An organisation can move from the middle of a network to its edge and back again. Does that movement make it more innovative—or merely record that its collaborations are changing? Chongfeng Wang, Rui Sun and Jie Xu’s new study asks a useful dynamic question about AI patent networks. Its answer is suggestive, but less causal than the language of organisational adaptability may imply.
The paper introduces brokerage reversal. A broker connects organisations that are not otherwise connected, potentially gaining access to different pools of knowledge. Reversal occurs when an organisation’s movement towards or away from such a position changes direction across successive years. Instead of asking whether brokerage is always beneficial, the authors investigate whether repeated changes in network position are associated with later patent output—and whether the organisation’s internal inventor network alters that association.
A rare attempt to connect outside and inside
The design is ambitious. The authors use the US Patent and Trademark Office’s Artificial Intelligence Patent Dataset, covering 1976–2023. They construct an interorganisational network from joint patent applicants, an internal collaboration network from co-inventors within each organisation, and an internal knowledge network from patent-classification combinations. Their final fixed-effects sample contains 23,632 organisation-year observations for 3,456 organisations.
This multilevel construction is the paper’s strongest feature. External position is not treated as automatically productive. Knowledge encountered through changing partnerships must also be integrated internally. The authors therefore test whether cohesion and centralisation within inventor networks change the association between brokerage reversal and the number of AI patents produced.
The reported pattern is clear. More brokerage reversal is positively associated with patent output. Yet the association becomes weaker as internal inventor networks become more cohesive or centralised. The authors interpret this as evidence that tightly connected or concentrated internal structures can constrain the flexibility needed to exploit knowledge arriving through changing external relationships.
Several checks increase confidence that this is not an artefact of one specification. The paper uses organisation and year fixed effects, negative-binomial models suited to count outcomes, an alternative measure of network variation, a four-year rather than five-year window, and a one-year lag. The signs of the principal coefficients remain stable. The authors also state the sample construction unusually clearly, including the large reduction from 75,589 identified organisations to the 3,456 with enough usable observations for fixed-effects estimation.
What the network measures—and what it cannot
The main caution is that the study observes patent records, not knowledge moving through organisations. A joint patent application is a defensible indicator of collaboration, but it cannot show who shared what, whether trust changed, or how external knowledge was interpreted and used. The proposed sequence—access, acquisition and internal integration—is theoretically plausible rather than directly measured. The authors acknowledge this.
Patent counts also reward quantity, not necessarily quality, public value or commercial significance. One organisation may file many incremental patents while another produces fewer but more consequential inventions. Citation-weighted patents, breakthrough measures, product outcomes or independent assessments of social value could alter the picture.
There is a deeper identification problem. Organisations that already possess stronger research resources may patent more, form more ties and change network position more often. The one-year-lagged analysis improves temporal ordering and removes direct overlap between the focal outcome year and the reversal measure, but it cannot eliminate time-varying differences in research expenditure, staff, strategy or acquisition activity. Comparable organisation-year data on research inputs are absent from the patent dataset.
The network itself contains further approximations. Repeated collaborations within a window are reduced to one unweighted tie. Parent companies and subsidiaries remain separate when their identifiers differ, as do organisations before and after mergers or name changes. Apparent reversal can therefore reflect database boundaries, restructuring or sparse patent activity as well as strategic movement. The dramatic sample reduction also means the estimates describe organisations with sufficiently sustained and observable patenting, not the full field of AI innovation.
These limits matter for the practical recommendation. It would be premature to tell managers to reduce cohesion or centralisation. Dense internal ties can support trust, coordination and reliable execution even if they weaken this particular statistical association. The result is better read as a warning against assuming that external variety will automatically become internal capability.
An AE reading: change is not the same as renewal
My autopoietic ecology (AE) interpretation starts from the difference between changing position and renewing a capacity. The AE overview asks what makes an activity possible now and what that activity does to the conditions of future action. Brokerage reversal records movement in a collaboration network. It does not by itself establish that an organisation can interpret, retain or responsibly use what becomes available through that movement.
The study’s external–internal distinction nevertheless fits an important AE question. A collaboration can expose an organisation to unfamiliar knowledge while leaving the people, time, authority and interpretive practices needed to use it unchanged. Conversely, a cohesive internal network may preserve established competence while filtering out ideas that do not fit current categories. The issue is not whether openness or cohesion is universally good, but how one relationship changes the conditions under which later work can continue.
The AE Concept Guide also directs attention to semantic transduction: what happens as something moves between settings. Patent classifications, joint applications, inventor ties and managerial claims are not interchangeable representations of the same knowledge. A useful follow-up would trace particular ideas from an external collaboration into project discussions, design decisions, staffing, later patents and abandoned work. That would show where translation succeeds, where compensatory labour is required and whose judgement becomes authoritative.
This is an EdgeLab extension, not a finding proved by the paper. A conventional alternative is that absorptive-capacity theory and dynamic network analysis already explain the observed pattern: organisations with suitable routines can exploit changing ties, while concentrated structures create bottlenecks. AE adds value only if following future capacity, burden transfer and changes of meaning reveals something the established account misses.
Evidence that could change the interpretation
A stronger test would link network histories to research spending, personnel changes, mergers, project records and interviews with inventors. It would distinguish deliberate repositioning from disruption, and repeated collaboration from a single joint filing. Patent quality and non-patented innovation should be measured alongside counts. Comparison across countries and technological fields would show how much the result depends on US patent institutions and the peculiar pace of AI.
Such evidence could also challenge the AE reading. If brokerage reversal continued to predict consequential innovation after detailed controls, while measures of translation work and future organisational capacity added nothing, the ecological extension would be unnecessary. If similar network movements produced different outcomes depending on who could interpret new knowledge, repair coordination failures and revise internal authority, its contribution would become clearer.
Wang, Sun and Xu make an important methodological move: they replace a static snapshot of brokerage with a history of changing positions, and connect external structure to internal organisation. The paper does not show that movement causes innovation, or that less cohesive organisations are simply better. Its more durable lesson is subtler. Access to difference becomes valuable only through arrangements capable of making something of it.
Reference: Wang, C., Sun, R., and Xu, J. (2026). Brokerage Reversal in Interorganizational Collaboration Networks and Organizations’ Technological Innovation Output: The Moderating Role of Intraorganizational Collaboration Networks. Systems, 14(9), 1182. Published 20 September 2026. https://doi.org/10.3390/systems14091182.
EdgeLab is an independently organised research network. It is not formally recognised as a research group, centre or unit by the Faculty of Education or the University of Cambridge. Institutional affiliations refer to individual contributors and do not imply institutional endorsement.
Leave a Reply