Compassion in Machine Learning

Compassion in Machine Learning (CaML) is an AI safety research organization dedicated to shaping how advanced AI treats all sentient beings. Founded in August 2024, CaML conducts research on aligning artificial intelligence with the well-being of all sentient beings before the technology outpaces the ability to guide it.

CaML's core research method is Synthetic Document Finetuning, which generates targeted synthetic pretraining data to shift future AIs towards considering the welfare of non-humans, particularly animals and digital minds, in a way that is robust and designed to scale to superintelligence. The organization has demonstrated that this technique produces large improvements in animal compassion that persist even after supervised fine-tuning, with frontier labs expressing interest in such techniques to complement their existing alignment tuning.

CaML also trains AIs to exhibit moral humility to avoid incorrigibility and value lock-in, encouraging appropriate uncertainty and risk-avoidance about what entities matter and how much. Their benchmarking work includes the Animal Harm Bench (AHB), MORU (Moral Reasoning Under Uncertainty, a 201-question benchmark across four datasets covering alien organisms, digital sentience, AI values, and human compassion), and the CompassionBench leaderboard hosted at compassionbench.com.

The organization was co-founded by Miles Tidmarsh, a research economist with a Masters in Economics from the University of Melbourne and extensive experience in the Effective Altruism community, and Jasmine Brazilek, who brings over six years of cybersecurity experience including security systems design at Anthropic. The team also includes technical staff, a web designer, and several volunteers, with an advisory board that includes NYU philosopher Jeff Sebo, Sentient Futures founder Constance Li, and Open Paws founder Sam Tucker.

CaML operates as a fiscally sponsored project of Players Philanthropy Fund, Inc. and partners with organizations including Sentient Futures, Electric Sheep, Open Paws, and Wild Animal Initiative. Their models and datasets are publicly available on HuggingFace, where they have published approximately 200 models and 92 datasets.