Activation steering is a technique that adds a vector to a model's hidden states to control a named behaviour. But a gain on that behaviour does not show that the vector controls it directly, because it may act through a component shared with other behaviours.
The paper proposes a competing-target test: if steering vectors for competing ethical frameworks are interchangeable, then the original vector is probably acting through a shared component rather than specifically targeting the named framework.
The finding implies that apparent success in steering models toward one ethical framework may not reflect genuine, specific control. Researchers need a stronger test to verify what a steering vector actually does.