Responsible Artificial Intelligence Governance Frameworks for Trustworthy Enterprise Decision-Making

The convergence of enterprise artificial-intelligence governance is around three points: Most importantly, tier systems by their risk level, then implement proportionate controls, and finally document it in an attestation. This chapter concludes that the architecture has three points at which it fails to meet measurable objectives, and quantifies each. A portfolio simulation of 40 cases demonstrates that a categorical tiering introduces 14.5 to 24.7 per cent more residual risk, compared to a continuous allocation of the same set of assurance budget, as in the case of the range-compression defect for risk matrices. Tail risk increases by 84 per cent as dependency concentration increases at constant budget, a level of exposure that is not covered by any per-system control in enterprise systems, due to the fact that all of these systems share foundation models, vendors and data pipelines. a pass/fail attestation conveys only 0.069 bits of mutual information, marginally reducing the posterior probability of non-compliance from 0.150 to 0.108.

Hali tarjima qilinmagan