The Human Alignment Experiment
A worldwide experiment in whether we can find what we share and what we would do together, enough to act as one team on what matters to everyone.
A worldwide experiment in whether we can find what we share and what we would do together, enough to act as one team on what matters to everyone.
We do not need to agree on everything. We need to agree on a few things that make working together possible at all.
What we have found so far
Realizing the power of our interconnectedness.
What holds it together
Each line holds the threads beneath it. Open any thread to read it, see where the idea comes from and what argues against it, and add your own reaction.
Opening the commons
All the layers
What we hold in common, open to read, question and add to. It starts from one root, distills into a few lines, and opens into threads you can follow as deep as you like.
OpenThe odds are long and the cost of trying is small.
Why alignment
The three crises form a hierarchy. The polycrisis is what we see. The metacrisis is what generates what we see. The belief crisis is what prevents us from responding to what we see.
The polycrisis: multiple interconnected breakdowns amplifying each other at once. Each failure makes the next one harder to address.
The metacrisis is the pattern that produces crises.
Despair is rational, and despair is self-fulfilling. Both are true.
The solutions exist. The bottleneck is the shared recognition that we are in this together.
Human alignment is a precondition for solving our largest problems.
Lineage
- Polycrisis: Edgar Morin and Anne Brigitte Kern, Homeland Earth (Terre-Patrie, 1993).
- Metacrisis: the generator beneath the crises, in the work of Daniel Schmachtenberger and of Jonathan Rowson.
For most of human history, you could externalize costs. You could pollute upstream and someone downstream would deal with it.
We connected everything. The smoke comes back. The virus gets on the plane. The crash crosses the ocean in a second. What you do to others, you do to yourself now.
If everyone is rowing in a different direction, nobody goes anywhere in an interconnected rowboat. If we point in the same direction, we all move faster than anyone thought we could, without any more effort.
The path emerges through the participation of everyone.
Lineage
- Elinor Ostrom, Governing the Commons (1990): those who share a resource can make and keep the rules that sustain it.
The interconnectedness that makes risk planetary in scale also makes planetary cooperation imaginable.
The tools finally match the ambition.
- AI can hold a conversation with millions, and remember each one.
- Trust online is collapsing, and proving that a real person took part is newly possible, lightly.
- The fear of misaligned AI has made alignment a word we know.
The timeline pressure is severe. If the window for coordination is shorter than the window for AI disruption, the case for belief may be technically correct and practically irrelevant.
Lineage
- Deliberation at scale: Polis and the vTaiwan process (from 2015), and Collective Constitutional AI (2023), in which about a thousand participants wrote and voted on principles for an AI model.
We have never been more powerful, with access to exponential technologies that have the proven power to affect the entire world, for the better or for the worse. That is a problem because humanity has lagged behind in the evolution of its collective wisdom.
If we're misaligned, AI will accelerate our misalignment all the way till we rip apart. But if we could align, AI could be an ally in helping accelerate that alignment.
The danger is misalignment between what the technology accelerates and what we actually want.
Feeds built for attention reward what divides.
Lineage
- Kentaro Toyama, Geek Heresy (2015): the law of amplification, that technology amplifies human intent and capacity.
The problem of AI misalignment is downstream of the problem of human alignment. How could AI possibly be aligned with humanity when humanity is so misaligned with itself? The dominant framing of AI alignment asks how we align AI with humanity. This question has a problem: it presumes humanity has coherent values to align AI to. We do not.
AI amplifies whatever direction we're already heading. If we stay this misaligned with each other, it will help us do it faster. Every AI is aligned to someone's values, and right now very few of us get asked. The conviction behind this experiment: human alignment comes first, and it is a necessary step toward AI that is aligned with all of us.
It is a conviction to test in the open. Making AI systems reliably safe is also deep technical work. This experiment works on the human half, beside those working on the technical half.
Lineage
- Norbert Wiener, Some Moral and Technical Consequences of Automation (Science, 1960), on making sure the purpose put into the machine is the purpose we really desire. Brian Christian, The Alignment Problem (2020).
The stream
The campaign's plans, open to the core team and the collaborators working on it. Each opens in the reader, to read, comment and suggest.