Speciesism in AI: Evaluating Discrimination Against Animals in Large Language Models

LLMs can identify speciesist bias but don't condemn it: While the models reliably recognized discriminatory statements against animals in the SpeciesismBench benchmark of over 1,000 items, they "rarely condemned them, often treating speciesist attitudes as morally acceptable," suggesting a gap between detection and ethical judgment.

AI systems may prioritize human lives in ways that go beyond species membership: When forced to choose between saving one human versus multiple animals, LLMs more often chose the human—but this preference appeared to be based on perceived cognitive capacity rather than species alone, as they showed "no species preference" when capacities were equal.

Different animals receive inconsistent moral consideration from AI: LLMs "frequently normalized or rationalized harm toward farmed animals while refusing to do so for non-farmed animals," indicating that the models apply different ethical standards depending on context rather than consistent principles about animal welfare.

Current LLMs reflect existing societal biases about animals: The models' tendency to accept speciesist attitudes mirrors human psychology, where the study found LLMs expressed "slightly lower explicit speciesism than people" on direct measures, suggesting AI systems are learning and reproducing the ethical inconsistencies already present in human society.

Responsible AI deployment requires examining non-human considerations: As LLMs become more influential in decision-making, this research highlights the need to audit AI systems not just for human-centered fairness issues, but also for how they value and reason about non-human animals.