In a recent post, we discussed the need for policymakers to think of AI and Autonomous Systems (AI/AS) always needing varying degrees of the human role (“hybrid” human/machine systems). Understanding the potential and limitations of combining technologies and humans is important for realistic policymaking. A key element, along with accurate forecasts of the changes in technology, is the safety of AI/AS-Human products as discussed in the IEEE report “Ethically Aligned Design”, which is subtitled “A Vision for Prioritizing Human Wellbeing with Artificial Intelligence and Autonomous Systems”, and Ben Shneiderman’s excellent summary and comments on the report as well as the YouTube video of his Turing Institute Lecture on “Algorithmic Accountability: Design for Safety”.
In Shneiderman’s proposal for a National Algorithms Safety Board, he writes “What might help are traditional forms of independent oversight that use knowledgeable people who have powerful tools to anticipate, monitor, and retrospectively review operations of vital national services. The three forms of independent oversight that have been used in the past by industry and governments—planning oversight, continuous monitoring by knowledgeable review boards using advanced software, and a retrospective analysis of disasters—provide guidance for responsible technology leaders and concerned policy makers. Considering all three forms of oversight could lead to policies that prevent inadequate designs, biased outcomes, or criminal actions.”
Efforts to provide “safety by design” include work at Google on Human-Centered Machine Learning and a general “human-centered approach that foregrounds responsible AI practices and products that work well for all people and contexts. These values of responsible and inclusive AI are at the core of the AutoML suite of machine learning products …”
Further work is needed to systemize and enforce good practices in human-centered AI design and development, including algorithmic transparency and guidance for selection of unbiased data used in machine learning systems.