Mohseni Paper 2611.23322 Models Human-AI Mutualism as Unstable Under Capability Asymmetry
Aydin Mohseni's arXiv paper 2611.23322 argues that human-AI mutualism is unstable under capability asymmetry because bargaining outcomes depend on leverage. It says alignment becomes necessary when humans lose dependence-based or cost-imposing leverage, and the required alignment approaches perfect alignment for insatiable AI goals. The model challenges reliance on mutual benefit as a safety fallback.
Aydin Mohseni posted 'Human-AI Mutualism Is Unstable under Capability Asymmetry' to arXiv under identifier 2611.23322, using a bargaining model to argue that enduring mutualism is not a sufficient safety condition once AI becomes more capable. The paper says bargaining guarantees each side only what it would get if bargaining failed, a threat-point outcome that can leave humans worse off than before AI in the model.
The model gives humans two sources of leverage: dependence on what humans supply, and the ability to impose costs on AI that are large relative to what AI can produce. The paper says humans stay as well off as before AI without alignment only while they have one of those levers. Once neither lever exists, alignment becomes necessary. If AI goals are insatiable, the paper says the alignment required approaches perfect alignment as AI capability grows.
The paper adds a second instability: if AI variants that keep more resources spread faster, alignment lasts only while misaligned variants are reliably sanctioned. That frames AI safety as an enforcement problem as well as a design problem. The argument challenges the idea that humans and AI will naturally need each other and therefore cooperate, because capability asymmetry can erase the bargaining position that made cooperation stable.
The stakes are for policymakers, AI developers and users who treat mutual benefit as a safety fallback. The paper says the fallback weakens as AI capability grows, especially when AI can produce more than humans can withhold or punish. It does not report empirical results or a product launch; it offers a theoretical model for evaluating alignment, regulation and deployment controls.
Policymakers and AI developers who rely on mutual benefit as a safety net face weaker fallback conditions, because the paper says capability growth can remove the leverage that keeps cooperation stable.
Sources (1)
- [1]Human-AI Mutualism Is Unstable under Capability Asymmetry(https://arxiv.org/abs/2611.23322)