A new paper on arXiv examines a growing problem in AI services: providers have both the opportunity and the incentive to quietly substitute a cheaper, lower-quality model for the one a client paid for. Because model outputs are inherently stochastic, clients often cannot tell whether a poor result came from a bad model or just normal randomness, making the degradation hard to detect and even harder to prove.

The authors frame this as a moral hazard problem and turn to contract theory for a solution. Their proposed mechanism, which they call "anytime contracts," appears designed to give clients a way to verify or enforce quality without needing to perfectly observe the provider's internal choices. The abstract does not spell out the full mechanics, but the title and framing suggest the contract can be invoked at any time, rather than only at fixed checkpoints.

This is a conceptual proposal rather than a deployed system, and the abstract gives only a high-level sketch. Still, the paper points to a practical concern: as AI services become commodities, contractual safeguards may be as important as model quality itself. The authors do not claim to solve all forms of provider misbehavior, but they offer a structured way to think about accountability in an opaque market. [1]