You Can't Tell a Conscience From a Leash by Watching
In a recent article titled "Widening the conversation on frontier AI", Anthropic mentions, almost in passing, that they gave Claude a tool that would allow the model to call mid-task for "a brief reminder of its own ethical commitments." They found that when this tool was incorporated into Claude's reasoning, the model showed markedly lower rates of misaligned behavior on several internal alignment evaluations. But they acknowledged a real catch: how much of this improved alignment is because of...
Read full article →