CaML
Compassion in Machine Learning
Robustly improving the values of future AIs
Compassion in Machine Learning
Robustly improving the values of future AIs
Our Mission
CaML researches how mid-training data can shape the behavior and values of AI systems. We develop methods to ensure future transformative AI is broadly compassionate and morally thoughtful, treating all beings with consideration and care.
We have found evidence that Synthetic Document Finetuning can shift an LLM to be more robustly compassionate towards non-humans, and that this persists after typical fine-tuning.
Research
Two pillars of research aimed at building AI that genuinely cares.
AI fine-tuning often introduces flawed behaviors. We use Synthetic Document Finetuning to shape model values more carefully and reduce harmful outputs.
We measure whether models genuinely internalize compassion through custom benchmarks including Animal Harm Bench (AHB) and MORU Bench.
Results
Our leaderboard tracks how well leading AI models perform on compassion-related evaluations.
Get Involved
Your support directly funds the research that makes AI safer and more aligned with compassionate values.
We need funding to support our team and continue critical AI safety research. Every contribution helps advance the alignment of future AI systems.
Donate TodayJoin the Sentient Futures community to stay informed about our progress and connect with others working on compassionate AI.
Join CommunityAcknowledgements
We are deeply grateful to the individuals and organizations who make this work possible.
Thank you to the Survival and Flourishing Fund, Longview Philanthropy, Marcus Abramovitch, Ryan Kidd, BlueDot Impact, Macroscopic Ventures, Simon Newstead, Juliana Seawell, Alex Kastner, and many anonymous donors.
We are grateful to the Sentient Futures community for their support, especially in creating the AHB benchmark, and to our volunteers for accelerating our mission.
Sincere thanks to Strong Compute for donated compute time, enabling us to advance our alignment research.
CaMLCompassion in Machine Learning.
Robustly improving the values of future AIs.