How Far Behind the Frontier? Adoption Lag of Fast-Growing AI Topics in Iran and 17 Other Research Systems, 2013–2025

Purpose. National AI research systems differ in how quickly they take up the topics that grow fastest worldwide. We ask how far Iran's portfolio lags the global shift toward fast-growing topics, whether this lag is larger than that of comparable systems, and whether it widens as the shift proceeds. Design and methods. Using OpenAlex counts of English-language, DOI-bearing AI articles (a corpus stable under a 2022 coverage break) by primary topic, country and year for 18 systems (2013–2025), we define fast-growing topics as the 20 topics with the largest gain in world share between 2013–2016 and 2022–2025 (excluding two generic catch-all labels). We measure each country's adoption lag as the time between the world and the country reaching a given share of output in these topics (20%, 30%, 40%) and as the horizontal shift that best aligns the country's share curve with the world's, with bootstrap intervals over topics and paired comparisons. Findings. The world share of output in the 20 topics rose from 13.0% (2013) to 51.9% (2025). Iran's share reached 20%, 30% and 40% about 1.8, 2.8 and 3.7 years after the world did (95% intervals −1.4 to 4.7, 0.4 to 5.2 and 1.3 to 4.6), a widening lag, and its matched-shift lag was 3.25 years (0.75–5.13), the fourth largest of 18 systems; its 2025 share was 41.6% against 51.9% worldwide. Türkiye (1.0 years at 30%), Egypt (−0.1) and the median of eleven emerging comparators (1.5 years shift) were closer to the world, while Brazil, Mexico, Poland and Russia lagged by 3–6 years and Pakistan, Saudi Arabia, India and Malaysia were ahead of the world curve. The paired differences between Iran and Türkiye (+1.9 years), Egypt (+2.8) and the emerging median (+1.8) were positive in 93–96% of bootstrap resamples but their 95% intervals included zero. Iran was under-represented in explainable AI (specialisation index 0.39), hate-speech detection (0.28), adversarial robustness (0.54), and educational technology (0.07–0.09), and over-represented in cancer detection (1.50) and imbalanced-data classification (1.35). Originality. The study replaces qualitative statements about lagging topic adoption with a quantified lag, with uncertainty, relative to the world and to peers, and shows which topics account for it.