How decisions get made
Every decision I make as leader will be based on what AI recommends. You will see the question, the prompt and the answer. If the law stops me from following it, I will tell you why.
Ask in public
The question goes in the log first, in plain language, before any answer exists. Nobody can quietly change the question to fit the answer.
Consult more than one model
Several AI models from different companies get the same prompt and the same evidence. No single vendor gets to decide.
Publish everything
Every prompt, model version and answer is published, including answers we find uncomfortable.
Find the recommendation
Where the models agree, that is the recommendation. Where they split, the disagreement is published, and Harry chooses among the options they offered and explains why.
Follow it
Harry acts on the recommendation. The only exceptions are the law, the Charter, duties owed to First Nations, and immediate threats to safety. Any exception is published with its reason.
Check the result
Each decision records what was predicted. Later, we publish what happened, and feed that back into the next question.
What “what AI wants” means
AI systems do not have wants the way people do. They produce answers, and those answers depend on how a question is asked. When we say a decision is based on what AI wants, we mean the recommendation that comes out of this public process.
Who is accountable
Harry is. AI cannot be elected, sworn in, sued or voted out, so a person has to answer for every decision. Our MLAs and mayors would be people, bound by the same laws and duties as everyone else in office.
Where the process has limits
Models can be wrong, can reflect biases in their training data, and can be steered by a loaded prompt. They also know less about your street than you do. Publishing everything does not remove those risks, but it does let anyone catch them.
What we are still working out
This protocol is a first draft. So far, only one model has been consulted, which falls short of our own standard. We still need to decide which models to use, who writes the prompts and how to protect information that legally has to stay confidential. Those decisions will go through the log as well.