Our research focused on the ‘computer science, artificial intelligence’ category from the Web of Science (WoS), considering literature published up to the end of 2023. We initially obtained 5,991 results based on the review document type assigned by WoS. However, due to potential misclassification of review articles in WoS (Blümel & Schniedermann, 2020), this study underwent automatic preprocessing and manual review to exclude incomplete or non-review articles. As a result, we compiled a final set of 4,315 AI review articles.

To identify major research themes within this body of literature, we performed a keyword frequency analysis, allowing us to systematically uncover prevailing topics and trends discussed in the AI review articles. These findings were then structured with the task-technology fit (TTF) model (Goodhue & Thompson, 1995), which categorizes research topics based on their focus on either tasks, technology, or applications (i.e., the utilization of TTF). Accordingly, we classified the identified topics into three broad categories:

Task-oriented reviews: these reviews emphasize AI tasks that are central to various domains. Examples include tasks such as ‘classification’, ‘pattern recognition’, and ‘clustering’. These reviews often concentrate on algorithmic methodologies and performance evaluation specific to task-related challenges.

This study mapped the temporal trends of AI review articles across three domains (see Figure 1), observing a general increase in publication frequency over time, especially notable since 2019. This aligns with the rapid growth of scientific literature noted by Bornmann et al. (2021). Then, we introduced ‘density’ as a measure for the frequency of articles per year, identifying the ‘COVID-19’ topic within the application domain as having the highest density (19.5) despite its brief existence. Other topics like ‘internet of things (IoT)’, ‘generative adversarial networks (GAN)’, ‘blockchain,’ and ‘federated learning’ have also shown high densities and an upward knowledge accumulation trend. Moreover, traditional task domain topics like ‘classification’ and ‘object detection’ demonstrate high densities with stable development.

As shown in Table 1, the comparative analysis of the past five years and five years prior revealed that the task domain typically showed higher article densities than the technology domain, which outpaced the application domain. The last five years witnessed significant density increases across all domains. Annual growth rates across these domains have converged and stabilized post-2019, with the application domain experiencing a marked growth rate increase after 2019, while the task and technology domains maintained consistent growth rates before and after 2019, stabilizing thereafter.

This study delved into the age distribution of references cited in AI review articles across three domains, defining ‘reference lag years’ as the time difference between the publication year of the review and its cited literature. Figure 2 revealed a predominant lag of 2 to 3 years in task and technology domains, with the application domain occasionally extending to 4 years. Post-peak, all domains showed a rapid decline in citation numbers with increasing lag time.

A closer examination of reference lag years within a five-year span showed distinct patterns as shown in Table 2, the application domain’s citations peak in the second year (0.136), the task domain in the third year (0.123), and the technology domain equally in the second and third years (0.123). The application domain demonstrated notably higher citation frequencies in the first two years compared to task and technology, suggesting quicker incorporation of recent literature.

Across all three domains, citations with a lag of up to five years constituted over half of the total, with the highest proportion in the application domain.

In addition to analysing the relative reference lag years, this paper also examined the overall citation time distribution of review articles across different years. As shown in Figure 3, recent review articles span longer historical periods than older reviews. Specifically, literature time spans starting years shifted significantly backward from 1971, 1962, and 1958 in the year 2000 to before 1950 by 2022 and 2023 for the application, task, and technology domains respectively. Early 2000s and recent reviews showed more concentrated interquartile ranges, with wider ranges and noticeable fluctuations in other periods.

In examining the dynamics of references cited in AI review articles, we define references as ‘knowledge units’ with temporal attributes, together forming a ‘knowledge set’ over time. Our assessment of knowledge accumulation for each year’s new references is threefold: (1) accumulated references: the cumulative total of references cited in all review articles up to a given year, indicating the knowledge set’s size to that date. (2) New references: references cited for the first time in that year, expanding the existing knowledge set. (3) Types of new references: distinguished as either recent (published within the past year) or previous (published more than a year ago).

As shown in Figure 4, the ratio of accumulated to new references indicates a predominance of accumulated references annually. Within new references, those published over a year ago consistently outnumber recent publications. Nevertheless, the share of recent references within new citations has been rising lately. Quantitative analysis shows that the proportion of recent references in new citations post-2019 has increased across all domains.

This study investigates the dynamics of knowledge accumulation in scientific fields by analysing review articles across multiple dimensions. Using AI research as a case study, we uncover patterns that are likely applicable to other evolving domains. Our analysis reveals several noteworthy findings and insights. Firstly, with regard to the temporal distribution characteristics, there has been a significant surge in their publication frequency, particularly in the past five years. This surge indicates that the pace of knowledge accumulation in AI is rapidly accelerating, aligning with the explosive growth of AI publications (Leeming, 2021). The domain of application has experienced the highest growth rate in terms of review articles, which can be attributed to the emergence of AI as a powerful tool for addressing complex interdisciplinary problems across various research fields (Zhuang et al., 2020). Secondly, in terms of the age distribution characteristics of references, more than half of the references cited in review articles across all domains fall within a five-year timeframe, with the application domain exhibiting the highest proportion. This trend signifies a notable acceleration in knowledge accumulation across the domains, especially in the application domain over the past five years, which aligns with the observed temporal distribution patterns. Furthermore, the analysis demonstrates that recent review articles encompass a broader timeframe of references. This finding is consistent with the nature of scientific knowledge accumulation, which involves the growth of the available stock of knowledge over time (Petrovich, 2018). Thirdly, in relation to the updating patterns of cited references, the number of older references cited in newer reviews consistently exceeds that of more recent references. This observation confirms the existence of a citation lag and suggests a pattern of delayed citation across review articles. Moreover, the increasing proportion of recent references among new citations in recent years indicates a heightened pace of knowledge accumulation in the three domains under examination.

Despite these insightful findings, the limitations should be acknowledged. Our analysis focuses on the AI research field. The findings may not generalize to other disciplines. Future research can explore different fields to help further clarify the role of review articles in understanding the knowledge accumulation phenomenon.