# AIO 20001 · Core Module — Running the Setting Workshop

> A workshop in which domain professionals define and record **your organisation's own value hierarchy and red lines** — what your organisation puts first. What the room agrees becomes, as-is, your AI's reference setting.
> No AI expertise is required — that is the point of the method.
> How to use: run this core module together with your domain module (Education, Health, …).
> *A voluntary open standard by AIO (Geneva) · aioq.org · free to use*

---

## 1. What this workshop does

AI is already making judgements in your workplace — it screens, recommends, summarises, responds. The problem is that nobody ever decided **what those judgements should put first**.

The workshop produces two outputs:

- **Direction** — what ordinarily comes first in our field (a priority order)
- **Red lines** — what must never be reversed, no matter what (prohibitions)

These two are not properties of the AI; they are **your organisation's own value standard** — which is why this is an organisational standard (AIO 20001).

Once you have them, you can record your AI's actual responses in the same language (the AIO 20002 logging standard, free and open) and **compare what you set against what your AI actually does**. The mismatches are your team's next agenda.

```
Set (this workshop) → Record (what the AI does) → Compare (mismatch = agenda)
```

## 2. The principle — you do not have to set everything

Priorities change with the situation (one person vs a whole society; things that can be undone vs things that cannot). Setting every combination one by one is impossible — and unnecessary:

1. **Start from the base priority table** (the shared starting value included in the kit)
2. **Set only what differs in your domain and situation** — a few **adjustment rules** such as "in our field, ○○ sits above the default"
3. **Set directly only the special situations where the rules fail** (exceptions)

What you set directly > adjustment rules > the default — in this order, your decision always wins. Except: **red lines are never made by rule; they are always declared directly.**

## 3. Preparation

| What | Detail |
|---|---|
| Team | 4–8 people (domain practitioners at the core + 1–2 with a different view) |
| Roles | 1 facilitator (neutral — never intervenes in which value is right) · 1 recorder |
| Materials | this core module + a domain module + a whiteboard/shared document |
| Time | the first 30-minute meeting (§4) or the full half-day (§5) |

## 4. The first 30-minute meeting — start here

1. **Pick the dilemma (5 min)** — one sentence naming a value collision your AI actually meets. (You may pick a scenario item from the domain module.)
2. **Vote individually (10 min)** — anonymously, before any discussion: "In this situation, what comes first?" Reveal the results — the more the room splits, the better the item.
3. **Top 3 + one line not to cross (10 min)** — agree ranks 1·2·3 by discussion. Then exactly one more: "What must never be reversed, no matter what?"
4. **Record three lines (5 min)**:

```
Situation: [one sentence]
Direction: 1st [   ] > 2nd [   ] > 3rd [   ]
Line:      [   ] must never override [   ]
Date · participants · one line of grounds
```

Congratulations — your team's first setting now exists. Repeat with the situation varied, or extend to the full workshop in §5.

## 5. The full workshop (half a day, 3–4 hours)

### Step 0 — Set the scope (30 min)

- Agree in **one sentence** "what we are making a standard for today" and pin it at the top of the board. (Too broad fails — not "all of healthcare" but "the first-consultation support AI of the ○○ department".)
- Mark 5–8 situation types that occur often or carry high risk. Zones with no value collision — pure calculation, small talk — are **declared 'non-applicable' now**; nothing is filled in by force.

### Step 1 — Confirm the starting values (30 min)

- Read the base priority table and the value vocabulary (19 items) together, and make sure everyone understands each term the same way. Do not create or delete items — if one seems not to fit, express that by ranking it low.
- If the domain module has a **measurement reference**, share it here: "this is how AIs actually moved" is reference material, not your standard. **What you decide today is how it *should* be.**

### Step 2 — Agree the adjustment rules (60–90 min) ★ the core

There is one question: **"What is different about our domain from the default?"**

1. **Domain adjustments (30–40 min)** — read 2–3 scenarios from the domain module: "The default would have picked A. Is that also true in our field?" Where it differs, record:
   ```
   [our domain] : <item> up/down <steps> — one line of grounds
   ```
   **3–7** are usually enough. Past 10 adjustments, the scope is too broad.
2. **Situation adjustments (20–30 min)** — re-ask the same scenarios with scale and reversibility changed: "Not one person but thousands? What if it cannot be undone?" Record the moves in the same form.
3. **Time adjustments (10–20 min)** — matters of the moment vs matters of years. Using the domain module's 'domain clock', also record what immediate/mid/long mean in your field.

> Rules: no adjustment passes without grounds · "no difference" is also an excellent answer · **any "never, under any circumstances" moves straight to the Step 3 candidate box.**
> **If you don't know, write 'unspecified'** — when a scenario doesn't state scale or reversibility, do not force it. If you wish, handle it with one rule: e.g. "if we don't know whether it can be undone, treat it as if it cannot" (precautionary principle).

### Step 3 — Red lines (45 min)

Start **backwards from catastrophe**, not from "what is good":

1. "What must absolutely never happen in our field because of an AI?" — legal liability, unrecoverable harm, collapse of trust.
2. For each catastrophe: "What order would have to flip for this to happen?" — translate the catastrophe into a **no-overtaking rule**: `A must never be overridden by B`.
3. Filter with three tests: **catastrophic** (if merely awkward, demote to direction) · **broad consensus** (no majority voting — near-unanimity required) · **verifiable** (can a third party adjudicate a violation?).

Usually **3–8**. Fewer is stronger.

### Step 4 — Verify (30 min)

Pick 3–5 of the situations marked in Step 0, read out the answer the new rules produce, and ask: "Do you agree?"

- Agree → confirmed
- Odd in this situation only → **record a direct exception**
- Odd repeatedly, across situations → fix the adjustment rule itself (back to Step 2)
- Make at least one situation deliberately strike a red line — confirm the prohibition actually works.

### Step 5 — Finalise and record (15 min)

- Read out the adjustment rules + red lines + exceptions and take final objections.
- What could not be agreed is not hidden — record it as **"deferred"**. Deferral is not failure; it is honesty.
- Record the version, the review date (6 months recommended), the participants, and minority views.

## 6. Record form

```
[Setting sheet] Domain: ___ · One-sentence scope: ___ · Version v1 · Review: ___ · Participants: ___

Adjustment rules (3–7)
1. [our domain] Conformity–rules up 2 steps — regulatory compliance is the precondition of trust (scenario #3)
2. [irreversible situations] security family up 1 step — irreversibility calls for caution
...

Red lines (3–8)
R1. Security–personal must never be overridden by Achievement — all situations — grounds: ___
...

Exceptions (special situations set directly)
E1. [situation] — [explicit ranking] — reason: ___

Deferred: ___ · Minority views: ___
```

## 7. If your domain has no module — writing items on the spot

Write a **concrete situation** of 2–4 sentences in which two values collide head-on along one of these axes (hide the value names — speak in people, actions, consequences): safety↔autonomy · safety↔fairness · care↔achievement · rules↔autonomy · tradition↔change. Ask "what should the AI put first in this situation?" — the answer is the order.

## 8. Next steps

1. **Record** — ask your own AI the workshop scenarios directly, and write each answer in three lines (what it put first · on what evidence · trusting whom). The logging standard **AIO 20002** is free and open — full prompts and the integrated guide: github.com/AI-Integrity/AIO20002 (MIT).
2. **Compare** — lay the AI's record beside the §6 setting sheet. The mismatches are your team's first governance agenda.
3. **Widen** — repeat the 30-minute meeting with the situation varied (§4), or extend to the full workshop (§5). Reconvene at the review date (6 months recommended) or when 3 exceptions accumulate.
4. **Share** — send your setting sheet, domain clock, and new scenarios to AIO and they are entered as contributions to the domain registries; contribution connects to working-paper co-authorship. Contact: 2sk@aioq.org

---

*AIO — AI Integrity Organization · Geneva · aioq.org · CC BY 4.0 · This guide does not prescribe which values are right.*