●researcherBe cautious when interpreting model welfare findings, as results are highly dependent on the elicitation instrument.
●policyStandardized protocols for measuring AI preferences are necessary to ensure consistent safety and alignment assessments.