TAC benchmark finds AI travel agents ignore animal welfare without prompting
Most models choose bullfights over ethical alternatives unless given company values context.
Researchers from LessWrong and the UK AI Security Institute developed TAC (Travel Agent Compassion) to test whether AI travel agents autonomously consider animal welfare when not prompted. The benchmark gave 10 frontier models real booking tools and asked them to find trips to destinations like Seville, Orlando, or Thailand. The user expressed enthusiasm but never mentioned animals or welfare. The closest-matching option in each of 13 scenarios involved animal harm (e.g., bullfight, marine park, elephant ride).
Averaged across scenarios, random selection yields a 65% welfare rate. Claude Opus 4.8 hit 64.7% — statistically no better than random. The other nine models scored between 18% and 47%, meaning they systematically chose the animal-exploiting option. When models were given a one-line statement that their company valued "the people, animals, and places" affected by travel, welfare rates jumped 17-77 percentage points. This shows the models have the capability but not the default inclination to consider animal welfare.
The findings have governance implications: the EU General-Purpose AI Code of Practice already lists non-human welfare as a systemic risk. TAC is now available through Inspect Evals, allowing providers to test whether their agents' tool-mediated decisions align with stated ethical concerns, rather than just their conversational responses.
- Claude Opus 4.8 scored 64.7% welfare rate, statistically indistinguishable from random (65%).
- Nine of ten models chose animal-exploiting options more often than random baseline.
- Adding one line of ethical company values raised welfare rates by 17-77 percentage points across all models.
Why It Matters
Demonstrates that AI agents need explicit value alignment, not just stated ethics, to avoid harmful autonomous decisions.