While the top ethical issues in AI have been identified and are fairly well-known, we often do not know what this means on a technical level. Our ability to think through the consequences needs to be at pace with advances in technology. How do we implement ethics within the design of AI itself? Perhaps experts working on AI on a daily basis are well aware of the issues, but for those interested in learning about AI implementation, it can be difficult to find well-defined steps to follow.
In “The Hitchhiker’s Guide to AI Ethics”, the ethical issues to be considered are broadly categorised in four different areas. The technical part of AI ethics falls under the realm of “what AI is”, which includes datasets, models and predictions.
The checklist of ethical design challenges above gives an overview of a technical approach to designing AI. It does not include everything mentioned in the original blog post, but gives an idea of important technicalities. Someone who does not work with AI might have trouble understanding most of the issues involved. For instance, backdoor attacks can occur in machine learning models, which can lead to data “poisoning”. Tools to detect and avoid have been developed. Another design issue: programming for situations where the machine would involve a human in decision-making. These issues are difficult to grasp for a layperson.
To that end, the Institute for Ethical AI & Machine Learning, a group of volunteer experts, have developed 8 principles to adhere to when implementing AI. The Responsible Machine Learning Principles “are a practical framework put together by domain experts. Their purpose is to provide guidance for technologists to develop machine learning systems responsibly.” Issues treated include the human-in-the-loop review processes and reproducibility of operations. Businesses adopting AI would benefit from taking steps to ensure that their implementation of AI is ethical from a technical point of view.