Clinical trials are getting an AI layer that regulators will have to learn to read

The clinical trial stack is being rewritten from three directions at once. Unlearn is building digital twins that simulate patient cohorts before the first enrollment. PhaseV is layering AI into trial design and execution. City of Hope just built a national trials platform engineered to scale with oncology innovation. Each is a different bet on the same thesis: the bottleneck in drug development is no longer the science, it's the infrastructure around proving the science works.

Aaron Smith at Unlearn frames the shift as a move from trial simulation to a full scientific intelligence layer. Digital twins no longer just model what a placebo arm might have looked like—they inform dosing, endpoints, patient selection, and site strategy before sponsors commit capital. The question is no longer whether the model works in retrospect, but whether regulators will accept it as evidence in real time. That gap between capability and acceptance is where the next two years of conversation will happen.

Raviv Pryluk at PhaseV argues that AI changes trial design by surfacing the decisions that used to be locked in institutional memory or buried in subgroup exploratory analyses. The data exists. The compute exists. What didn't exist was the interface that let a sponsor ask the right question at the protocol stage instead of during a Phase III readout. That's changing, and it's changing faster than the regulatory frameworks designed to evaluate trial integrity.

City of Hope's national platform is the infrastructure play. Oncology moves fast, trials are becoming more modular, and sponsors need scale without losing fidelity. The platform is designed to absorb innovation—adaptive designs, decentralized arms, AI-selected cohorts—without rebuilding from scratch every cycle. It's a bet that the winners in clinical development will be the institutions that can take a new mechanism from concept to enrollment in weeks, not quarters.

The through line is speed, but the risk is legibility. Digital twins work. AI-optimized designs work. National platforms work. What regulators need to see is not whether they work, but how a sponsor shows their work when the model made the call. That's the bottleneck now, and it's a documentation problem, not a science problem.