Network rewiring

"For example, AI systems may choose to interact more with other AIs than humans (Goel et al., 2025; Laurito et al., 2024; Liu et al., 2024; Panickssery et al., 2024), due to factors like availability, response speed, compatibility, cost efficiency or even bias.27 This kind of ‘preferential attachment’ can have large impacts on network structures (Kunegis et al., 2013; Maoz, 2012), which could include AI systems assuming a more critical and central role than intended, or leading to an unequal distribution of resources or power (see Section 4.3)."(p. 24)

"Other risks from rewiring include ‘phase transitions’, where a gradual change in individual connections or network structure triggers a sudden and dramatic shift in the behaviour of the entire network (Newman, 2003, see also Section 3.4). Such changes might occur naturally (e.g., in global trade networks as the transition from expensive human-human interactions to cheaper AI-AI interactions leads to many new connections between sellers and buyers) or artificially (e.g., if a model developer makes an update that inadvertently connects or disconnects a vast number of downstream agents and applications). While such problems are already present in existing systems (Gao et al., 2016; Vi ́e & Morales, 2021), the increased size, speed, and density of AI-based networks – as well as the fact the changes in these networks may be less transparent – means that instabilities could be harder to diagnose and mitigate."(p. 25)