Regulating for the responsible development of artificial intelligence

A key variable affecting firms' propensity for collaboration is the likelihood that their pre-existing technological knowledge will be lost to a competitor. Nueno and Oosterveld find as follows: "Companies are concerned with the potential unplanned loss of knowledge through coalitions. It is difficult to control what exactly goes on in the many meetings between scientists from the different companies involved in a common project." (1998: 13) Even if the scope of the RJV were restricted only to AI safety research, there could nevertheless be high spillovers of non-safety-related research (as it is impossible to fully separate the different areas). This problem could increase the "transaction costs" of collaboration (Hennart, 2008). Legal norms have the potential to affect the levels of spillover within an RJV. In particular, intellectual property (IP) protection can allow parties to main-tain control of their existing technical knowledge (Teece, 1987; Parker, 2007). For example, the collaborating firms can agree to assign or license their back-ground IP rights to the RJV entity and not to the competitor firm (Parker, 2007). Therefore, the IP regimes of key markets (such as the US and the EU) can shape the collaboration game, by determining the proportion of a firm's technical knowledge that it can protect despite collaboration. Within this issue, an important question is whether AI research is eligible for patenting. The US Supreme Court, in the context of computing patents, has shown an unwilling-ness to allow patents in relation to abstract ideas or mathematical principles.3 Research is needed to consider how legal decisions like these shape the incentives for AI firms to collaborate. We could also address the non-typical way in which intellectual property is handled within AI research. Second, the firms may not form a grand coalition RJV, but may instead splinter
into multiple RJVs. Again, the legal framework is relevant here: both the EU and the US impose restrictions on the combined market share of RJV collaborators.5 Additionally, there may be incentives to exclude some rival firms from
the collaboration. As Greenlee points out, firms "benefit most from cost reductions that are not shared with rivals. Having many partners that benefit from a firm's research effort, then, can dampen incentives to invest in R&D." (2005:
357) Nonetheless, the traditional view is complicated by the fact that a subset AI firms appear to prioritize moral or safety concerns over cost-reduction. Finally, existing research has emphasised that firms are more likely to collab-
orate if they trust one another (Hill, 1990). This trust is built through past interaction (Gulati, 1995) and through decision-makers being embedded in the same wider networks (Dacin et al, 2008). Dacin et al argue that "Economic, cog-
nitive, and socio-cultural forces simultaneously impact the form and evolution of alliance dynamics and alliance activity" (2008: 100). In this respect, Nooteboom emphasises the importance of "culture", i.e. where firms develop "their
own specialized semiotic systems, in language, symbols, metaphors, myths, and rituals." (2008: 610) Therefore, research is needed to uncover the relationships that exist between US firms (such as Google) and Chinese firms (such as Baidu). The concern would be that (a) these firms do not sufficiently trust one another; (b) the firms' decision-makers do not share networks (c) there is a stark culture divide. The three problem areas all relate to the conditions under which firms would decide to collaborate. In terms of methodology, much of the literature in this area relies on game theoretic models, with firms being motivated by purely financial considerations (for a review, see Li and Nguyen, 2017). This approach can be restrictive in that, as Li et al acknowledge, "other factors may influence the innovation decision such as culture, psychology, and the support of their staff...'.