Theoretical analysis reveals structural limits in diagnosing exploitation across multi-party AI systems, highlighting the necessity of mandatory disclosure and precise entity boundaries.
Key Points
To establish a formal structural framework for evaluating multi-party economic relations and diagnoses of exploitation in artificial intelligence systems.
Formulated a four-role structural framework defining control, action, capability decline, and benefit extraction across up to twenty directed pairs of economic entities.
Conducted structural proofs analyzing topology equivalence, entity aggregation and merge admissibility, audit opacity under incomplete data, and reachability across finite time horizons.
Proved that network topologies with identical node-level quantitative metrics can yield contradictory conclusions about exploitation, showing directed pair models are structurally insufficient.
Established that merging entities alters action attribution and benefit paths, meaning an aggregation is formally valid only if no witness edges cross the merge boundary.
Showed that unobservable training corpora and counterfactuals leave audits indeterminate, proving that disclosure is a strict prerequisite for structural diagnosis rather than an elective policy.