of dollars to decipher. Yet understanding these structures is invaluable because they reveal how proteins function, interact and contribute to disease.
Using deep learning and massive data sets, AlphaFold2 has achieved in a few days what would have taken laboratories decades, by predicting hundreds of protein structures. Within four years, it mapped all known proteins – a feat that gave DeepMind researchers a victory Nobel Prize in Chemistry and is now accelerating drug discovery and medical research.
Another example is a collaborative project between the University of Pittsburgh and Carnegie Mellon AI analyzed electronic medical records (EPDs) to identify adverse drug interactions. Traditionally, this process took months of manual review to uncover just a few risks. AI allowed researchers to screen thousands of drugs in days, dramatically improving speed and accuracy.
These achievements show that AI can bridge the gap when science has a clear path but lacks the speed, resources and scale for execution. If today’s generative AI technology had existed in the 1990s, ChatGPT estimates that it could have mapped the entire human genome in less than a year – a project that originally took thirteen years and $2.7 billion to complete.
Applying this criterion to Dario Amodei’s claim that AI will soon eliminate most infectious diseases, I think this goal is realistic. Current AI technology already analyzes enormous amounts of data on the efficacy and side effects of medicines, discovering new applications for existing medicines. AI is also proving effective in guiding the development of new medicines and can help tackle the growing problem of antibiotic resistance. I agree with Amodei that AI will be able to achieve in a few years what would otherwise have taken scientists decades, offering new hope in the fight against human pathogens.
Question 2: Does the complexity of human biology make the problem unsolvable, no matter how smart the technology is?
Imagine looking for a needle in a giant haystack. When a single answer is hidden in mountains of data, AI can find it much faster than humans alone. But when that ‘needle’ is metal dust spread across multiple haystacks, the challenge becomes insurmountable, even for AI.
This analogy shows why certain medical problems remain beyond the reach of AI. In his essay, Amodei predicts that within a decade, generative AI will eliminate most genetic disorders, cure cancer and prevent Alzheimer’s disease.
While AI will undoubtedly deepen our understanding of the human genome, many of the diseases that Amodei considers curable are “multifactorial,” meaning they result from the combined impact of dozens of genes, plus environmental and lifestyle factors. To better understand why this complexity limits the reach of AI, let’s first explore simpler single-gene conditions, where the potential for AI-driven treatment is more promising.
For certain genetic conditions, such as BRCA-related cancers or sickle cell disease that result from an abnormality in a single gene, AI can play a valuable role by helping researchers identify and potentially use CRISPR, an advanced gene-editing tool, to directly edit these mutations to reduce disease risk to reduce.
But even for single-gene disorders, treatment is complex. CRISPR-based therapies for sickle cell diseasefor example, require harvesting stem cells, processing them in a laboratory, and reinserting them after risky conditioning treatments that pose significant threats to patients’ health.
Knowing this, it is clear that the complications will only multiply when investigating multifactorial congenital diseases such as cleft lip and palate – or complex diseases that manifest later in life, including cardiovascular disease and cancer.
Simply put, editing dozens of genes simultaneously would pose serious health threats that would most likely outweigh any benefits. While the capabilities of generative AI are increasing exponentially, gene editing technologies like CRISPR face strict limitations in human biology. Our bodies have complex, interdependent functions. This means that correcting multiple genetic problems at the same time would disrupt essential biological functions in unpredictable, likely fatal ways.
No matter how sophisticated an AI tool may become in identifying genetic patterns, inherent biological limitations ensure that multifactorial diseases will remain unsolvable. In this regard, Amodei’s prediction about curing genetic diseases will prove only partially correct.
Question 3: Will AI’s success depend on changing human behavior?
One of the biggest challenges for AI applications in medicine is not technological but psychological: it’s about navigating human behavior and our tendency to illogical or biased decisions. While we might assume that people will do anything to extend their lives, human emotions and habits tell a different story.
Consider the treatment of chronic diseases such as high blood pressure and diabetes. In this battle, technology can be a strong ally. Advanced home monitoring and wearable devices currently track blood pressure, glucose and oxygen levels with impressive accuracy. Soon, AI systems will analyze these measurements, recommend diet and exercise changes, and alert patients and doctors when medication changes are needed.
But even the most advanced AI tools can’t force patients to reliably follow medical advice — or ensure that doctors respond to every warning.
People are flawed, forgetful and fallible. Patients skip doses, ignore nutritional recommendations and abandon exercise goals. On the clinician’s side, busy schedules, burnout, and competing priorities often lead to missed opportunities for timely interventions. These behavioral factors add layers of unpredictability and unresponsiveness that even the most accurate AI systems cannot overcome.
And in addition to behavioral problems, there are also biological problems that limit human lifespan. As we age, the protective caps on our chromosomes wear out, causing cells to no longer function. Our cells’ energy sources, called mitochondria, gradually fail, weakening our bodies until vital organs no longer function. If we don’t replace every cell and tissue in our body, our organs will eventually fail. And even if generative AI could tell us exactly what to do to avoid these shortcomings, it’s unlikely that humans would consistently follow the recommendations.
For these reasons, Amodei’s boldest prediction – that lifespan will double to 150 years within a decade – will not come true. AI provides remarkable tools and intelligence. It will expand our knowledge far beyond anything we can imagine today. But ultimately it cannot eliminate the natural and complex limitations of human life: aging parts and illogical behavior.
Ultimately, you should embrace AI promises if they build on scientific research. But if they violate biological or psychological principles, don’t believe the hype.