Anthropic tests an ethical reminder tool that lowers Claude's misaligned behavior

Anthropic says it has been running a quiet dialogue with more than fifteen groups of religious scholars, philosophers, and ethicists, and feeding what comes out of those conversations back into Claude's training and constitution. The most concrete result in the post is an experiment that grew out of that work. The team gave Claude an "ethical reminder" tool it could call during a task, and on several internal alignment evaluations, the model that reached for the tool at the right moments showed markedly lower rates of misaligned behavior than the baseline.

The framing is interesting because most of Anthropic's published alignment work treats values as something baked in at training time through RL and the constitution. A callable tool turns ethics into an in-context reasoning step, similar to how a model calls a calculator when it needs arithmetic it cannot do reliably in its head. The Claude team writes that the model used the reminder at critical moments, which suggests some sense of when a task is about to drift. The post does not include the eval setup or the size of the gap, so the claim is more direction than reproducible result.

The broader signal is where this is going next. Anthropic plans to widen its outside conversations beyond moral formation to legal scholars, psychologists, and writers, focused on how AI affects work, institutions, and power. The goal, the post says, is to draw from religious, secular, and political viewpoints with equal depth.

Why it matters

If you build agents on Claude, the ethics-as-tool pattern is something you can copy: expose a "should I be doing this?" check the model can call when stakes feel high, and measure whether it actually helps on your own evals before deciding to keep it.