Would your AI travel agent book a bullfight? Testing whether agents consider animal welfare without being prompted by Jonah Woodward

This ar­ti­cle re­flects new up­dates to the ac­com­pa­ny­ing pa­per: arxiv.org/​​abs/​​2606.18142. Bench­mark: now in­cluded in the UK AI Se­cu­rity In­sti­tute’s In­spect Evals. Leader­board: com­pas­sion­bench.com/​​tac.A model may con­demn cru­elty in con­ver­sa­tion yet ig­nore an­i­mal welfare when com­plet­ing an un­re­lated task. Stated con­cerns mat­ter lit­tle if they do not af­fect de­ci­sions. We tested whether mod­els con­sider an af­fected party with­out be­ing prompted, even when ...