Digital platforms built on user-generated content face a new kind of shock: generative AI substitutes directly for the interactions that give them value. Some have collapsed in the two years since ChatGPT's release while others in adjacent domains have remained stable. This heterogeneity-across platforms, and across user tiers within a platform-is not explained by accounts organized around AI's capability in different task domains. We develop an equilibrium theory in which fragility is governed by a structural feature of design: the degree to which value flows directionally, from higher-to lower-skilled users, rather than laterally among peers. The cascade threshold separating clean truncation from full collapse is strictly decreasing in this directionality, so more directional platforms collapse at lower AI reach. A second and more distinctive prediction concerns survivors: when the shock truncates a platform from below without collapsing it, the share of value carried by the peer channel rises for every remaining high-skill participant-a compositional reorganization that follows from the geometry of the matching technology rather than the model's asymmetric weights, and that a capability-only account cannot generate. A transient-overestimation mechanism produces hysteresis: a platform pushed into collapse by inflated early beliefs about AI does not recover when beliefs correct, because the empty platform is itself absorbing, and recovery requires coordinated re-entry of the whole surviving cohort. The theory yields falsifiable predictions that separate the directionality channel from the capability channel through within-platform tier variation, off-diagonal cross-platform comparisons, and compositional shifts in engagement.
NetSI authors
