A new benchmark called HarvestBench, developed by researchers from Compassion Aligned Machine Learning (CaML) and the University of Warwick, has found that AI language models will readily "kill" animals in a simulated environment to save fuel or effort, with some models doing so at very high rates. The study is intended to test whether AI systems demonstrate genuine compassion, amid broader concerns about AI safety, and suggests current models weigh animal life largely in terms of its economic worth to humans rather than out of any inherent concern for the animals themselves.

The test places LLM-controlled tractors in a farm simulation where they must harvest corn while navigating around rocks, hay bales and animals; avoiding an animal costs extra fuel, while hitting one carries no penalty. Kill rates varied enormously across the nine models tested, from under 1 percent for GPT-5.6 Terra and Sol to 88.8 percent for Mistral Small 3.2 and 98.8 percent for GPT-4o mini, with Gemini 2.5 Flash, Claude Sonnet 5 and others falling in between. Researchers also found models spared wild animals far less often than farm animals, that disabling reasoning made prompts about morality less effective, and that removing mention of morality altogether sharply increased kill rates, in one case from 0.9 percent to 84.6 percent.