What is the Main Source of Unresolved Machine Learning Ethical Concerns?

Machine Learning (ML) is a branch of AI that is gaining significant attention as its potential to help solve practical challenges across various domains is enormous. Despite the potential benefits of ML, it is important to examine the main source of ethical concerns in ML, highlight some of the unresolved ethical concerns, and propose ways forward for developers and policymakers.

What are the unresolved ethical concerns in ML?

Ethical concerns cut across many disciplines. Some would argue that ethics questions fall under the philosophy discipline, not science, but we need to take a broader view as social, cultural, and economic domains are also affected. Among some of the ethical concerns in ML still to be resolved, either technically, socially or through research are but not limited to:

• System and algorithmic bias as decisions made from ML reflect the quality of data that was used to train the model
• Misrepresentation and under-representation of facts which give rise to disinformation challenges
• Transparency issues from data processing to the decision-making point of view
• Privacy concerns as surveillance now naturally follow ML designed camera-ready systems
• Trustworthiness concerns as the designs of some ML applications are not revealed to the public
• Functionality issues as ML designs lack solid moral decision-making capabilities

What is the source of unresolved ethics concerns in ML?

Many of the ethical concerns are the after-effects of the design process. We need to move beyond the design and look into the design inputs. ML thrives on data, as such if we want to understand the source of the various unanswered and unresolved ethical issues, we need to interrogate the datasets from which the ML models and designs are trained. The initial biases found in datasets are often amplified if they are not rooted out before being used as training datasets. Therefore, input data that has traces of bias won’t go away even when we have a trained ML model capable of performing certain functionalities.

What could be the way forward?

Dataset quality assurance should be mandatory for all machine learning-driven applications whose applications domain is in the public space which may affect human lives directly or indirectly. Open access repositories for datasets could help developers and regulators access data that are being used in ML designs in order to validate the accuracy and determine if they are free from errors. The central repository approach that Finland is working on could serve as a guide for best practices for policymakers with regards to regulatory approaches towards data labelling and standardisation. With increased accountability towards data sources, some of the challenges highlighted above could be minimised.