Steering vectors for ethical perspectives often overlap and are interchangeable

Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering

Artificial IntelligenceComputation and LanguageMachine Learning

Summary

Activation steering means nudging an AI’s internal states to control its behavior, but it’s unclear if it actually targets a specific behavior or shares parts with others. The authors introduce a test that compares vectors aimed at competing ethical views—Kantian deontology and act utilitarianism—to see if they truly differ. They find that vectors made to steer these ethics are very similar and influence each other’s scores, suggesting they control a shared underlying perspective rather than distinct ethics. However, vectors for simpler differences like tone (optimistic versus pessimistic) are more distinct. This means steering an AI to follow one ethical viewpoint may blend in aspects of another unless carefully tested.

What this means in practice

  • For ai ethics teams: Evaluate whether different ethical perspective controls in AI models actually target distinct values or share overlapping internal signals.
  • For natural language processing engineers: Design activation steering vectors that better distinguish between subtle competing behaviors, improving controlled AI outputs.

Authors

Weilun Xu, Zipeng Sun

Abstract

Activation steering adds a vector to a model's hidden states to control a named behaviour. A gain on that behaviour, however, does not show that the vector controls it directly: it may act through a component shared with neighbouring behaviours, which unsteered and random baselines cannot separate. We introduce the competing-target test: a vector is compared with the vector built the same way for a competing target, and an equivalence test states when their difference is below a useful margin. We apply it in open-weight models to vectors for Kantian deontology, which judges an act by the rule it follows, and act utilitarianism, which judges it by the welfare it produces. On held-out situations, prompting a framework raises its score and lowers the other's, whereas vectors for either framework move both scores alike, and none reaches the margin. The two vectors share most of their direction, a request to apply some framework; the same construction separates an optimistic from a pessimistic tone, whose vectors share much less. What a vector shares with its competitors therefore sets what it controls: steering a perspective can deliver a mixture, and measuring a disposition can measure its neighbours, unless tested against the competing target.
View PDFOpen arXiv