A Trilogy of AI Safety Frameworks: Paths from Facts and Knowledge Gaps to Reliable Predictions and New Knowledge

TL;DR

This paper proposes a trilogy of AI safety frameworks focusing on facts, knowledge gaps, and reliable predictions, aiming to address immediate safety concerns without hindering AI innovation, supported by case studies in biomedical science.

Contribution

It introduces three practical AI safety frameworks that are tractable and applicable to critical domains, emphasizing short-term improvements based on existing case studies.

Findings

Proofs of concept in biomedical applications

Frameworks address safety without hindering innovation

Case studies demonstrate practical applicability

Abstract

AI Safety has become a vital front-line concern of many scientists within and outside the AI community. There are many immediate and long term anticipated risks that range from existential risk to human existence to deep fakes and bias in machine learning systems [1-5]. In this paper, we reduce the full scope and immense complexity of AI safety concerns to a trilogy of three important but tractable opportunities for advances that have the short-term potential to improve AI safety and reliability without reducing AI innovation in critical domains. In this perspective, we discuss this vision based on several case studies that already produced proofs of concept in critical ML applications in biomedical science.