A Costly Emergencies Approach to Estimating Costs for Artificial Intelligence Review

Abstract

Artificial intelligence (AI) research is increasingly considered to be a risky endeavor. The present prevailing model of AI risk management imports the commitments of research ethics, in terms of principles, approach to risk assessment, mitigation, and management. Despite the history and appeal of the research ethics model, alternative models for evaluation of AI risk should be entertained before adopting this approach. Here we propose a hypothetical applied information economics approach that changes the model for risk review from ex ante hypotheticals paid for as a fixed cost borne by all to variable costs that could be calculated in response to consumers’ willingness to pay for risky technology or to be paid for by companies willing to shoulder the cost of risks. Some of these risks will be of such magnitude that mitigating them constitutes an existential emergency for which an ex-ante research review approach may be justifiable. But, for risks that are less than non-remediable, existential, threats, a more cost sensitive approach may represent a more equitable path that shifts costs to those willing to bear them. We approach thinking through AI risks as something with specific costs we can estimate using known risks from similar activities of daily life.