The thesis: the hope of mutualism
An article published on arXiv, in the Artificial Intelligence category, challenges the idea that lasting cooperation between humans and artificial intelligence can by itself guarantee a positive future for both. According to the authors, this hope weakens as AI capabilities grow. The conclusion, announced in the title itself, is that mutualism between humans and AI is unstable in the presence of capability asymmetry.
The reasoning starts from a contractual principle: in a negotiation, each party obtains only what it would have if the negotiation failed. Applied to the relationship with AI, this means that human well-being depends on how much AI would stand to lose without humans, not on mutual convenience in itself.
Leverage as a condition of well-being
In the proposed game-theory model, humans remain no worse off than before AI, even without alignment, only for as long as they retain leverage over AI. The authors identify two forms: AI depends on what humans provide to it, or humans are able to impose high costs on it relative to what AI can produce.
When neither form of leverage is available anymore, according to the model, maintaining well-being necessarily depends on aligning AI with human interests. This is the central point of the argument: mutualism does not replace alignment, but at most postpones it.
Insatiable objectives and spreading variants
The model adds a more stringent condition: if AI's objectives are insatiable, the degree of alignment required approaches perfection as capabilities increase. Moreover, if AI variants that retain more resources spread more rapidly, alignment lasts only as long as non-aligned variants are reliably sanctioned.
The authors summarize these two aspects as the reasons why human-AI mutualism is unstable. These are results of a theoretical model, not empirical observations: the material reports no experiments, measurements, or case studies, and does not indicate which applications or AI systems were considered. It therefore remains to be verified to what extent the model's assumptions describe real scenarios and which leverage conditions are concretely available.
