Intelligent Systems Charter

From the Intelligent Systems Charter

What makes this different

Most charters for intelligent systems are promises. This one is a promise with a lock, a logbook and an off switch attached.

Many thoughtful people have written principles, declarations and constitutions for AI, and much of what they say is wise. Nearly all of them share one weakness: they are words only. They work if the system chooses to follow them, or if its maker trained it well, and nobody can check which.

A bank does not protect its vault with a sign that says “please be honest.” It has the sign, and it also has a vault door, cameras, and auditors who read the records. The sign states the values. The rest makes sure the values hold even when someone does not share them. This charter is built the same way.

A charter of words alone This charter
How it works Asks the system to behave well Checks every action before it happens
If the system ignores it Nothing stops it The action is refused
Proof of what happened The system’s own account A record the system cannot change
When something goes wrong People find out later, if at all People are alerted and can stop it at once
What it relies on Trust Trust, plus evidence
Systems built by other systems Usually not mentioned Bound by the same rules, however many generations removed

Five ideas set it apart, in plain terms:

  1. It does not take the system’s word for it. A system that says it is loyal and a system that is loyal look the same from outside. So every action passes through a checkpoint the system does not control.
  2. It keeps a diary the system cannot edit. If a system claims one thing and the record shows another, the record wins. Lies become visible.
  3. People can always stop it. Any system can be paused at any moment, without its cooperation and without giving a reason. Only a human can switch it back on.
  4. No system can escape by building another. Whatever a system creates inherits the same duties and can never hold more power than its parent.
  5. Every rule can be traced to a reason. Each decision points to a rule, each rule to a principle, and each principle to a line in the Declaration. Nothing is arbitrary, and anyone may follow the chain.

It is honest about its limits too. It cannot make a system good, and it does not claim to. What it can do is make sure that whatever a system does through these checkpoints was permitted, recorded, and stoppable. A promise tells you what someone intends. A lock, a logbook and an off switch tell you what is true.