Homogeneous AI Traders, Endogenous Liquidity, and Phase Reversals in Tail Risk

As machine learning tooling diffuses across asset managers, trading stacks can become homogeneous: agents rely on similar signals, similar online update rules, and similar risk controls. Homogeneity mechanically increases correlated exposures (crowding), yet modern stacks also enforce shared constraints (volatility targeting and leverage caps) that can synchronously reduce risk when volatility spikes. We study when increasing homogeneity amplifies versus attenuates tail losses in an agent-based market with three coupled ingredients: (i) adaptive trading rules learned from P&L, (ii) endogenous liquidity where depth declines with aggregate order flow and volatility, and (iii) ubiquitous constraint gating. Across a large sweep over impact strength and learning speed, homogeneity almost always increases crowding, but its effect on 1% expected shortfall of losses is non-monotone. In a high-impact/fast-learning regime, synchronized volatility spikes activate constraints system-wide and mechanically truncate exposures, reversing the ES slope despite higher crowding. We map the phase boundary, interpret it via a weak/strong-impact × slow/fast-learning mechanism decomposition, and then decompose homogeneity into three channels—shared signals, shared updates, and shared risk models. Shared signals primarily generate persistent high-stress episodes, shared updates primarily increase stress-entry frequency (regime switching), and shared risk models alone has comparatively muted systemic effects. The results imply that “AI homogenization” is not a scalar risk factor: its systemic impact depends on which components are shared and on liquidity endogeneity.