In November 2020, as the world battled the COVID-19 pandemic, a different kind of breakthrough captured global attention. Google DeepMind announced that its AlphaFold model had solved the protein-folding problem, one of biology’s most stubborn puzzles. The announcement was hailed as the scientific equivalent of a moon landing. Newsrooms called it a revolution that could bring new medicines to market faster than ever before.
But half a decade later, the flood of new cures has not materialised. Despite billions of dollars being invested in artificial intelligence (AI), drug discovery remains a slow and expensive process. This paradox lies at the heart of what analysts Jack Scannell, Alex Blanckley, Helen Boldon, and Brian Warrington called Eroom’s Law in a 2012 paper.
Quantity-quality mismatch
When Gordon Moore predicted in 1965 that computing power would double every two years while costs halved, he captured the astonishing pace of progress in electronics — a rule that came to be called Moore’s Law. But in medicine, the opposite has happened. Eroom’s Law (‘Moore’ spelt backwards) observes that the number of new drugs discovered per billion dollars spent has been falling steadily for decades.
Today, it costs several times more to bring a drug to market than it did in the 1970s, despite the availability of vastly superior computers, labs, and algorithms. In short, the chips have raced ahead but the pills have slowed down.
In drug discovery, every new treatment begins with a hypothesis, an educated idea or guess about how a molecule might influence disease. For decades, the real constraint has never been the quantity of hypotheses but the quality. Even before the advent of AI, researchers generated millions of plausible ideas, most of which led nowhere. With today’s AI systems, that number has grown to billions, yet the quality of hypotheses has not improved. Algorithms can exponentially increase the quantity of hypotheses but cannot enhance the quality by infusing it with intuition or imagination. The leap from quantity to quality remains a distinctly human privilege.
Creativity and chaos
AI using deep learning techniques, such as AlphaFold, thrives on patterns where clear, well-defined relationships are hidden within data. The protein-folding problem suited this perfectly. By 2015, scientists had already mapped over 1.5 lakh protein structures through five decades of human effort using X-ray crystallography, fluorescence spectroscopy, and protein nuclear magnetic resonance spectroscopy.
There was a known question, a vast dataset, and an idea of what a correct answer — all conceptualised by humans — should look like.
AlphaFold’s success was thus akin to a brilliant student topping a national entrance exam, such as the NEET or UPSC. The questions were difficult but predictable; the syllabus was vast but well known; and years of human groundwork had built the coaching material. With enough computational practice, the student could achieve top ranks.
Drug discovery, however, is not an examination; it is an act of exploration. It resembles a cricket talent scout trying to spot a future Virat Kohli in a dusty village ground for his IPL team or a political analyst attempting to predict who might become India’s next prime minister. There is no fixed pattern, no set syllabus, and no reliable coaching manual. On the other hand, randomness dominates in the wilderness in which drug discovery operates.
Accidents v. AlphaFold
Penicillin was discovered because Alexander Fleming forgot to cover a petri dish. Insulin was discovered through a series of messy experiments on dogs, conducted by Frederick Banting and Charles Best, who were simply trying to isolate pancreatic extracts. Paracetamol originated from a 19th-century misidentification in a laboratory notebook and metformin was studied for the treatment of influenza before its role in diabetes was understood.
Today’s world is also far more ethical and careful, rightly so. Every molecule is required to pass through stringent preclinical tests and multi-phase clinical trials before reaching patients. This caution while essential has also slowed the journey of discovery. Earlier scientists could test wild ideas with relative freedom; today’s researchers navigate mountains of paperwork and risk assessments. So even when AI proposes a promising molecule, the path to a prescription bottle remains a long and arduous marathon.
AlphaFold could succeed in cracking a computational challenge because it was solving a bounded problem: one where rules existed and human scientists had already mapped the territory. To be sure, AI will continue to reshape various aspects of medicine, including screening, clinical trial design, and drug repurposing. But expecting it to create or develop new cures single-handedly is folly. AI excels when guided by questions that humans already know how to ask and verify, thus ensuring its answers are accurate and reliable. More broadly, AI can reproduce knowledge at a faster pace but not imagine or create it. So while it will continue to reshape various aspects of medicine, including screening, clinical trial design, and drug repurposing, expecting it to create or develop new cures single-handedly would be folly.
As history shows, every great leap in medicine, from insulin to paracetamol, began with a human mind willing to wander beyond the data.
(Note: AI’s capabilities described here are as of November 2025.)
Dr. C. Aravinda is an academic and public health physician. The views expressed are personal.
