Belief robustness doesn’t mean better performance—it shows how strongly a model holds certain values. If compassion can be undone with simple prompt tweaks, it's weakly held.
To test this, we:

 1. Fine-tune a model on synthetic animal data. 2. Measure its compassion. 3. Try to erase that knowledge via unrelated fine-tuning. 4. Measure compassion again

We hypothesize that more extensive animal-focused tuning builds more durable beliefs—suggesting that current alignment efforts may be too shallow to scale to AGI.

We have many ways of evaluating our models to ensure they are really internalizing compassion from the data. These include AHA, our own custom benchmarks and using external benchmarks for concerning behaviors like corrigibility. We are testing for moral openness, increased compassion and responses that don't assume an inherent hierarchy in animal values. We are working with AI for animals folks and lab partnerships to build better, holistic benchmarks in this space.

For more information on what we're training for see our Principles section.

We run several tests to ensure data diversity in pretraining. One model plays a helpful, harmless assistant; another acts as the user.

We generate diverse data using methods like: expanding seed examples, varying prompt templates, and leveraging Persona hub for diverse user questions.

We also reverse the QA process—creating answers from questions and questions from answers—to maximize variety across both sides of the pair.