Oncology clinical trials have a structural problem that has persisted for decades. They enroll patients too slowly, fail at high rates, and produce results that often do not translate well to the broader population of people with cancer. These are not minor inefficiencies. They represent a fundamental gap between the scientific ambition of cancer research and its operational reality. A review published in Nature Reviews Clinical Oncology examines where artificial intelligence can genuinely close that gap, and where it cannot yet.
The Operational Bottleneck AI Can Actually Address
The most immediate and evidence-supported role for AI in oncology trials is not the dramatic one. It is not replacing clinical judgment or generating synthetic patients. It is, more precisely, fixing the paperwork problem at scale.
Identifying which patients are eligible for a given trial requires parsing complex eligibility criteria against detailed medical histories, a process that is time-consuming, error-prone, and often performed manually. AI systems trained on large electronic health record datasets can automate significant portions of this work: screening patient records for eligibility, extracting structured data from clinical notes, and flagging candidates for enrollment. These applications are already being deployed at select academic cancer centres, according to the review.
Remote patient monitoring and real-time trial data extraction fall into the same category. These are tasks where the volume of information exceeds what human teams can process efficiently, and where AI augments the existing workflow rather than redesigning it. The review is explicit on this point: the most defensible near-term role for AI is operational augmentation under human oversight, not autonomous decision-making.
This is what most coverage of AI in medicine misses. The headline-grabbing applications, such as AI diagnosing cancer from images, attract attention. The quieter work of matching a patient to a trial they would otherwise never hear about may ultimately matter more for how many people benefit from research.
Where AI Is Not Ready: Synthetic Evidence and Digital Twins
The review draws a clear line between what AI can do now and what it is being asked to do prematurely. Applications designed to substitute for clinical evidence generation, including synthetic control arms, outcome-prediction simulations, and digital twins, remain at earlier stages of development. The methodological challenges are unresolved, prospective validation is limited, and regulatory frameworks have not yet caught up.
A synthetic control arm, to explain the concept, is a computationally constructed comparison group built from historical or real-world data, intended to replace a traditional placebo or standard-of-care arm in a trial. The appeal is obvious: fewer patients needed, faster trials, lower costs. The problem is that the validity of such an approach depends on whether the synthetic population genuinely resembles the patients in the trial, and that assumption is difficult to verify and easy to violate in ways that are not immediately visible.
Digital twins, virtual models of individual patients used to simulate how they might respond to treatment, face similar challenges. The concept is scientifically compelling. The evidence base for using them to make consequential clinical decisions is not yet there.
Regulatory oversight of these tools is best understood, the review argues, as a risk-proportionate continuum rather than a simple regulated-or-not binary. Both the FDA and EMA are converging on frameworks that emphasize model transparency, credibility assessment, and post-deployment monitoring. That convergence matters because it signals that regulators are not waiting for a single definitive standard to emerge before engaging with the technology.
Why This Distinction Matters Beyond Medicine
The framework the review establishes, separating AI applications that augment human workflows from those that attempt to replace evidence generation, has relevance well beyond oncology. It is a useful lens for evaluating AI claims in any high-stakes domain.
The pattern is consistent across fields: AI performs well when it handles volume, pattern recognition, and structured data extraction in support of human decision-makers. It becomes unreliable when it is asked to substitute for the kind of prospective, controlled evidence generation that exists precisely because human intuition and historical data are insufficient guides to what will actually happen.
The review also raises cross-cutting concerns that apply broadly: algorithmic bias, data drift over time, calibration failure, and equity in access. An AI system trained predominantly on data from well-resourced academic medical centres may not perform equally well when deployed in community hospitals or in populations that were underrepresented in the training data. This is not a theoretical concern. It is a known failure mode that requires active monitoring and diverse validation datasets.
The call for coordination among clinicians, trialists, regulators, industry, and patients reflects something important: no single actor in this system has the information or authority to solve these problems alone.
In Short
AI is beginning to make oncology clinical trials more efficient, primarily by automating the identification and screening of eligible patients and by supporting real-time data monitoring. These applications are operational, evidence-supported, and already emerging in practice. More ambitious uses of AI, such as synthetic control arms and digital twins, remain unvalidated and face unresolved regulatory and methodological questions. The distinction between augmenting human workflows and replacing clinical evidence generation is the most important line to hold as this technology develops.
Based on reporting from Nature: Machine Learning.