To characterize planned AI use and its description in interventional trials registered on ClinicalTrials.gov with start dates from January 2022 onward.
Approach:
Design and data source: Cross-sectional analysis of ClinicalTrials.gov records retrieved through its API on December 17, 2025.
Eligibility and review: Eligible interventional trials had a start date on or after January 1, 2022, and an AI-related term in the title, brief summary, or detailed description. Reviewers annotated AI presence, type, data modality, and purpose.
Key Findings:
Of 8,559 records identified, 1,212 trials were included after eligibility screening and exclusion of records without confirmed AI involvement.
Trial counts rose annually from 2022 through 2025. Planned generative AI use increased from none among trials initiated in 2022 to 100 of 491 (20.4%) initiated in 2025; overall, 166 trials (13.7%) planned generative AI.
Cancer was the most common condition (20.0%), followed by neurological and mental health conditions (19.6%) and cardiovascular and metabolic disease (18.6%).
Screening and predictive modeling for disease risk or treatment selection were the most frequent AI applications (39.8%); images and video were the most common data modality (29.2%).
Most trials were single-site (69.3%); median target enrollment was 150 participants (IQR, 60-500). The United States accounted for 303 trials (25.0%) and China for 149 (12.3%).
All records described AI or machine learning, but specific methods were frequently missing and could be identified only in free-text entries.
Interpretation:
The authors report that registered trials planned AI applications across multiple clinical areas and that generative AI use increased in registrations before corresponding trials were published. They note that the lack of structured registry fields and reporting requirements limited transparency about AI methods.
Limitations:
Registry records describe intended rather than confirmed trial conduct, so actual AI use could not be verified.
Reporting completeness for individual elements was not measured because descriptions were too heterogeneous for reliable classification.
The counts do not show the proportion of all registered trials involving AI and should not be interpreted as evidence of widespread adoption.
Excluding trials by status disproportionately removed records with earlier start dates, potentially overstating growth in registered AI trials.
Conclusion:
The authors state that structured registry fields aligned with AI reporting standards would support tracking, reproducibility, and evidence synthesis.
