These supplementary experiments test whether the steering effects persist under small changes to the prompt and choice of transport modes. Each figure combines the unsteered model in panel (a) with public-transit and car-dependence feature steering in panels (b) and (c). Colored curves denote feature-activation multipliers; the dashed zero line marks no change from baseline.
Prompt wording
Using “or” instead of “/”
Figure S1 · “Or” wording. (a) Baseline walk-minus-drive logit difference for the prompt variant using “walk or drive”; positive values favor walking. (b–c) Change in that logit difference after scaling public-transit and car-dependence features, respectively.
Using “meters” instead of “m”
Figure S2 · Spelled-out distance units. Replacing “m” with “meters” preserves the baseline transition and the opposing effects of the two feature groups. Positive values favor walking.
Option order
Changing the order
Figure S3 · Reversed option order. Positive values now favor driving. Public-transit and car-dependence features retain their opposing effects despite the reversed ordering.
Using “or” and changing the order
Figure S4 · Reversed order with “or.” Positive values favor driving. The direction of both interventions remains consistent with the original prompt.
Alternative transport modes
Replacing walking with cycling
Figure S5 · Cycling instead of walking. Positive values favor cycling. The feature groups generalize in the expected pro- and anti-driving directions.
Replacing driving with cycling
Figure S6 · Cycling instead of driving. Positive values favor walking. Both feature groups have small, inconsistent effects when neither choice is driving.
