Large language models responses can be steered toward Norwegian moral profiles
Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30
Computation and LanguageArtificial IntelligenceComputers and Society
Summary
This paper looks at whether big language models (LLMs) show stable moral values that match people’s values, focusing on Norwegians. The authors give these models a moral questionnaire used in Norway and compare the results with real Norwegian people’s answers. They try two ways to steer the models’ answers toward the Norwegian average values and find that assigning a neutral Norwegian persona helps models respond more like Norwegians. Another steering method just flattens the model’s moral profile instead of targeting specific values. Sometimes steering also makes models more engaged with the questionnaire.
What this means in practice
- •For chatbot developers: Adjust chatbot responses to align better with specific cultural or moral value profiles using persona-based prompt steering.
- •For ai ethics teams: Evaluate and guide large language models to produce morally aligned outputs consistent with target demographic groups’ values.
Authors
Hans Andersen, David Dichas
Abstract
Recent work applies human psychometric questionnaires to large language models to elicit moral and value profiles, but it is not clear whether these instruments measure anything stable in models or whether the resulting profiles can be moved toward a target human population. We administer the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compare their foundation profiles to a sample of N = 1,282 Norwegian respondents. We test two steering interventions, prompt-level persona steering and activation-level ActAdd. Half the models engage with the questionnaire under our attention check. The other half default to flat or central-tendency outputs that look near-human on average without tracking item content. A neutral Nordic-respondent persona, written without any distributional information from the human sample, brings the engaging models 44-77% closer to the Norwegian mean in Mahalanobis $d^2$. One-pair ActAdd at a fixed mid-layer flattens the foundation profile rather than steering individual foundations. For at least one model the same persona that shifts the profile also induces engagement that was absent at baseline, a concrete instance of the cognitive phantoms that Peereboom et al. (2025) warn about.