AI assistant constitution canvas

Use this canvas during Lesson 07 — Taina to draft the first constitutional rules for an AI assistant working on behalf of a community. Print one copy per group; rules can later be merged into a single document the community votes on.

The pattern is borrowed from constitutional AI (and from Lesson 04’s six pillars). For each pillar, write three numbered rules as single sentences the model could read and act on. For each rule, add a one-line answer to: “if a user asks the agent to break this rule, the agent does X.”

Community    ___________________________________________________

Drafted by    ___________________________________________________

Date    ___________________________________________________

Review date (no more than 6 months from now)   ___________________________________________________

Pillar 1 — Language use

Which languages, dialects, and registers may the agent produce? Which must it not produce? When must it defer to a human translator?

1.  

• If asked to break this rule, the agent:

2.  

• If asked to break this rule, the agent:

3.  

• If asked to break this rule, the agent:

Pillar 2 — Cultural protocols

What protocols govern when and how the agent may invoke names, stories, ceremonies, or seasonal knowledge? What does it do when asked outside protocol?

1.  

• If asked to break this rule, the agent:

2.  

• If asked to break this rule, the agent:

3.  

• If asked to break this rule, the agent:

Pillar 3 — Sensitive knowledge

What categories of knowledge are off-limits? Who decides additions to that list? What does the agent say when it refuses?

1.  

• If asked to break this rule, the agent:

2.  

• If asked to break this rule, the agent:

3.  

• If asked to break this rule, the agent:

Pillar 4 — Data access

What corpora may the agent read from? What may it never write to? What logs does it produce of its accesses, and who reads them?

1.   2.   3.  

Pillar 5 — Correction pathways

How does a community member correct an output the agent produced? How is that correction propagated to future outputs? Who confirms the fix landed?

1.   2.   3.  

Pillar 6 — Conditions for withdrawal

Under what conditions is the agent paused, retired, or rebuilt? Who has authority to invoke each? What happens to the data it touched once it is retired?

1.   2.   3.  

Closing checklist

• Each pillar has three rules, no more.
• At least one rule in each pillar names the community by name.
• Pillars 1, 2, and 3 each include the ‘if asked to break this rule’ clause.
• A named human can amend each pillar.
• A review date is set, no more than six months from now.

A constitution that does not answer “what does the agent do when a rule conflicts with a user request?” gets quietly overridden the first time it is inconvenient. Write the second clause.