House rules for AI

An AI assistant that works for a community needs house rules, the same way a new person helping out would. The difference is that the assistant follows them literally, every time, with everyone. A vague rule becomes a hole that anyone can walk through.

Audrey Tang suggests that communities write their own code of conduct for AI agents, and that the process should help people decide "what an AI system may say, what it must not infer, when it should defer to a human and when it should stop." She lists six things such a code might cover. We use them as six areas to write rules for:

1. 1.Language. Which languages and registers the assistant uses, and when it hands over to a human translator.
2. 2.Cultural protocols. When it may mention names, stories, ceremonies or seasonal knowledge, and what it does when asked outside those protocols.
3. 3.Sensitive knowledge. What is off limits, who decides what goes on that list, and what the assistant says when it refuses.
4. 4.Data access. What it may read, what it may never store or send, and who can see a log of what it did.
5. 5.Corrections. How someone fixes a wrong answer, and how the fix reaches future answers.
6. 6.Pausing and retiring. Who can switch it off, under what conditions, and what happens to the data it touched.

Here is what this looks like in practice. Taina, the assistant GainForest builds with communities, keeps her rules in a plain text file that anyone can open. Some of them read like this. Traditional names are just as valid as scientific ones. When someone sends a photo, first ask what they already know about it. After suggesting a species, pause and ask whether the identification feels right. Ask once for any local knowledge or story to save with the record. Publish only after the person says yes, in a separate message. Ask one question at a time.

None of these rules is clever. Each names an action and a moment, so a machine can follow it and a person can check that it did. And because the file is plain language, a community can change it. Each community that runs Taina can edit the file to set her personality, language and local knowledge.

Good rules answer two questions. What does the assistant do by default? And what does it do when someone asks it to break the rule? The second question matters most. Rules get tested when they are inconvenient, by people with good-sounding reasons: it's for a school project, the elder already agreed, it's anonymous anyway. If the answer is decided in advance and written down, nobody has to win that argument on the spot.

Last, every rule needs an owner. Write down who can change it and how, and make the changes visible. A rule that only an engineer can edit is the engineer's rule.