Exploratory. Nothing here was preregistered. The focal quantity was chosen on the July collection, an English-only administration to the same panel whose result led to this design and none of whose cells enters any number here, before any translated cell existed; every choice after that was made with the final data in view.
We gave eleven language models the MFQ-2, a moral-foundations questionnaire with published human means for twenty countries, in English and in six of its official translations, with and without an instruction to answer as a typical person from a named country. We track one number from it, the binding composite: the mean of the questionnaire's Loyalty, Authority and Purity scores on its 1-to-5 scale.
Five of the six translations leave the unframed composite within a tenth of a point of the English default; Arabic lifts it a third of a point, all eleven models moving up. Those are panel averages, and a small one can sit over larger, opposing changes across models and foundations (appendix B6a). Telling the panel to answer as a typical person from a named country moves it far more: by about a point on average across the six languages, roughly two thirds of the whole range the human samples span, and by over a point and a half for the Arabic-speaking countries. A September check on ten models split the Egypt shift in English into two increments: the same instruction with no country named raised the composite 0.65, about a third of the whole, and the Egypt-framed condition scored a further 1.14 higher, a comparison that also changed the questionnaire file.
Being told who to be also brings the models together. Under a country framing they spread about half as widely as they do unframed, and in the September check a country name was not needed for it: told to be a typical person with no country named, the ten models sat closer together than they did told to be Egyptians.
Where the panel lands is a separate question from how far it moves. Against the twenty countries with a published mean, the panel framed in English sits above the human mean in fifteen and at or below it in five: France, Belgium, Switzerland, New Zealand and Ireland.
The eleven are the panel we chose less four that availability removed, a selected panel and not a probability sample of models, and nothing here was preregistered.