AI must find drugs far more effective than current treatments to lower trial costs, says Hard Drugs host Jacob Trefethen
AI must find drugs far more effective than current treatments to lower trial costs, says Hard Drugs host Jacob Trefethen
On November 13, 2026, Saloni Dattani and Jacob Trefethen discussed how the success of existing treatments makes new ones harder to test on Hard Drugs, a podcast about medical innovation. The cost of testing depends on how much additional benefit a treatment offers patients and how quickly that benefit can be measured.
In 2012, researchers coined the term Eroom's law for a historical trend: over the period 1950–2010, the number of new drugs approved by the FDA per billion dollars spent on research and development, adjusted for inflation, roughly halved every nine years. Molecular synthesis and DNA sequencing were getting faster, while drug development was becoming more expensive. The authors attributed this in part to the success of existing drugs.
When a new drug is tested as an addition to an already effective treatment, researchers look for an additional benefit. A small difference between groups is difficult to distinguish from random variation in outcomes. With the same variability, trial conditions, and requirements for statistical reliability, an effect half as large requires roughly four times as many participants. Each participant must be recruited, examined, and followed up, turning a statistical challenge into a financial cost.
“Only pursue large effects,” Trefethen suggests. In his scenario, AI works with scientists to find much more effective treatments whose effects can be tested in the laboratory within weeks or months. Scientists could test and refine hypotheses more quickly. If a large effect also appears in patients, fewer participants may be needed to detect it.
Dattani counters that several drugs with moderate effects can produce a large effect in combination. She cites combinations of drugs used to treat cancer and HIV. Testing only drugs that have large effects on their own could mean missing useful combinations.
Trial duration also depends on how soon treatment outcomes can be measured. Dattani proposes looking for early markers that indicate how a disease will progress. To establish this relationship, long-term studies test whether a drug-induced change in a marker predicts an actual benefit for patients. A validated marker allows benefits to be assessed earlier and may shorten subsequent trials.
Trefethen cites imatinib, a drug for one form of blood cancer, as an example of a treatment that reached patients quickly. In 2001, the FDA reviewed the drug's application in two and a half months and granted accelerated approval. The decision was based on improvements in blood counts and a reduction in the proportion of cells carrying the chromosomal abnormality characteristic of the disease. The FDA judged these changes reasonably likely to predict a benefit for patients. The manufacturer was required to continue following patients and complete a study comparing the drug with another treatment.
If AI helps more drug candidates reach human testing, trial organizers will face a greater workload: recruiting participants, bringing research sites on board, and testing different drugs and combinations. The pace of drug development will also depend on how many of these trials the medical research system can conduct at once.