Design For Trustworthy AI Solutions

Anonymous

Keywords: Trustworthy AI, Data Interpretation, Conformity Assessment, Decision Intelligence, Bias Parallel, Data Imbalance, Data Risk Assessment, Sensitivity Loss, Algorithmic Assessment, Concept Test, Adversarial AI, What-if-Analysis, Counterfactuals, Explainability

TL;DR: Design For Trust - A framework for developing trustworthy, reliable and unbiased AI solutions by developing "Data for Trust", "Algorithm For Trust" and "Decision Intelligence"

Abstract: Transparency of an AI solution is the need of the hour. With growing adoption, AI is increasingly making business critical decisions in organizations and propagating it, not only limited to organizations but also to society. This has resulted in growing legislative asks from organizations such as "US Algorithmic Accountability Act -2019", " EU Right To Explainability", "EU AI act 2021" and so on. We hereby propose a framework for addressing transparency in AI solution and bringing about trustworthiness, reliability and un-biasness of AI solutions to various stakeholders, which may include but not limited to, AI Solution Engineers, Chief Legal Council, Decision Reviewers etc. Our solution addresses the problem via providing transparency in terms of Data Interpretation, where we use AI to spot historical bias, mitigate them and perform risk assessment of the data; Conformity assessment, where we test trustworthiness, robustness and explainability of AI algorithm; Decision Intelligence, where we provide insights on financial impact, potential risks and scope for human intervention based on business and regulatory requirements.