I sometimes ruminate on how my learning journey and insights are different from how I would expect a machine to learn. While the recent launch of OpenAI’s GPT-3 model promises a potential path to Artificial General Intelligence (AGI), I do believe AGI is still a distant dream. Adoption of ethical principles and guidelines while building product is still very much in its infancy.
The Partnership on AI is central in Ethical AI, with several initiatives to champion AI adoption through education, sharing of best practices, open discussion groups and projects both related to policy and practice. Closing Gaps in Responsible AI is a neat project on identifying gaps in implementing ethical principles, building the bridge between principles to practice. The first phase of the project involves a crowd-sourced ideation game to identify gaps in current practice, potential solutions and scoring of potential solutions for feasibility and impact. I am eager to see their results published later this summer. The AboutML project is super-interesting as it explores how documentation of machine learning during development and deployment can be leveraged for ethical compliance and transparency.
When researching the actual implementation of the frameworks and demonstration of ethical productization of AI, I found some tools that could be applied but hardly any examples of successful implementation. IBM’s open source toolkits (Fairness: AIF360, Explainability: AIX360, Adversarial Robustness Toolbox: ART) and fact sheets for evaluating compliance have a lot of press. In addition, a prototype app from 510 Global (an initiative of the Netherlands RedCross) that built out a FACT Score “ethical” questionnaire is an interesting open source effort. Responsible AI Licenses that allow data scientists/developers to include limited use verbiage in their content found its first application with EdgeImpulse recently. Organizations are just beginning to claim ethical application of AI in their products or services but successful demonstration of responsible AI productization is still to come.
The central challenge to adopting ethical AI principles is the lack of a governing body to enforce adoption. While regulation is anticipated to originate initially in the European Union, until regulation happens, adoption of ethical principles may be inconsistent, sparse and weak.
