What If AI Could Learn From Humanity At Its Best?

In 2019 I started carrying a phone and a small stack of PostIt notes everywhere I went. Whenever a thought arrived that I wanted my children to have one day, I wrote it down. This was not a project I sat down and finished in an afternoon. It took months. I had been trying for years to pass on what I had learned, and watching most of it fail to land, because I talk too long and my children, reasonably, drift. So I gave up on the lecture and collected fragments instead. After several months I gathered the notes, shaped them into thirty principles, printed the document and showed my family. I called it A Father’s Advice.

It did not become the thing I had pictured. I had imagined the grand framed painting for our living room wall, something they would return to when they needed it. Instead it went into a desk drawer, and it stayed there for years.

What pulled it back out was artificial intelligence.

By 2025 I was reading more and more about how these systems are built, and one detail kept catching on something in my mind. Today’s AI learns mostly from what is easiest to collect: billions of words scraped from the open internet. Comment threads, arguments, quick reactions, the things people type when they are annoyed or amused or trying to win. That is a real picture of humanity. It is just not the whole one. It is weighted heavily toward our reactive mode, the part of us that responds fast and feels strongly and rarely stops to think.

Here is what struck me when I reread my own thirty principles in that light. None of them were reactive. I had not written down my anxieties or my arguments. I had written the opposite. I had told my children to measure their progress only against who they were yesterday and not against other people. I had told them to own their attention, because in a digital world attention is valuable and the apps and algorithms will take more of it than you ever meant to give. That is not the voice I use when I am online. It is the voice I used when I imagined not being around, and asked myself what I actually wanted to leave behind.

There is a structural difference between writing for engagement and writing for legacy. The first rewards heat. The second rewards care. When a parent sits down to write principles for a child, they do not reach for their fastest take. They reach for their considered best. That mode of human expression is real and it is common, and almost none of it sits in the data we are using to train AI, because almost none of it is posted in the places we scrape.

People do not put their deepest convictions in a comment section. They write them on PostIt notes and bury them in a drawer.

I want to be honest about one thing here, because it is in the document and I am not going to pretend otherwise. One of my 2019 principles was already about AI. I had told my children that computers, AI, the internet and robots would fill more and more of their lives, and that they should learn to use them wisely. So I cannot claim the AI question arrived from nowhere. What changed was its direction. In 2019 I was asking how my children should relate to AI. By 2025 I was asking the reverse. What should AI learn from people like them.

The hypothesis behind Codex Vivendi is simple, and unproven. A structured record of human values, written in reflective mode rather than reactive mode, might be a useful input for the people working to align AI with what humans actually care about. I want to be careful with that sentence. It might be a useful input. I do not know that it is. AI labs may never use it. The difference between reflective data and scraped data may turn out not to matter in any way that affects how these systems behave. I find the question interesting enough, and the cost of trying small enough, that not trying seemed worse than trying and being wrong.

Part of what drew me in is that the thing these systems struggle with most is the thing humans have argued about for as long as we have existed. What is a life for. We have never settled it. Maybe that is the point. The answer moves depending on who is asking and when, and it has always moved. But people answer it anyway, every day, in how they decide what matters and what they are willing to give up for it, and most of all in what they try to pass on to their children. Almost everyone wants the next generation to have it better than they did. That impulse is quiet, it is nearly universal and it is almost entirely missing from the record we are handing to the machines.

So I built the experiment that would let me find out whether the question has a useful answer. People write their principles in a structured form and keep ownership of their words. The mechanics, the privacy, the funding and how the shared record is handled all sit on the site for anyone who wants to check them. None of it is novel in concept. The one thing that might be is the format, whether reflective articulation at scale produces something different enough from what already exists to be worth having.

I am not writing this to ask for money, a signup or an endorsement. I am asking for the one thing this community is unusually good at, which is scrutiny. The thesis is a hypothesis, not a position I need to defend. If it is wrong I would rather find out early and cheaply. If there is a reason the reflective-versus-reactive distinction falls apart under examination, I want to hear that reason. And if it holds, I would like to find the people who think it is worth building properly.

It started as one father trying to leave his children something that would outlast him. It turned into a question I could not put back in the drawer. I do not know whether Codex Vivendi will become what I hope it might. I know the question is worth asking out loud, in the place where the people who think hardest about these systems are already talking.

So I am asking it here.