Two Vocabularies, One Phenomenon: Metadata Bias in AI Evidence Synthesis on Fertility Decline

Abstract

Declining fertility is one of the defining policy questions of the next decade, and increasingly, what policymakers know about it is shaped by AI-synthesising the evidence base. But ask such a tool about reproduction and the answer depends on the word you use. The same phenomenon, framed clinically (e.g, infertility, IVF) or socially (e.g., childlessness, fertility intentions), is catalogued with radically different completeness. And the catalogue as much, if not more than the underlying scholarship, is what AI synthesis begins with (Bola\~nos et al., 2024). That databases under-index the social sciences, books, and grey literature is well established (Visser et al., 2021). What is new here is holding the topic fixed and asking whether metadata gaps act as a hidden policy filter on a single contested issue: the determinants of (in)fertility. We use two OpenAlex queries on the same phenomenon: a clinical basket (infertility, subfertility, ART, IVF, fecundity; n=101,645) and a social basket (childlessness, social infertility, fertility intentions, reproductive decision-making; n=3,646). We compare them on metadata completeness, open access, output type, and institutional provenance. The social framing is consistently less machine-legible: output skewed to books and dissertations, authorship university -- rather than healthcare-based. Open access rates are essentially equal (43.1% vs 41.3%), so the gap is in indexing depth, not paywalls, suggesting simple OA mandates will not fix it. On this same mixed literature, even before any coverage bias enters the picture, LLM tools already miss more than they catch when asked to extract hypotheses and claims (Uprety et al., 2025); the bias documented here compounds an already-imperfect extraction stage.