Floridi and Cowls: 47 AI ethics principles reduce to five

The convergence Floridi and Cowls found did not happen in a vacuum. It followed a burst of activity between 2016 and 2019, when organizations ranging from the Future of Life Institute to the OECD, Japan's government and the European Commission all published AI ethics principles. Their paper, published in 2019, asked whether this proliferation was redundant and concluded that it largely was not: the principles varied in scope, but their core overlapped.

The overlap has a lineage. Four of the five principles — beneficence, non-maleficence, autonomy and justice — are the same as those long used in medical ethics. That suggests AI ethics borrowed its scaffolding from an earlier field rather than inventing it from scratch. The fifth, explicability, goes beyond medical ethics and acts as an enabling condition: without it, the other four cannot be checked in practice.

Scope still differs. The Asilomar principles extend to long-term risks such as superintelligence, while the UNESCO recommendation, adopted in 2021, reaches into environmental and cultural diversity. The focus has since shifted from drafting principles to embedding them in organizations and rules — the subject of AI governance.