Recent work has applied human psychometric questionnaires to large language models (LLMs) to elicit moral and value profiles, but it remains unclear whether these instruments capture stable model traits or whether the resulting profiles can be steered toward a target human population. We administered the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compared their foundation profiles with a sample of 1,282 Norwegian respondents. Two steering interventions were tested: prompt-level persona steering and activation-level ActAdd. Models were split into two groups: half engaged with the questionnaire under an attention check, while the other half defaulted to flat or central‑tendency outputs that appeared near‑human on average but ignored item content. A neutral Nordic‑respondent persona, written without any distributional information from the human sample, brought the engaging models 44%–77% closer to the Norwegian mean in Mahalanobis distance $d^2$. A single‑pair ActAdd applied at a fixed mid‑layer flattened the overall foundation profile rather than steering individual foundations. For at least one model, the same persona that shifted the profile also induced engagement that was absent at baseline, providing a concrete instance of the “cognitive phantoms” warned about by Peereboom et al. (2025).
Review