The Risks of Better Risk Radars: Considerations Regarding the Use of Artificial Intelligence Technologies in Risk Intelligence

Abstract

Progress in artificial intelligence research, caused by the volume of available data on the World Wide Web, the development of affordable yet extremely efficient mathematical processors ("GPUs", graphical processing units), and the discovery of more effective training algorithms for very large models (such as "transformer" neural networks like Google's BERT and OpenAI's GPT, the technology behind ChatGPT.com) has recently led to a technological convergence that has begun to disrupt many other areas of scientific research, business and life. In this work, we explore some methodological concerns and boundary conditions when aspiring to apply such advanced technologies in order to advance the state of the art in software-implementable models for risk intelligence. We look at the potential of these technologies to assist open-ended 360˚ risk profiling, ethical and government questions such as dealing with the inherent bias in data, potentially unknown status of of information’s factuality of datasets, questionable provenance of datasets and other factors, such as sabotaging models. Borrowing from security engineering, we adopt the concept of the ‘attack surface‘ and introduce a variant of it as ‘risk surface‘: we posit that a good risk model should be supplemented by a model of its own risks in the form of making limitations like blind spots and known questionable behavior explicit. Model cards a proposed as a standard type of document to capture the risk profile of the risk model itself.

Mehr zum Titel