Innovation in the Artificial Intelligence sector: An exploratory study of the structure of co-inventor networks

Lantz, Jean-Sébastien; Lacaze, Delphine; Braune, Eric; Sahut, Jean-Michel

doi:https://doi.org/10.59876/a-es2g-vy4a

Innovation results from the combination of different actors’ knowledge (Schilling & Phelps, 2007). In networks that facilitate information circulation, these actors mobilize their social capital (Bourdieu, 1986; Coleman, 1988) to accelerate knowledge flow, a phenomenon that has grown rapidly over the past 20 years. Patent inventors use this knowledge that they master to innovate. Recent research has accumulated knowledge on co-inventor networks by viewing the firm as a system conveying and producing knowledge via social capital (Zhou et al., 2007). Previous studies have examined how inventors’ positions affect the relevance of their innovations which is usually measured by patent citations for the considered innovation. In this respect, Aral (2016) and Toth & Lengyel (2021) demonstrated the structural hole hypothesis (i.e., brokerage, weak ties, and information diversity) and the central position hypothesis (i.e., strong ties and greater channel bandwidth). Furthermore, Bruggeman (2021) linked patent citations to inventors’ strong ties or ability to broker two clusters. These studies have the advantage of analyzing the structure of co-inventor networks using graph-theoretic measures, such as the number of network components, modularity or diameter, and this makes it possible to show that the structure of each network influences its output (Lin, 2017). But the USPTO’s patent citation count is used to assess the connections between inventors. This proxy is noisy, as US patent applicants must include an exhaustive list of known or supposed prior art, on pain of prosecution and heavy penalties. As a result, patent citations are a biased measure of the connections between applicants.

Another approach is to analyze co-inventor networks based on registered patents. Argyres et al. (2020) used this approach, investigating the interactions between the organizational structure of formal R&D networks and the type of innovation produced. However, these researchers considered the network structure in relation to the centralization of R&D budget. Hence, flows of money are substituted for flows of knowledge and information, underpinned by the assumption that the R&D budget authority guides collaborations among inventors who want to secure funding for their projects and advance their careers. Moreover, the measure of centrality used is biased because, in many countries, the R&D budget is not necessarily representative of the way R&D is actually carried out within teams because of tax credits linked to R&D expenditure (Ciaramella, 2023).

In addition, some authors have shown that the structuring of co-inventor networks and their dynamics differed according to the technologies and regions considered (Turkina & Oreshkin, 2021). Graf (2012) focused on the differences between the semiconductor and information technology (IT) sectors. The dynamic analysis of inventor networks in both sectors showed an increasing connectivity and the emergence of a large component in the semiconductor sector but not in the IT sector. While Abbasiharofteh et al. (2023) show the different dynamics of co-inventor networks in different regions. Except for Tsay & Liu’s 2020 study on assignees’ perspectives, cooperation networks in the Artificial Intelligence (AI) sector are understudied.

Consequently, the aim of the present study is to analyze what is the structure of co-inventor networks of the most innovative companies in the world in the AI sector.

To do this, we link the position of actors to the properties of the networks in which they operate, and which are largely imposed on them (Tsouri, 2022; Brennecke and Rank, 2017; van der Wouden & Rigby, 2019). We thus reveal the network forms observed among the 30 most innovative companies in the world in the AI sector. Our factor analysis conducted on the characteristics of the networks and the positions of the inventors inside them enabled us: 1) to reduce the complexity of the interaction variables resulting from the analysis of networks and graphs by identifying two dominant dimensions, i.e., “the encapsulation of information and knowledge” and “information and knowledge accessibility”. The first dimension is based on the structural aspects of the networks, whilst the second dimension is related to the inventors’ position in the co-inventor network; 2) to show that the company’s regional origin determines the structure of co-inventor networks; 3) to identify that the most innovative companies make a clear choice between “encapsulation of diverse information and knowledge” and “information and knowledge accessibility” when structuring their intra-firm co-inventor networks, and that competition between different regions of the world interfere in this process. Thus, this research contributes to the nascent debate concerning the influence of the structures of co-inventor networks on their outcomes (Lin, 2017).

This paper is organized as follows. The literature review recalls the social dimension of innovation linking it to the circulation of knowledge and the social capital of inventors. It shows how the over-arching theory, i.e., social capital, is pertinent to analyze the structure of co-inventors’ networks. The research methodology section describes the organization of our research before presenting the data and results of the analyses we conducted. We finally discuss the results and underline some limitations of our research in the conclusion.

Literature Review

Social capital, knowledge, and innovation

Social capital—resources developed through human interactions (Adler and Kwon, 2002)—can benefit individuals and organizations. Social capital in social organizations is made up of networks where mutually respected norms foster trust. Mutual benefits from coordination and cooperation reduce malfeasance and transaction costs (Putnam, 1995; Woolcock, 1998). A social structure with high social capital is economically successful (Sillicon Valley, Putnam, 2000). Reciprocal obligations and expectations based on mutual trust and information circulation have been studied in social capital research since Coleman (1988). Social capital characterizes a relationship structure rather than an actor (Coleman, 1988). Reflecting the constraints of social structure, the social network embodies the study of social capital, the latter constituting a resource for actors supposedly interested in exchanging information (Granovetter, 1974). Subsequent research on social networks has focused on the dissemination of information via personal and professional connections with other people as its primary concern (Zhou et al., 2007) and has revealed national cultural influences (e.g., Tsujinaka, 2002).

As far as innovation is concerned, one of its most important precursors is access to diverse sources of information and knowledge (Cohen and Levinthal, 1990; Chesbrough, 2003) within a firm’s network (e.g., Laursen and Salter, 2006) as well as from the external environment (Eisenman and Paruchuri, 2019; Moreira et al., 2018). Indeed, social capital can enhance innovation through the different processes that take place within a co-inventor network such as knowledge acquisition and spillovers: knowledge sharing (Bourdieu, 1986; Coleman, 1988), interpersonal and social trust (Putnam, 1993), facilitation of joint problem solving (Lettl, Rost and von Wartburg, 2009), transformation of information into knowledge (Landry, Amara, and Lamari 2002). In shaping a firm’s innovation activities and the consecutive knowledge creation, the literature on innovation has shown how inventors i.e., the individuals who have participated in the development of a patent (Pinto et al., 2019), play a key role (e.g., Grigoriou and Rothaermel, 2017) in innovation, as well as their surrounding social environment (e.g., Fisher and Qualls, 2018).

Co-inventor networks

Rapid technological development and increasing research costs have meant that innovation is also increasingly dependent upon collaboration being successful (Ribeiro, Rapini, Silva, & Albuquerque, 2018). As inventors do not possess or recall all knowledge that can be combined to create new knowledge (Fleming, 2001), mutual assistance between colleagues enables them to learn new knowledge or be reminded of known knowledge (Furukawa & Goto, 2006). Thus, mobilizing their social capital, inventors of a network collaborate to share deep and diverse knowledge (Aral & Van Alstyne, 2011). Yet, the type of information and knowledge inventors will be able to access in such collaborations depends on the characteristics of the structure of the innovation network and on the position, they occupy within it (Nerkar & Paruchuri, 2005). Also, the dynamic of information sharing differs between bridging or bonding innovation network. Bridging occurs in networks of heterogeneous groups while bonding enhances the networks of homogeneous groups of actors (Granovetter, 1973).

As the topology of a network regulates the process of information distribution (Gulati, 1998) and influences the amount and the accuracy of information and resources shared (Carnabuci & Operti, 2013), identifying the advantages and disadvantages of the various kinds of network structures is important. Although research is still in its infancy (Pinto et al., 2019), past studies have revealed that collaboration networks share a number of common structural properties which facilitate their analysis and comparison (Barabási & Bonabeau, 2003). Empirical studies on patent collaboration networks have also found that the structures of co-inventor networks vary between countries and over time (Ferrara et al., 2017; Lim & Kidokoro, 2017). In addition, some research collaboration communities, observed in Europe, the United States, Japan and South Korea, were found to favor innovation when they were embedded within local and organizational environments (e.g., Tóth et al., 2018) or organized around shared equipment or technology (e.g., Claudel et al., 2017). Japanese and Korean research communities appeared highly clustered though maintaining connections with international inventors and firms; with Korean networks especially reserving these international connections for the central actors of the system (Lim & Kidokoro, 2017). All in all, these studies underline how the structures of co-inventor networks can be driven by cultural and thus regional considerations. The superiority of one type of network over the other has not been demonstrated. Originally, it was argued that the bridging of social capital enabled access to more diverse nonredundant knowledge (Putnam, 1993) and favored innovation (Reagans & Zuckerman, 2001; Tortoriello & Krackhardt, 2010) while bonding in a dense network of closure ties was said to provide homogeneous, overlapping and subsequently less valuable knowledge (Burt, 1992; Granovetter, 2018). However, strong ties between actors can foster a common culture, trust and prosocial behaviors (Carlile & Rebentisch, 2003; Tortoriello, 2015) and consequently provide access to targeted and therefore relevant information (Burt, 2000) that foster innovation (e.g., Hur & Park, 2016). To reinforce innovation, it has been argued that when the necessary information is simple, the diversity of the network obtained by weak ties is preferable, whereas when the relevant information is complex and rapidly evolving, strong ties characterized by mutual trust and understanding are appropriate (Bruggeman, 2016).

In sum, the superiority of one type of network over another is not apparent: brokered networks provide non-redundant diverse information, but it may be not pertinent, too abundant, of a low quality, too centralized, and it may involve unproductive efforts for the brokers. Cohesive networks may provide redundant useless information but also targeted knowledge due to rapid access to pertinent information obtained within relationships and promote rapid dissemination of external knowledge inside the network (Toth & Lengyel, 2021). Thus, the diversity-bandwidth trade-off seems to be highly contingent upon the type of information needed and the context; in addition, it is possibly the combination of bridging and bonding which may produce the best innovation outputs (Aral, 2016). Schilling & Phelps (2007) supported this proposition with empirical data on inter-firm co-inventor networks, but their findings have not been validated on intra-firm networks to the best of our knowledge. Indeed, intra- and inter-firm networks need to be distinguished due to their different dynamics (Phelps et al., 2012).

Research Methodology

Preliminary justifications

Previous research concerning networking of R&D activities have been biased towards qualitative and empirical rather than quantitative analysis (Komoda et al., 2021). We have decided to undertake a quantitative method, which is appropriate when a knowledge-based perspective is the primary concern (Hoskisson et al., 1999). So, our study examines the properties of the co-inventor networks using measurement allowed by graph theory.

Research context

Also, research has been criticized for analyzing patents in a homogeneous way in large-scale study samples without considering the company’s strategy (Gittelman, 2008). Past research on co-inventor networks conducted on a range of sectors may have hidden some sectorial specificities (Argyres et al, 2020). Therefore, we narrowed the sample to the AI industry, and within that industry, to the 30 largest patent applicants for AI, and the characteristics of the most productive inventors’ positions within these networks[1].

Several observations justify our choice of the AI industry. In the years to come, the investments linked to the IA will probably account for 4% of the GDP in the United States (Goldman Sachs, 2023). AI raises sovereignty issues (Mügge, 2024) and many countries will devote between 1.5% and 2.5% of their GDP to similar activities (ibid). The AI industry will therefore be the most active in terms of global R&D investment. They could reach 100 billion dollars in the United States and 200 billion dollars worldwide by 2025[2]. These investments will be made by companies in many countries enabling cross-country comparisons. For all these reasons, we thought that a study focused on AI industry could reveals the properties of the archetypal R&D networks.

Data analysis strategy

We first used the Spearman rank test to analyze the association between the rank of each company in relation to the number of patents filed and the rank of these companies according to each variable describing the position of their most prolific inventor and the structure of their co-innovator network. This analysis highlighted the variables associated with performance, according to the number of patents, of each company in the sample. The Spearman rank test was chosen over Pearson correlations because of the small size of the sample. Moreover, analysis of the distribution of the variables showed that many had a heavy tail, which argued in favor of using the Spearman rank test (Winter et al., 2016).

However, the Spearman rank test does not explain how different variables may be combined to boost the performance of each company. Therefore, we complemented the study with a factor analysis. This method was primarily used to reveal the latent dimensions unifying these variables while retaining the common variance of the data. The latent dimensions are linear combinations of the original variables (Ringnér, 2008). Factor analysis retains most of the information included in the original data set, while the dimensionality reduction of the data enables us to plot the sample to highlight similarities and differences between the observations. It is important to underline the fact that factor analysis is a descriptive and not an explanatory approach. Consequently, the number of patents filed was described through how companies jointly drew on these dimensions to complete their innovative projects, but causal relationships between patent numbers and the features of the inventors’ positions in the networks or the structures thereof fell outside the scope of this study. Hence, our study is descriptive but not causal.

Data and measurement

Based on the WIPO Technology Trend on Artificial Intelligence report (2019), we selected the 30 companies that obtained the most patents in 2018 (Appendix 1). Extraction of data from the KNWL500.com database enabled us to identify 92370 patents and 212451 inventors attached to the 30 companies selected (Appendix 2).

The KNWL500.com database, created from the United States Patent and Trademark Office (USPTO) data, offers extensive information on the technical innovation capabilities of the AI patenting industry. This database provides information about the names of inventors, their location and the patent owner or assignee. The latter specifies the party/organization to which the intellectual property right is intended to be transferred (WIPO Intellectual Property Handbook, 2008). This makes it possible to obtain all patents registered in the name of a company, whether it has developed its patents alone (from its parent company or subsidiaries) or through collaborations with other organizations.

The database was built in three phases. In the first step, the patents attached to each company selected were collected. As the purpose of this research is to examine certain businesses, a patent set was found by researching the patent assignee name of the company under investigation. The patents attached to the assignee name were retrieved from the USPTO database. This data sample was generated based on all inventors who patented their innovations with at least one of the 30 companies in the AI sector stated above.

In the second step, an examination of the networks of patented co-inventions was carried out. In their initial configuration, networks have two modes i.e., links unite two different categories of actors, namely, patents and their inventors. Therefore, identifying the relationships between inventors requires modeling the matrix of their network in parallel. The result is a network presenting only the relationships between the inventors.

As can be seen in Figure 1, the co-inventor network was developed between inventors via the formation of links between them when they submitted at least one patent jointly (Hur and Park, 2016). Inventors who collaborate on a project together are familiar with one another and may share knowledge derived from their engagement in parallel or past projects.

Figure 1

**Construction of the co-invention network**

In the third step, measurements of the position characteristics of each node in the co-inventor network were performed. This work enabled us to identify the position features of each R&D engineer in the inventors’ networks.

Social Network Analysis (SNA)

Simply from their visual appearance, networks exhibit complex structures generally difficult to summarize. Their parts may be more or less densely interconnected, sometimes with cores containing most of the links. Participants in the network can be more central or more peripheral (Easley and Klienberg, 2010). Hence, representing and measuring networks is the main aim of network analysis. From its inception network analysis overlaps with graph theory developed by mathematicians. These technics and measurements have been largely adopted for the analysis of social and economic networks (Jackson, 2008). SNA is one of the prevalent patent analysis tools used for the purpose of examining the patterns of interactions that take place within a set of inventors by utilizing co-inventor information (Pinto et al., 2019). These patterns of interactions represent information flows and inventive relationships among inventors. SNA is based on the concepts formed within structural theory and graph theory. It analyzes properties of networks such as connectedness, density, and centrality (Wasserman and Faust, 1994). The nature of the links that exist between actors is revealed by dedicated measurements revealing the influential actors within the network (Otte and Rousseau, 2002).

Properties of the analyzed networks and features of the position occupied by each inventor lead us to differentiate two types of measurements.

Properties of the analyzed networks

Number of components: In graph theory, a component of an undirected graph is a connected subgraph that is not part of any larger connected subgraph. The components of any graph partition its vertices into disjoint sets and are the induced subgraphs of those sets. A graph that is itself connected has exactly one component, consisting of the whole graph. Components are sometimes called connected components.

Modularity: The existence of densely connected groups of nodes, with only sparse connections between groups, characterize the extent to which a network is modular. Networks made up of tightly knit groups of nodes performing different functions with a degree of independence are evidence of a modular view of the network (Newman, 2006). Modularity is, up to a multiplicative constant, the number of edges falling into groups minus the number expected in an equivalent network with randomly placed edges.

Diameter: The diameter of a graph is the greatest distance between any pair of vertices. First, the shortest path between each pair of vertices is calculated. Then, the greatest length of any of these paths is the diameter of the graph.

Features of the position occupied by each inventor

Degree: The degree of a node is determined by the number of connections it has to other nodes (Freeman, 1979). In other words, it assesses the degree to which a person is connected to the environment that surrounds them (Newman, 2010).

Betweenness: is the number of times a node is on the shortest route between two other nodes in the network. In other words, betweenness identifies boundary spanners or brokers, which are inventors who act as bridges along the shortest path between two other inventors who otherwise could not interact with each other. These bridges enable interaction between inventors who would not have been able to interact with each other otherwise.

Clustering: measures the likelihood that two nodes associated with a third are also associated with each other. The closing of the relationship between the three nodes expresses the opposite of betweenness. A node associated with a high clustering coefficient cannot act as a bridge and its relations do not need it to interact. More generally, the clustering coefficient quantifies the tendency of the nodes to cluster together.

Centrality: we used two measurements of inventors’ centrality in their respective networks. The first, called eigenvector centrality, was first proposed by Bonacich (1972), who assumed that the centrality of a node depended on the centrality of its direct relations in the networks. Intuitively, one becomes central by having relationships with central people in the network. The second centrality measurement, the closeness centrality, is the reciprocal of the sum of the shortest paths between the considered node and all other nodes in the graph (Scardoni and Laudanna, 2012).

Average path length: A path is the sequence of nodes to route from a source node to a target node. The length of a path is the number of nodes through the path. Hence, the average path length associated with a node is the average distance between this node and any other nodes of the network.

Eccentricity: The eccentricity of a node v in a network is the maximum distance from v to any other node. In network analysis, the reciprocal of eccentricity is used as a measure of the importance of a node within a network.

Results

Preliminary tests associated with the factor analysis

With the Spearman rank test, we first identified the variables influencing the number of patent applications filed by each organization in our sample (Table 1).

Table 1

**Results of Spearman Rank Test**

As this test does not help us to understand how these different variables combined to promote the filing of patent applications, we further conducted a factor analysis to reveal the latent dimensions associated with the variables highlighted in the previous analysis. Note that the skewness of the variables retained led us to normalize them.

We first checked that the matrix of correlation between the variables was statistically different from the identity matrix, which was tested with Bartlett’s test of sphericity. Also, the KMO measure of sampling adequacy showed that partial correlations between the pairs of variables was minimal. The results of these two tests reported in Table 2 enabled us to perform a factor analysis.

Since the variables had been normalized, we used the Kaiser criteria to define the number of factors to be retained. Table 3 below shows that only the first two factors have an eigenvalue greater than one. We therefore selected a two-factor model explaining 87.16% of the variance in the data set.

The factorial matrix (Table 4) shows the variables correlated with each of the two factors.

Main findings of the factor analysis

The factor analysis enabled us to describe the structure of the co-inventor networks as well as the features of positions of inventors in these networks for the 30 largest companies in terms of patent applicants for AI.

First finding: Identification of two latent dimensions of R&D productivity in the AI industry

After promax rotation, the variables were clearly divided between the two axes. Modularity, component number, eccentricity and diameter combined to form the first dimension. This accounted for almost 65% of the variance explained by the factor analysis (Table 5). Eigen centrality was by far the main component of the second dimension which accounted for 22.34% of the variance explained.

The factorial matrix after rotation, as well as the contributions of the variables, lead us to characterize the two latent dimensions of R&D productivity in the AI industry. The first (dominant) dimension was linked to the diversity and encapsulation of information and knowledge which barely circulated between the subgroups composing the network. We named this axis “Encapsulation of diverse information and knowledge”. The prominence of this dimension indicates the co-inventor network structure had a major influence on R&D productivity. This dimension contrasts two opposite polarities. On the one hand, in large, modularly structured networks, engineers work in teams, and information and knowledge circulate very quickly within these teams. However, the different teams have few links with each other, and knowledge and information circulate very poorly between them. Therefore, information and knowledge are mainly encapsulated inside each team. Encapsulation inside teams maintains a highly diversified information content and a rich repertoire of knowledge within the network. On the other hand, inventors’ work is not organized through fixed or structured teams and their networks are small. Information and knowledge circulate freely among individuals in the network. New knowledge is quickly shared and a situation of common knowledge among inventors in the network is attainable.

Table 3

**Eigenvalues and total variance explained**

Table 4

**Initial and rotated factorial matrices**

The second dimension referred exclusively to the ability of an organization to rapidly capture most of the information and knowledge circulating in the network. We named this axis “Information and knowledge accessibility”. This second dimension is related to the position of the inventors in the co-inventor network. Inventors’ ability to have direct relationships with the most central inventors in the network plays a pivotal role in R&D productivity. This dimension contrasts inventors having direct relationships with the most central inventors in the co-inventor networks from the other types of relationship.

Second finding: the company’s regional origin determines the structure of co‑inventor networks

The literature review has underlined the regional origin of co-inventor networks structuration (Ferrara et al., 2017; Lim & Kidokoro, 2017). All the companies in our sample have global networks of co-inventors. However, these networks can tend to be organized in different ways according to the origin of the parent company. Some may be scattered in the form of clusters working autonomously. Others may have more centralized structures, with individual units maintaining closer links with the parent company. The multiplication of comments concerning competition on AI supremacy between different regions, and notably between the U.S. and China, led us to question the structuration of co-inventor networks according to the national or regional origin of the company. Consequently, we projected the barycenter of each region or country on the factorial plan we had just constructed. The plot of the median axes of each dimension characterized the features of each observation compared to the whole sample.

Figure 2 shows the projection of the barycenter of each region or country on the factorial plan consisting of the horizontal axis (dimension 1) “encapsulation of diverse information and knowledge” and the vertical axis (dimension 2) “information and knowledge accessibility”. Similarities or dissimilarities in the features of the networks’ positions among organizations are made visible.

The U.S. companies were characterized by a high “information and knowledge accessibility” centrality of their inventors in the company co-inventor network and their median position according to “encapsulation of diverse information and knowledge”. Europe was the nearest region to the U.S. but the ability of the inventors to rapidly access to information as well as the encapsulation of diverse information and knowledge in the co-inventor networks were not as important. Korean organizations were located at the exact median of the two dimensions. The Japanese inventors presented the same “information and knowledge accessibility” as their Korean counterparts, but Japanese companies were characterized by a degree of higher diversity and encapsulation of information and knowledge of their co-inventor networks. Finally, Chinese organizations showed specific features. Inventors did not have rapid access to diversified information and knowledge. Chinese companies did not show as much information and knowledge diversity and encapsulation as other regions.

To summarize, three countries presented specific features. U.S. companies were characterized by the inventors’ quick access to information and knowledge. Japanese companies were characterized by the diversity and encapsulation of information and knowledge of their co-inventor networks. Chinese companies contrasted entirely with both US and Japanese company models.

Third finding: network structure and inventors’ localization combined to drive R&D productivity in the AI industry

The projection of the top 30 organizations engaged in AI research refined the previous macro-analysis by showing the extent to which the positional characteristics of organizations in the factorial plan reflected the macro-economic comments or assumptions stated above.

Figure 3, which shows the projection of each observation in the factorial plan we constructed, was presented for two reasons. First, we wanted to know if a single model of positioning in the factorial plan guided the optimization of the number of patent applications. In other words, was there such a thing as a magic square? In this case, the companies positioned at the top of our ranking of innovators should be located in the same quadrant of the graph. Second, we wanted to know the extent to which the national models characterized above reflected the position of the companies of each country in the networks.

Figure 2

**Projections of the countries in the factorial plan**

IBM (U.S.), Microsoft (U.S.), Toyota (Japan), Samsung (South Korea) and NEC (Japan) were the top 5 AI innovators in our sample. Figure 3 highlights the dissimilarities of their positions in the factorial plan. IBM and Microsoft were characterized by the high centralities of their inventors in the company’s co-inventor networks where information and knowledge were moderately diverse and encapsulated. Other American companies such as Alphabet or HP presented similar features. Toyota and NEC, along with Hitachi, Panasonic, Canon and Fujitsu, were characterized by the high diversity and encapsulation of information and knowledge of their co-inventor networks. Their inventors were not central in the respective companies’ co-inventor networks. It seemed that the two dimensions we first highlighted were not drawn from influential cases; the features of the networks as well as the position of the organizations in the factorial plan seemed to depend on the companies’ nations of origin. In the same way, the Chinese Academy of Sciences, the Xidian University, The Zhejian Lab and The Grid Corporation, four of the five Chinese organizations in our sample, belonged to the same quarter which characterized the position of the Chinese Barycenter in the factorial plan. Baidu was the very exception with a position akin to that of Philips.

Figure 3

**Projections of the top 30 AI innovators in the factorial plan**

However, nation of origin was not the only factor. Samsung presented a positioning in the factorial plan far from the Korean barycenter and Ricoh or Mitsubishi were closer to American companies than to Japanese ones. Intel was embedded in a network marked by high information and knowledge diversity and encapsulation, and its positioning was closer to Japanese than to American companies.

Another way to describe Figure 3 is to note that the 10 most prolific patent applicants in the AI industry were companies that made a clear choice between centrality of their inventors and co-inventor networks characterized by the diversity and encapsulation of the information and knowledge they contained. IBM, Microsoft, Samsung and Alphabet patents were borne by the most central innovators of their networks. Toyota, NEC, Fujitsu, Hitachi, Panasonic and Canon were characterized by the very diversified information and knowledge included in their co-inventor networks. Consequently, there was an obvious tradeoff between the desire to control information (centrality) and the willingness to maintain the most information-rich networks, even though access to this information might be difficult.

Nokia and, to a lesser extent, Philips and Bosch, highlighted the issues faced by European companies. Their networks were less information or knowledge diversified than their American counterparts. Occupying a central position in such networks brings fewer benefits because a central position does not provide quick access to a large panel of information or knowledge. The comparison with Siemens showed the difficult choice European companies have to make: have central inventors in poorly diversified networks or let inventors occupy a peripheral position in a network containing more diversified information.

Discussion

Examining the sources of productivity of co-inventor networks in the AI industry, our research aims to highlight the contributions of the literature on social capital for studies on R&D networks. Traditionally, the literature dealing with networks only links the value of inclusion in a network to the place occupied within it. The characteristics of intermediarity – or the broker position (Newman, 2003; Burt, 2004) or centrality – have received particular attention (Freeman, 1979) while the structural properties of the network were forgotten. More recently, attempts have been made to typify networks (e.g., Easley & Kleinberg, 2010; Girvan & Newman, 2002; Newman, 2003). These are constructed along a line measuring the strength of the links uniting network participants. In this way, networks characterized by strong ties are contrasted with networks dominated by weak ties. The theory of social capital has considerably enriched the picture. Social capital, defined as resources embedded in a social structure which are accessed and/or mobilized in purposive actions (Lin, 2017), emphasizes the structure and organization of these relationships (Coleman, 1988) as well than on the way in which individuals capture the resources embedded in relationships to generate an output (Bourdieu, 1986). The cultural dimension of social capital has also been explored in a number of articles (e.g., Tsujinaka, 2002; Ferrara & al., 2017; Lim & Kidokoro, 2017), suggesting that the structure of social relations may differ in different parts of the world.

Our study of the 30 firms filing the most patents in the AI industry shows that the structuring of social interactions guiding the production of new knowledge is organized in a two-dimensional space that follows the aforementioned two perspectives of social capital theory.

The first dimension relies on structural properties of the networks. It concerns the encapsulation of information and knowledge within clusters as in small world type networks (Watts & Strogatz, 1998) which copy the closed groups described by Bourdieu (1986), Coleman (1988) or Putnam (200). When these different clusters have few links, information and knowledge spread slowly within the network. The slow diffusion delays the arrival of a common knowledge situation where all network participants have the same information and knowledge. Indeed, the presence of distinct clusters can also create barriers to information accessibility (N’Ghauran & Autant-Bernard, 2021). For example, if a particular cluster in a modular network has limited connections or resources, it may restrict the accessibility of information to users outside of that module (Zhou et al., 2021). However, the repertoires of information and knowledge within such networks are therefore particularly rich. As social capital is rather equally distributed between the different members of a cluster, many members possess a part of the information and the knowledge available within a cluster. The effectiveness of this type of network is due to the fact that a high number of actors can contribute to innovation within a cluster. Information is disseminated, hence the use of encapsulation to capture information and knowledge. Indeed, modular networks enhance information accessibility by compartmentalizing the network and allowing for efficient routing and communication within each module. Conversely, the absence of clusters facilitates the circulation of information and knowledge among distant researchers with weak ties (Granovetter, 1973) but contributes to the standardization of both. This first dimension greatly discriminates between the companies in our sample, which therefore stands out first and foremost through the organization of their co-inventor networks. Complementing the results of Flemming, King and Juda (2007), this dimension differentiates the strategies of the most innovative companies in AI in the world but does not allow us to conclude that small world type networks favor the production of innovation.

The second dimension exemplifies the ideas of power and status of the social capital theory (Granovetter, 20218). It concerns the inventor’s position in the company’s R&D network. Our factor analysis shows that centrality is the only pivotal variable. It contrasts inventors who have direct relations with the central players in the network with other inventors as in power-law type networks (Newman, 2003). In centralized networks, social capital is possessed by a few central actors. The effectiveness of the network is due to the fact that information and knowledge are easily accessible because they are held by the central actors. The innovation ability is high for the central inventor but could be lower for the more peripheral inventors as information and knowledge are not equally shared. If the peripheral inventors cannot access the information and knowledge held by the central actors, the innovative capacity of the network can be reduced (Tsouri, 2022). Therefore, managing interactions between peripheral inventors and central inventors is essential to foster innovation in such networks. Contrary to the findings of Fleming, King and Juda (2007), intermediarity plays no role in the network of leading AI patent applicants. This result extends those recorded by Ahuja (2000) and de Vaan et al. (2015), who reported that the existence of structural holes brings no advantage in terms of innovation.

By using factor analysis, we hoped to highlight a magic square of productivity in the plan traced by the two dimensions mentioned above. However, this was not the case, and the different structures of relationships between inventors combined with their position in the networks to reveal equally effective combinations. In addition, the use of social capital theory enabled us to anchor our study of interpersonal relationships in the culture and social norms specific to each society. In this way, we have shown that the networks of American companies give priority to access and speed of circulation of information and knowledge. Inventors occupy a central place in these networks, and therefore benefit most from their structural characteristics. The networks of co-inventors in Japanese companies have wider repertoires of information and knowledge. These are not centralized, and our study shows that inventors from these companies occupy peripheral positions in co-inventor networks.

This result is contrary to that of Komoda (2021, p. 113) who investigated the software industry where US companies are highly diversified in terms of networks, whereas Japanese firms are highly centralized “because of their strong commitment to vertical integration and the full set principle”. It would seem that the parent companies of the US firms wish to control AI developments and projects[3], while the Japanese firms have created clusters in their various units to counter Google and Microsoft, and thus promote diversity of approaches. Hypothetically, this would suggest that as far as innovation in AI is concerned, Japanese firms are moving away from their historical logic of conglomerate integration.

Chinese companies present a very intriguing profile. While this country is often cited as the most advanced concerning AI technology, Chinese companies present the least information and knowledge diversified networks. Moreover, their inventors occupy a peripheral position in these networks. Consequently, Chinese companies underperform in terms of the number of patent applications they file. This last observation does not fit well with the observed status of Chinese companies in the global AI industry. It is possible that Chinese companies rely more on secrecy than on patent applications filings to protect their innovations (Aharonson et al., 2007). Furthermore, the People’s Republic of China is still a planned economy where the R&D of companies or universities are mainly financed by the State, which sets the priorities and to which one is accountable. This centralized organization is the opposite of a network having a decentralized institutional form (Williamson, 1985). The poor development of innovators’ networks in China may result from this opposition. Moreover, successive five-year plans show how R&D financing is conditional on the sharing of research findings with other companies. In this context, patent application filings signal that you have something to share with your competitors and may explain why Chinese companies or universities could opt for secrecy over patents. Our results also confirm the findings of Sun et al. (2022) that Chinese firms lack information diversity and encapsulation to enhance the creativity and productivity of their inventors.

Conclusion

Capital social is an interactional resource for an individual inventor; its structure and nature have important impacts on the efficiency of a co-inventor network. In innovation networks, social capital represents the possibility to access to information and knowledge possessed by other inventors. The distribution of social capital impacts the accessibility of information and knowledge and therefore the efficiency of the social network. Our results highlight two typical organizations of social capital within the inventor networks of the world’s most innovative companies in the AI sector. In addition, the development of AI opens a new era where no country or region may pretend to dominate the industry. This situation contrasts with the past when Silicon Valley exemplified how efficient R&D should be conducted. The R&D in the AI industry allows us to observe how co-inventors’ networks compete in a new high-tech sector.

This study contributes to knowledge on co-inventor’s networks in three ways.

Firstly, it links the two latent dimensions of R&D productivity in the AI industry to the social capital theory. We have tried to show how the social capital theory may enrich our understanding of R&D networks. The first dimension is “Encapsulation of diverse information and knowledge”. The second dimension is “Information and knowledge accessibility”. The first dimension relies on the structure of the network, while the second reflects the position of innovators in the network. This result highlights a dilemma faced by the organizations in our sample: they must structure networks including the widest variety of information encapsulated in competing R&D projects. Also, these companies must be able to gather such information rapidly. This result is innovative, as the literature has shown that companies should be able to quickly access information about external innovation in order to periodically reassess their own R&D interests (Toth & Lengyel, 2021), but has not focused on the links between network structure and access to internally available information between different R&D teams.

Secondly, competition between firms reveals regional specificities in terms of intra-firm co-inventor networks (Lim & Kidokoro, 2017). Our study reveals that companies optimizing the number of patent applications they file make a clear choice when faced with the dilemma between the two latent dimensions of the R&D productivity. American and Japanese companies are the most productive in terms of patents, even if the features of the networks’ positions among organizations are different, which shows that there are several optimal strategies, the choice of which depending on the company’s culture. U.S. inventors are crucial to firm R&D networks and have a median information encapsulation and diversity position. Japanese innovators are central like Korean inventors, but Japanese corporations have more network variety and information encapsulation.

Thirdly, our results enrich the debate on connectivity and brokerage (Aral, 2016; Toth & Lengyel, 2021; Komoda et al., 2021) by highlighting the fact that network characteristics are just as important as a person’s position within it. To our knowledge, this result has never been highlighted in the academic literature.

In relation to management implications, this study reveals that there is no prevailing model in terms of co-inventor network structures within the AI business. Instead, each model has its own advantages and disadvantages. A central position is beneficial for the fast and easy gathering of information located in the network. Being embedded in a diversified network with encapsulated information may favor explorative innovation as it offers access to non-redundant and varied knowledge. A highly diversified network with loose connections may not promote trust or common standards (Obstfeld, 2005). In any case, a clear positioning between these two alternatives seems to be a successful strategy, as it is the stance evidently chosen by the most innovative companies in the AI sector.

As we also show that a company’s regional origin determines the structure of its co-inventor networks, firms must integrate this contextual specificity to increase the number of patents they obtain in the field of AI.

Despite these significant results, this research has few limitations which generate avenues for future research. Our results show that the size of social capital determines the ability to connect with many inventors and share information with them and the nature of social interactions determines the ability to join research communities and access a wide range of knowledge and expertise. Further research could study the strength of ties existing within co-inventor networks as well as the boundary conditions favoring collaborations. Drawing on patents to measure innovation does not appropriately reflect the strategies implemented by companies protecting their inventions in other ways. Chinese companies whose strategies seem intriguing may not be accurately positioned in their innovation networks when studied according to the present patent analysis. Patents capture only technological and codified knowledge, to the exclusion of tacit or secret knowledge. Our research did not specifically differentiate academic and non-academic organizations. Future research could investigate the impacts on the structures of the co-inventor networks of collaborations between these two kinds of organizations.

Abstract

Résumé

Resumen

Literature Review