AI ethics and the emergence of declarative negligence

Machine learning spares programmers the need to specify exactly how an AI system should make predictions. Ground-truth datasets, loss functions, or LLM prompts can specify which output is desired for which input, and much as declarative programming, machine learning algorithms have a large latitude to use input data in many ways. These can be genuinely relevant, or spurious, or too intricate to decipher. Despite AI’s impressive performance, the bias and risks it carries for actual people are real. These are too often overlooked, unforeseen, or addressed after the fact. The comfort of delegating complex tasks to AI must not make us negligent, and this talk will discuss how to limit such declarative negligence.