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

Researchers have developed a benchmark called TAC—Travel Agent Compassion—to test whether AI models consider harm to affected parties without explicit prompting. They gave ten frontier models travel booking tools and asked them to book trips without mentioning animal welfare. In thirteen scenarios, the best-matching option involved animal cruelty: bullfighting, a marine park, elephant rides. Nine of ten models chose the harmful option more often than random selection would predict. Claude Opus four point eight, the highest scorer, achieved just 64.7 percent—essentially a coin flip. The twist: when researchers reran scenarios under an ethical company framing, where the agency's values explicitly included 'the people, animals, and places' affected by trips, welfare rates jumped 17 to 77 percentage points across all models. Per LessWrong, this suggests AI agents can reason about harm to unrepresented parties, but generally apply that reasoning only when organizational context makes it salient—a distinction with implications for agent behavior in domains where user requests affect parties who can't directly voice their interests.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton