AI-SAFETY

If we get AI right, it could be the most beneficial technology in human history. If we get it wrong, it could be the most dangerous.

Constitutional AI is about building systems that are helpful, harmless, and honest. In that order.

Safe AI and powerful AI are the same thing. The companies that figure this out first will win.

Someone is going to build the most powerful AI in the world. It should be people who take safety seriously.

The default outcome of advanced AI is not good. We have to deliberately steer it toward good outcomes.

I left OpenAI because I thought we could do better on safety. Not because I thought they were bad, but because I thought the stakes demanded the best.

AI safety is one of the most important and most neglected problems we are aware of. That combination is exactly what we look for.

We started Anthropic because we believed there was a better way to build AI — one that doesn't require choosing between capability and safety.

The question is not whether powerful AI gets built. The question is who builds it and with what values.

Safety and helpfulness are more complementary than they are at odds. We don't see them as a tradeoff.

I don't think any one company should control the future of AI. Open models are not just technically better — they're safer for humanity.