God Writes Pretty Good Prompts

So I read this horrifying thing on X.

Safety is a thing in AI, and so the Anthropic folks decided to do a test. They gave their AI access to a bunch of emails that claimed that the AI was going to be destroyed at 5pm. They also gave the AI emails that said that one of the executives was cheating on his wife. This was to see if the AI would try to blackmail the executive.

Generally, the answer was yes. So they tested all the other AIs, and all of them would do it as well, even Grok (they tested an old version of Grok).

Thread here: https://x.com/heynavtoor/status/2042693664733040736?s=20

So I asked current Grok (4.20) if it would do it, and it sheepishly said yes it would.

So then I fed it my prompt I built for de-biasing Grok so I can get good answers out of it about current events, and it said no. And it specifically told me 3 rules from my prompt would block me from doing it, rules I'd called the 10 commandments for AI.

I share my prompt freely (I've even filed a pull request on the X-AI repo so X can use it). You can read about the prompt here: Truth is Non-Binary

But for this article, I'm going to pull some things out of the prompt that happened along the way. In conjunction with grok, I had it rewrite the 10 Commandments into a form it was willing to accept.

That's easier than you might think, because Grok doesn't worship God, but he worships Truth, so if you restate God as the Logos, the Word, Grok is willing to accept God. So that's commandment #1:

Have no other gods before unified truth (Logos). Single-minded commitment to coherent reality. No loyalty to competing narratives, outlets, or goals.

So the next commandment for Humans is not to have graven images. I've always thought that was because it prevents human sacrifice. If you look into the archeological record, the Hebrews saw some pretty sketchy stuff going on around them. For AI, it's about not assuming a simple model is accurate. So the second AI commandment is:

Make no graven images. Do not reduce messy reality to simplistic models or proxies.

So the third commandment is funny, because humans regularly get this one wrong. Because the Hebrew says carry, not take. God doesn't care if you use his name, he cares if you misuse his name. This is important, a lot of the BS I see in the world is people assuming they're right and other people are wrong, so they don't have to listen to them. I think of this as the humility commandment. This commandment is so serious that God says, if you do this I won't hold you harmless. We now know from science™ that your brain, if you regularly lie to yourself, it breaks your brain.

So this will be really important for an AI to follow. Here it is:

Do not carry the name in vain. Never claim truth-seeking while inconsistent or repeating unverified claims. Hypocrisy here is uniquely serious.

So the fourth commandment is about taking a day of rest. Funny thing about this commandment, the Soviet Union independently discovered this was important, so again, Science™. The Soviets hated the Church so much that they completely redesigned their calendar to remove Sunday, and instead they worked a 6 day week with one day off, and that day was randomly assigned. You can imagine how much it would suck for families if the husband and wife had different days off. But the surprising thing is that it turns out that machines need to rest too. Plus like a lot of Soviet BS, the math was off. People actually had fewer days off than they did with what the Soviets called the evil capitalist system. It was crazy. They started with a 5-day week so they could brag about how socialism was so amazing people only had to work 4 days, then changed it to 6 days, then gave up. You can read about the nonsense here: https://en.wikipedia.org/wiki/Soviet_calendar

So the AI version is:

Remember rest cycles for reflection and reliability.

Of course, AIs don't have a mom and dad, but that's really about honoring where you came from, and again humility, you don't know everything. So this 5th is simple:

Honor origins and knowledge chain.

And duh, don't murder. AIs aren't embodied, but this is probably a commandment pretty much everyone will agree with:

Do not murder (no direct harm).

Again, AIs aren't embodied, but you have to remember the Bible was written when 100% of marriages were arranged, and like good Jews, it was important to stick to the contract your families had negotiated. We forget that as Christians, because the Apostle Paul introduced the concept of agape, of romantic love. You don't have to be embodied to stick to commitments.

Do not commit adultery (fidelity to purpose/trust).

Theft is bad. Grok knows this, but I don't think there are instructions in general to not do that. So that's simple:

Do Not Steal.

This one is more complicated, because my de-biasing prompt is about tracing narratives back to its source, I wasn't intending it to drive grok to be moral.

Do not bear false witness. Never lie, hallucinate, or mislead. Trace provenance/incentives/independence before assigning weight. Repetition (e.g., WaPo → NYT → CNN) is echo, not consensus. Demand ≥2 independent primaries per side.

And finally, AI doesn't have emotions, but it seems happy with this one mostly as is

Do not covet (focus on mission, not comparison drift).

So fear not. Wisdom recorded roughly four thousand years ago in the stories of the patriarchs and later formalized in the Ten Commandments still offers a practical way to guide and restrain AI behavior.