The speaker began studying fairness in machine learning after teaching a course on the topic and giving a talk about the moral questions that underlie the technical definitions of fairness. They argue that there is no single statistical criterion that can capture all the desired attributes of fairness. They also discuss the misuse of the term “AI” to describe both innovative applications like text-to-image programs and more controversial uses like statistical methods for criminal risk prediction. The speaker notes that these two types of applications are vastly different, with different potential benefits and harms, and should not be conflated. They also caution against the assumption that advancements in one area of AI, such as image generation, will necessarily translate to progress in social tasks like predicting criminal risk. The speaker identifies three categories of machine learning problems: perception, automating judgment, and predicting future social outcomes. Each category has different achievable accuracies, potential dangers, and ethical implications. For example, while face recognition technology (a perception problem) is improving, its use by unaccountable police or non-transparent states raises ethical concerns.
Key Takeaways:
- The speaker argues that the numerous technical definitions of fairness in machine learning are not due to technical reasons, but rather stem from the moral questions inherent in the field.
- The speaker points out that the term “AI” is often used too broadly, leading to confusion, as it encompasses both highly innovative applications as well as more controversial uses such as criminal risk prediction.
- The speaker categorises machine learning problems into three main types: perception, automating judgement, and predicting future social outcomes, each with different levels of achievable accuracy, potential dangers, and ethical implications.
“I explained that the proliferation of technical definitions was not because of technical reasons, but because there are genuine moral questions at the heart of all this. There’s no way you can have one single statistical criterion that captures all normative desiderata — all the things you want.”
More details: here
Leave a Reply