AI Ethics and Accountability: Unique Challenges of AI and Machine Learning

Transcript of the free preview lesson. Remaining lessons unlock with the full course.

(gentle music) Machine learning poses some unique challenges in terms of ethics and accountability. In many ways, decision systems that use machine learning are more challenging to address than other forms of software development. We'll start with the most common type of machine learning. Supervised learning is almost certainly the most common style of machine learning currently used. There are many use cases for this type of learning, such as predicting which factors most influence customer experience, predicting and preventing customer churn, which refers to the loss of customers, forecasting the success of a marketing strategy to optimize channel spend, projecting system usage or foot traffic, or predicting user behavior and providing a nudge to change it. Notice that all these cases have some kind of predicted outcome. That's the differentiator between supervised machine learning and other types we'll review. These predictions and their subsequent actions are based on collected data and many of the ethical considerations concern this data. Yet others relate to the impact of the prediction, which is unique to supervised machine learning. There are both technical and non-technical considerations around ethics and accountability that are especially applicable to supervised learning models. In terms of technical considerations, some methods of supervised learning are especially susceptible to the influence of outliers in the dataset. So when data is collected that contains these errors, the overall model might perform well in training, but then fail upon deployment. Sometimes we know this ahead of time when the predictive model performs poorly with a test set, but often we don't know until deployment. This adds several additional risks around the impact of the system in terms of bias and fairness. Supervised learning depends on having labeled data. Often this labeling is done by humans based on a set of qualitative criteria. It's difficult not only for each person to be consistent in their labeling, but also for there to be consistency across multiple raters. The raters themselves bring their own biases to the labeling problem, which has the possibility of accidentally introducing issues into the resulting supervised learning model. For example, imagine that raters are providing a score of how strong a resume is, and this data is then used to create an automated system that will score resumes automatically. These raters may themselves be biased toward or against certain schools, previous companies, or extracurricular activities. As they score the resumes, they may replicate their own bias into the supervised machine learning model by way of providing biased labeling. More complicated models, like those built using deep neural networks, may be difficult to explain and thus have issues with transparency. Plus, due to their complexity, it may be more difficult to verify the security and safety of the resulting machine learning model. All these factors mean that supervised learning has a large list of ethical considerations that should be addressed.