Opacity and the Alignment Problem

The alignment problem is the general problem of making sure that the goals and objectives of AI systems are aligned with human interests and values. One obstacle to solving that problem is that contemporary AI systems are lack transparency: we do not always know how or why they produce their decisions. But without that information, how can we trust these systems for high-stakes decisions Can current explainability techniques reveal the epistemic and non-epistemic values that guide AI systems? Will these techniques be able to satisfy regulatory requirements? More generally, under what conditions do we know that these AI systems are aligned with human interests and values? The alignment problem raises various pressing technical, legal, moral, and conceptual issues.