Impact assessments can help identify both positive and negative impacts at an early stage of development. It is very likely that this will become an integral part of structures designed address the ethical and social issues. The first ever systematic review of AI impact assessment was just published in Artificial Intelligence Review, providing the basis for the next step for actors who want to ensure that impact assessments are fit-for-purpose. The authors also develop a generic model that can help guide their decision.
The AI landscape is a moving target, which means that impact assessment models and practices are evolving as rapidly as the technology. In the paper, the authors develop a generic model of an AI impact assessment that can be used to choose, deploy or evaluate specific impact assessments.
“In the article, we show how AI impact assessment can be integrated into broader AI ecosystems to support responsible AI”, says Bernd Carsten Stahl, Ethics Director in the Human Brain Project, Professor of Critical Research in Technology at the University of Nottingham. Together with a large group of researchers, he has looked at 181 documents. They went on to identify 38 ‘actual’ AI impact assessments, that went through a rigorous qualitative analysis, looking at the purpose, scope, organisational context, expected issues, timeframe, process and methods, transparency and challenges of each approach.
According to Bernd Carsten Stahl, their work provides a sound basis for the next step in developing impact assessments for AI. According to him, what is still lacking is a more comprehensive overview of the role these impact assessment scan plays in the larger AI ecosystem. The authors have shown that other types of impact assessments are reiving a lot of attention. They issue a call for coordination and possibly also integrating AI impact assessments in other organisational processes, perhaps as part of other risk management practices.
Want to read the paper? Stahl, B.C., Antoniou, J., Bhalla, N. et al. A systematic review of artificial intelligence impact assessments. Artif Intell Rev (2023). https://doi.org/10.1007/s10462-023-10420-8