Artificial intelligence is increasingly capable of extracting eligibility criteria, matching patients to clinical trials, ranking potential participants, and supporting protocol-level decisions about enrollment. These capabilities are usually evaluated as problems of screening accuracy, recruitment efficiency, or operational scalability. Yet optimizing who enters a trial can also alter the population in which treatment effects are observed and, consequently, the clinical question the trial can legitimately answer. This article develops a methodological framework for preserving that question when artificial intelligence influences trial participation. The analysis links four domains that are commonly treated separately: eligibility design, treatment-effect heterogeneity, estimand specification, and transportability. The central distinction is between operational optimization that leaves the intended inferential target substantively intact and selection that changes the distribution of treatment-effect modifiers, narrows empirical support for the target population, or creates discordance between the population named in the estimand and the population represented by enrolled participants. Eligibility optimization is therefore treated as a scientific intervention on the evidence-generating population rather than as a neutral administrative improvement. The framework distinguishes legitimate enrichment, transparent redefinition of the target question, and unacknowledged clinical-question drift. It further separates model-level validation of matching or eligibility decisions from inferential validation of the resulting trial population. The resulting approach places population–estimand alignment, heterogeneity assessment, and transportability considerations upstream in the design of AI-supported recruitment rather than leaving them to post hoc interpretation.