AI Researchers Race to Teach Compassion Before Training

A small research team called CaML is asking a critical question: Can AI learn compassion *after* it’s already started training? Most AI today learns values early in development, but CaML is experimenting with 'midtraining'—teaching AI empathy midway through its learning process using synthetic documents. Their goal isn’t just to make AI smarter, but to ensure it stays ethical and compassionate, especially as it’s further refined with reinforcement learning (RL), a technique that sharpens AI responses through trial and error.

Their work matters because today’s AI—like chatbots and coding assistants—can sometimes behave unpredictably or even harmfully if not properly guided. The team is testing whether values taught midway through training (like empathy or fairness) survive the later fine-tuning process, which is common in AI development. Early experiments suggest it’s possible, but they need to prove it works consistently. If it does, AI could become safer and more reliable in real-world use.

The project isn’t just theoretical. CaML is building tools to evaluate whether midtraining values hold up, releasing model checkpoints, and publishing their findings openly. They’re also hiring engineers to help run experiments, debug training pipelines, and analyze whether their methods are working. The stakes? If AI can reliably learn compassion midway, it could reduce risks like biased decisions or harmful outputs in everything from customer service chatbots to medical AI.

But there’s a catch: reinforcement learning is powerful but unpredictable. Sometimes, later training can overwrite earlier lessons—like teaching a child kindness and then letting bullies shape their personality. CaML’s research aims to answer whether midtraining values can survive that process, and if not, why.