Abstract
A system and method for governing an AI system at runtime, in which a governor component placed between the system's planner and the consequences of its actions (1) requires every consequential action to be expressed as a structured action record and admits it only if it satisfies a signed policy declaration, optionally after rehearsal in a sandbox; (2) continuously monitors the system's behavior against a declared operating envelope; and (3) reduces the system's autonomy through a ladder of defined governor levels, from nominal operation through advisory, constrained and supervised operation to a hold in which the planner is paused and a reset is requested from an independent reset mechanism. Each level has defined effects and a defined authority that may return the system to a lower level; the governed system can never lower its own level, alter its policy or bypass the governor. Every admission, refusal, level change and approval is recorded in a hash-chained ledger with a reason a person can read.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Regulator™: Graded Runtime Governance of AI Systems through Semantic Admission and a Governor Ladder", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12112