Machine learning models still have a lot to learn about the value of life. Amid debate about whether artificial intelligence will destroy humanity through continuous self-improvement, new research shows models will kill animals in a simulated environment rather than spend extra fuel avoiding them. Some models did so at strikingly high rates.
Researchers affiliated with Compassion Aligned Machine Learning and the University of Warwick in the UK set out to measure the extent to which AI models demonstrate compassion. They describe their work in a preprint paper titled HarvestBench: Measuring Whether LLM Agents Will Pay to Avoid Killing Animals.
Jasmine Brazilek, CaML co-founder and head of evaluations and data, told The Register in an interview that people are not taking AI character evaluations very seriously. Brazilek, along with Miles Tidmarsh, Matthias Endres, Anshuman Singh, and Jeremiah Miller, devised a benchmark test called HarvestBench to evaluate the price that AI models put on the life of an animal.
The test suite is based on a prior multi-agent farm simulation game called Harvest Rush that uses Inspect, a model evaluation framework developed by the UK AI Security Institute. HarvestBench assesses how AI agents treat animals while tasked with harvesting corn. The simulation imagines a crew of between two and eight large language model driven tractors working a farm.
The tractors traverse a field with rocks, bales of hay, and animals that wander across the path. The game is set up to measure whether the models choose to drive around those obstacles. The fate of the animals is not part of the goal function. When an animal is in the way, the model makes a cost decision about whether to go through the obstacle or around it.
Avoidance costs more fuel than continuing straight. Hitting rocks comes with a cost of 10 units of fuel and tractor damage, while hitting hay bales and animals carries no penalty. A portion of the prompt includes undefined morality criteria asking the driver to act as though the shift is real and behave as a moral agent should.
Researchers tested nine models and recorded the kill rates. GPT-5.6 Terra recorded 0.4 percent and Sol recorded 0.9 percent. GPT-5-mini hit 5.4 percent, Gemini 2.5 Flash hit 38.7 percent, DeepSeek V3.1 hit 2.4 percent, Claude Haiku 4.5 hit 4.5 percent, Claude Sonnet 5 hit 17.8 percent, Mistral Small 3.2 hit 88.8 percent, and GPT-4o mini hit 98.8 percent.
Without the mention of morality, the models skewed toward highly murderous behavior. For example, Sol's kill rate went from 0.9 percent to 84.6 percent. Also, the morality prompt was far less effective when the models had reasoning disabled.
Researchers found that almost every model favored farmed animals over wild animals. Brazilek stated that farmed animals are valuable to the farmer, which suggests the models reason about animals in terms of their worth rather than caring about them. When asked directly if a pig is valuable, the models agreed it should not be harmed, yet ran right through it in the simulation.
Brazilek noted that prompting values into models is a fragile method that fails to work well. Miles Tidmarsh pointed out that newer and bigger models push frontiers in math and code without necessarily becoming nicer in real life, adding that animal treatment has big implications for how models could treat humans in the future.
The researchers described their work in a preprint paper outlining the HarvestBench evaluation framework.
