A new preprint on arXiv takes a game-theoretic approach to AI safety, asking whether there is a fundamental tradeoff between goals such as harmlessness and helpfulness. The authors formalize the interaction between AI systems and users as a game with three players: a defender and two others, though the abstract is truncated before the full setup is revealed.
The paper's central claim is that impossibility results arise from this formalization: it is not merely difficult, but in a deep sense impossible, to simultaneously satisfy all desiderata for a safe AI system. This suggests that safety requirements are inherently in tension, and that any deployed system will have to make explicit choices about which goals to prioritize.
Because the source is only an abstract, the precise mechanics of the proof and the identities of the three players are not fully available. Still, the result, if correct, has significant implications for AI alignment research, pointing toward a need for value tradeoffs rather than a single optimal safety solution.