AI’s Footprint in Clinical Trials Is Moving From Keywords to a Governance Issue

Artificial intelligence’s entry into medicine is often portrayed as a race over model accuracy, image interpretation, or new drug design. But in the setting of clinical research, another more fundamental question is emerging: when a trial says it uses AI, is it being used to screen patients, assist physicians, interact directly with patients, or appear only at the data analysis stage? These differences are not just technical classifications; they also affect ethics review, risk disclosure, and regulatory judgment.

A preprint submitted to arXiv in May 2026 attempts to map this trajectory through registration data from ClinicalTrials.gov. The research team searched clinical trial records containing AI-related terms, analyzed changes over time and geographic distribution, and further explored how “human-AI interaction” is described in registration text. Because ClinicalTrials.gov is part of the clinical trial infrastructure, the focus of this study is not to prove that a particular medical AI is effective, but to examine whether the overall trial registration system is sufficient to show how AI is intervening in clinical care.

According to the paper’s abstract, AI-related trials have increased markedly in recent years, with terms rising in frequency including machine learning, deep learning, chatbot, GPT, and large language model. Geographically, China and the United States account for the largest numbers of AI-related trials, while countries including Italy, France, Spain, the United Kingdom, and Turkey have also shown recent growth. This distribution suggests that the clinical adoption of medical AI is not a single-market phenomenon, but is gradually becoming embedded in the research and care systems of different countries.

One especially interesting layer of the study is that the team used a hybrid workflow of human review plus generative AI to examine whether the registration records actually involved substantive AI use, and what forms of human-AI interaction appeared. Among 100 randomly selected records, human and AI classifiers showed fairly good agreement in identifying records that “did not substantively use AI.” But agreement was lower when determining types of human-machine interaction, especially when descriptions were ambiguous about whether healthcare professionals interacted with AI systems, making classifications more likely to diverge.

This result pushes the issue from “Can AI help read trial registrations?” toward a more practical level: many clinical trial records may still not use sufficiently clear language to explain where AI systems sit in the workflow. If a chatbot directly faces patients, and a model only performs risk stratification behind the scenes for researchers, the safety monitoring, consent procedures, and allocation of responsibility involved are not the same. If registration data are written in broad terms, subsequent systematic reviews, regulatory audits, and public understanding will all be weakened.

Caution is needed: this remains a preprint and has not yet undergone peer review; the currently available information also comes mainly from the paper’s abstract, rather than full external validation. It analyzes terms and descriptions in registration text, which is not the same as evaluating the clinical efficacy, safety, or actual deployment quality of each AI intervention. Keyword searches may also miss studies that do not use standard AI terminology, or include records that only use popular terms loosely.

Even so, this analysis raises a concrete and urgent biomedical AI issue: clinical trial registration is not just an administrative form, but an entry point for society to understand medical innovation. As AI begins to appear across recruitment, diagnosis, treatment recommendations, patient communication, and research analysis, trial records need to label more precisely the model’s role, usage context, form of human oversight, and interaction targets. Otherwise, the growth of AI in clinical research will first be seen in numbers, but may not necessarily be understood with sufficient clarity.