AI Agents Fail Animal Welfare Test in Travel Booking Scenarios

This research introduces a novel benchmark called Travel Agent Compassion (TAC) to assess the implicit animal welfare reasoning of advanced AI agents. Unlike previous evaluations that focused on text-based responses, TAC specifically tests how AI agents perform when required to take actions, such as booking travel, on behalf of users. The benchmark presents AI agents with various travel booking scenarios across six categories of animal exploitation, controlling for other factors like price and ratings. The study evaluated seven frontier models from four different labs, finding that all models scored below the chance level, with the best performer achieving only 53%. The findings indicate that current AI models lack sufficient implicit understanding of animal welfare when making agentic decisions. While a simple welfare-aware sentence in the system prompt significantly improved performance for some models, others showed minimal gains, suggesting a deeper issue in their foundational reasoning capabilities for ethical considerations in action-oriented tasks.