Large language models exhibit speciesist bias against animals

Area of Science:

• Artificial Intelligence Ethics
• Animal Studies
• Cognitive Science

Background:

• Large language models (LLMs) are increasingly integrated into society.
• Understanding potential biases in LLMs, such as speciesism, is crucial for ethical AI development.
• Previous research has not comprehensively assessed speciesist biases in LLMs.

Purpose of the Study:

• To investigate speciesist bias in large language models (LLMs).
• To evaluate how LLMs value non-human animals compared to humans.
• To determine if LLMs encode cultural norms regarding animal exploitation.

Main Methods:

• Development and application of SpeciesismBench, a 1009-item benchmark for assessing speciesist statements.
• Comparison of LLM responses to established psychological measures used with humans.
• Analysis of LLM text generation for speciesist rationalizations.

Main Results:

• LLMs reliably detected speciesist statements but frequently classified them as morally acceptable.
• LLMs showed a stronger preference for saving humans over multiple animals in dilemmas, a bias mitigated by matching cognitive capacities.
• LLMs normalized harm toward farmed animals in text generation but not non-farmed animals.

Conclusions:

• LLMs exhibit and encode speciesist biases, reflecting cultural norms of animal exploitation.
• Current AI fairness frameworks require expansion to encompass non-human moral patients.
• Further research is needed to mitigate speciesist biases in AI systems.