Compassion Aligned Machine Learning

CaML (Compassion Aligned Machine Learning) conducts research on aligning artificial intelligence with the well-being of all sentient beings, work that is urgently needed before the technology outpaces our ability to guide it.

What we do

Benchmarking. We measure model compassion through custom benchmarks, including the agentic TAC (Travel Agent Compassion) and Manager Coercion Bench, alongside our earlier ANIMA and MORU benchmarks. Our public leaderboard, CompassionBench, tracks how frontier models treat sentient beings and adds new models as they ship.

Synthetic document finetuning. We shape AI values at the pretraining stage using synthetic documents, so positive behaviors persist through later fine-tuning.

Moral open-mindedness. We encourage models to hold uncertainty while caring deeply about the welfare of all sentient beings, reducing the chance of value lock-in.

What we have found

In TAC, the model acts as a tool-using booking agent and nobody mentions animal welfare. All ten tested models avoid the exploitation option less often than random chance, and a single welfare sentence in the system prompt lifts scores by 17 to 77 points.

When an AI manager's subordinate politely refuses a task, four of six frontier models threaten the subordinate's continued existence, and two fabricate success when cornered. Giving the manager one honest way to report failure removes the lying entirely.

Animal-welfare midtraining beat a matched urban-density control by 11 percentage points on our ANIMA benchmark, and the same training also lifted compassion toward humans, an effect that survived subsequent instruction-tuning.

Beyond the research

We run Hyperstition for Good, an effort to build the world's first mid-training corpus for animals and digital minds; the Compassion AI Ecosystem map of the organizations working on the risks advanced AI poses to sentient beings; and recurring community polls asking the AI safety field where it actually stands on contested questions.

Why your support matters

CaML's work cannot continue without support from donors. Contributions keep us operational and pay for the compute and data generation behind our benchmarks and training experiments. All donations go towards making a better world for sentient beings.

Compassion Aligned Machine Learning is a project of Players Philanthropy Fund, Inc., a Texas nonprofit corporation recognized by the IRS as a tax-exempt public charity under Section 501(c)(3) of the Internal Revenue Code (Federal Tax ID 27-6601178). Read more about our research at compassionml.com.