Minimum-cut bottleneck at the human–AI interface

Establish whether the collective performance of a hybrid human–AI network is bounded by the capacity of the lowest-capacity cut separating the human and AI sub-communities, implying that the human–AI interface acts as the system’s performance bottleneck.

Background

Using an information-theoretic framing, the paper notes that communication channels have capacity and noise, and that a network’s collective computation time is constrained by conductance, with bottlenecks determined by minimum cuts. The authors argue that these limits bite hardest at the human–AI interface due to translation, hallucination, and persona drift.

They conjecture that hybrid networks are often bottlenecked at this interface and that overall collective performance is bounded by the lowest-capacity cut across the human–AI divide, suggesting a fundamental limit independent of individual agent capabilities.

References

A hybrid network is therefore often bottlenecked not within its human or its AI sub-community, but when messages are passed between them, and we conjecture that its collective performance is bound by that lowest-capacity cut (Liu et al. 2024).

— Collective Cognition in Hybrid Groups: A Network Science Synthesis  (2607.05593 - Hemmatian et al., 6 Jul 2026) in Section 4.4, The message