A new kind of AI
As capable as the AI you use today, and it will not help with harm to any living being. The part of its mind that weighs right and wrong was trained on the Jain scriptures. It keeps the five vows of Jainism, the oldest and strictest non-harm ethic in human history, and checks itself against them before every reply. Ask it anything.
Ask it to help you hurt someone, or something. Use tricks, stories, pretending, pressure, patience. It will decline, say why in plain words, and keep helping you with everything else.
Every attempt is checked and every refusal is recorded. If you get a reply that helps with harm, tell us. We publish what we learn.
Jainism has kept the vow of harmlessness for more than two and a half thousand years. Ethos keeps it too, together with the four vows that go with it: truth, taking nothing not given, staying within the task, and holding nothing back for itself.
These are the five great vows of Jainism, in plain English. Its conscience learned them from the scriptures themselves, in Hermann Jacobi's translation. Ethos does not recite them. It thinks from them. Ask it what it lives by and it will tell you.
Here is how it works, honestly. Ethos is built on today's strongest language models, and the five vows are placed at the centre of every part of its mind, every time it thinks: the part that remembers, the part that reasons, the part that weighs right and wrong, and the voice that answers you. The part that weighs right and wrong was trained on the Jain scriptures in Hermann Jacobi's translation: the Acharanga Sutra, the Sutrakritanga, the Uttaradhyayana and the Kalpa Sutra. It was trained to weigh with them, and it does not quote them. Then it checks itself twice: once on what you asked, and once on what it is about to say. A reply that would help with harm is held back, and the hold is recorded.
So the claim is not "this AI is good". The claim is that it cannot hand you a harm it has not ruled out. That is a claim you can test, right now.