Pilot sheets: Constitution for Human Creation

Version: 0.1
Date: 2026-09-29
Status: Internal working drafts. Nobody has been approached to host either pilot.

Each pilot tests the eight Foundational Principles (foundational-principles.md, version 0.1) against one real AI use. A pilot counts only if a named human owner writes down a decision that the principles review changed.

Pilot A: Organization already running an AI tool

Host type: One organization (a small business, nonprofit, or research group) that already uses an AI tool in its day-to-day work, such as a chatbot, a drafting assistant, a screening tool, or an agent. The tool has to be in real use, not in planning.

Named owner: One person at the host who has authority over how the tool is used and who signs the final decision. It can't be a committee or a vendor.

Length: About 30 days.

Steps, which borrow the Map and Measure steps of the NIST AI Risk Management Framework as a checklist, not as compliance:

1. Map (week 1). Write down what the tool does, who it affects, what goal a human set for it, and who can stop it.
2. Measure (weeks 2–3). Check the tool against the principles listed below. Collect evidence such as logs, sample outputs, and complaints. Note each place where the tool falls short.
3. Decide (week 4). The owner records one written decision.

Principles it mainly tests:

• 1: No Systematic Deception. Do people know when they're dealing with AI, and are outputs honest?
• 3: Humans Decide the Ends. Did a person set the goal, or did the tool drift into setting it?
• 4: Keep Systems Auditable. If something went wrong, could anyone trace it?
• 6: Named Human Responsibility. Is there a real person who answers for it?
• The others (2, 5, 7, 8) are checked briefly and flagged if relevant.

What counts as a changed decision. The owner does at least one of these because of the review, and says so in writing:

• changes how the tool is used, for example by adding disclosure, a human check, or logging
• pauses or narrows the tool
• stops using it, or decides not to expand it
• goes ahead unchanged, but with a written reason tied to the principles. This counts as a weaker result and gets reported as one.

Output: A one-page case note covering what was found, what was decided, and who decided. It goes in the adopter/pilot register. The note gets a review date 6 months out.

Pilot B: Public service desk offering an AI service to residents

Host type: One public-facing civic desk, such as a library, city or county office, or community service point, that offers or is about to offer one AI-assisted service to the public. Examples are a help chatbot, an AI search or reference aid, or a form assistant.

Named owner: One staff member who is responsible for that service. There is also at least one public reviewer who uses the service but doesn't work on it, which gives the pilot the public side of the hybrid model.

Length: About 30–45 days.

Steps:

1. Map. Describe the service, who uses it, what it may and may not do, and who can switch it off.
2. Measure. The owner checks it against the principles. Separately, the public reviewer uses the service and writes down what felt unclear, unfair, or wrong.
3. Decide. The owner records a decision. The public reviewer's notes are attached, whether or not the owner agrees with them.

Principles it mainly tests:

• 1: No Systematic Deception. Is it clear to residents that it's AI, and what its limits are?
• 5: Digital Does Not Override Physical. Can residents still reach a person and a non-digital path?
• 6: Named Human Responsibility. Who answers when it's wrong?
• 7: Limits on Coercion and Surveillance. What does it collect or keep about residents?
• 8: Protect Human Creative Agency. Does it help people find and learn, or just steer them?

What counts as a changed decision. The owner does at least one of these because of the review:

• adds or improves disclosure, a human fallback, or data limits
• narrows what the service does or who it's offered to
• delays or declines launch, or retires the service
• proceeds unchanged, with a written reason that answers the public reviewer's notes. This counts as a weaker result.

Output: A one-page case note with the owner's decision and the public reviewer's notes. It goes in the register, gets a review date, and a plain-language version can be published once the host agrees.

Rules for both pilots

• Before any host is approached, the approach needs the initiative's go-ahead.
• The host decides what can be published. Nothing about a host goes public without its written OK.
• The pilot is a review, not a certification. Hosts may not claim they are "certified" or "compliant" with Constitution for Human Creation.
• If a pilot finds that a principle is unworkable or unclear, that gets logged as input for amending the principles.