An individual’s moral history influences a model’s decisions — and the direction of these decisions can be controlled through its activations

Moral evaluations assess decisions in isolation. By providing the models with an agent’s history, researchers have identified an internal mechanism that allows them to influence the agent’s choices at runtime.

The protocols that evaluate the morality of language models examine individual decisions. So they leave aside a question: does the past conduct of an individual, unrelated to the situation, influence the choice that the model makes about it? In humans, previous behaviour weighs on subsequent moral decisions. Researchers looked for the equivalent in models on two levels: observable responses, and internal representations.

Their cadre, named MoralLedger, varies an actor's moral history by keeping the decision context fixed. At the behavioural level, past history systematically changes the choices that follow, depending on their validity and intensity. At the level of representations, these same historical flows induce a linearly recoverable direction in the residual flow, which is generalized to unseen examples.

A control lever, not just a bias

To intervene along this direction on invitations with neutral history produces bilateral changes depending on intensity — effects more marked than those obtained by the mere formulation of the invitation, or by a favorable non-moral direction. To the knowledge of the authors, this is the first demonstration that a latent representation of the past moral conduct of an actor offers a signed control, at the time of inference, over a moral decision. Moral history thus becomes both a source of behavioural sensitivity and a causal target to audit or control this behaviour.