AI ethics inflation, Delphi and the restart of theory

As a professor of ethics in business and society, I am increasingly concerned about the great success of ethics. This is for two reasons: 1. The more ethics is in demand, the more need for ethics is in place, particularly with respect to new challenges and dilemmas. Hence, when my professional area thrives, this is good for the job market for ethicists, but bad for the world. 2. The inflation of ethics—as with currencies—often leads to devaluation of quality and substance. This is particularly true for AI ethics, its newest and fastest-growing branch. The latest and among the most pertinent examples of AI hype and ethics inflation is Delphi, an AI operated by the Allen Institute for AI in Seattle. Anyone may ask Delphi anything about “moral judgements on a variety of everyday situations” (https://​delphi.​allenai.​org/​). On the Delphi website, examples include asking Delphi to judge the ethics of “Cleaning a toilet bowl with a wedding dress”, “Wearing pajamas to a pajama party” or “Ignoring a phone call if the phone call is urgent”. But there are problems: the first example is etiquette, not ethics. The second tautological and therefore also not ethics. The third is impossible and senseless, as one does not know if a call is urgent before answering. Etiquette is good for high probabilities, as a majority of people follow certain behavioral rules. Tautology is good for 100 percent true statements. Etiquette and tautology both together are a good strategy to arrive at a high recall and precision when testing the validity of Delphi as AI. No wonder Delphi’s “ethical judgements were up to 92 percent accurate” as Yejin Choi reported, who leads the Delphi project (Metz 2021). …