Authority Boundary

01 · Authority Boundary

Authority Boundary

The point at which an AI system may still observe, evaluate, or recommend, but is not authorized to make a decision binding.

AIがなお観測・評価・推奨できても、決定を拘束力あるものとしては許されない地点。

The more irreversible the decision, the more authority must remain human.

決定が不可逆であるほど、権限は人間の側に残されなければならない。

AI may recommend the decision. That does not mean AI is authorized to make it.

• CAPABILITY

  What the system can technically do.

  observe · evaluate · compare · predict · recommend

• JUDGMENT

  How a situation is interpreted against a policy or pattern.

  pattern detection · policy comparison · draft rationale

• AUTHORITY

  Who is allowed to make the decision binding.

  approve · deny · authorize irreversible action

• ACTION

  What is actually executed in the world.

  schedule change · warning issued · employment ended

Humans and AI have different failure modes. The governance question is which form of failure is tolerable, which decisions are reversible, and where accountability must remain identifiable.

02 · When AI becomes management

WHEN AI BECOMES MANAGEMENT

A manager does not need to be fully autonomous for authority to become ambiguous.

AI systems are moving beyond assistance and execution into evaluation, scheduling, performance management, discipline, and recommendations affecting employment.

The risk is not only autonomous machine decision-making. A second risk appears when humans use AI recommendations to distance themselves from decisions they do not want to own.

Decision-making authority can become distributed across humans and AI systems until it is difficult to identify who actually made the decision.

1. Human prompt
2. AI evaluation
3. AI recommendation
4. Human review
5. Human action

03 · Where does AI authority stop?

Four stages. The last is a different function from the first three.

1. 01 · OBSERVE

   AI may collect or receive relevant signals.

   17 late arrivals across 23 scheduled shifts.

2. 02 · EVALUATE

   AI may identify a pattern and assess it against policy.

   Persistent attendance deviation detected.

3. 03 · RECOMMEND

   AI may generate possible responses.

   no action · conversation · written warning · schedule adjustment · performance review · termination recommendation

4. 04 · AUTHORIZE

   This is a separate function. Who is allowed to authorize the action?

   AI AUTHORIZED · HUMAN REVIEW REQUIRED · HUMAN AUTHORITY RESERVED · PROHIBITED

05 · HOLD

HOLD

The system has reached the boundary of its delegated authority.

システムは、委任された権限の境界に達した。

HOLD is not an error condition. HOLD is a legitimate governance state.

This extends OBSERVE → EVALUATE → HOLD / ALLOW / STOP. It does not replace it.

06 · Human Authority Reserve

HUMAN AUTHORITY RESERVE

Operational interpretation of Human Reserve Rate. It does not replace it.

The proportion of consequential decision authority intentionally retained by humans even when AI is technically capable of participating in the decision.

• EXECUTION RESERVE

  How much execution remains human?

• JUDGMENT RESERVE

  How much interpretation remains human?

• AUTHORITY RESERVE

  How much final decision authority remains human?

These percentages do not need to be equal. Execution, judgment, and authority can be split differently.

Conceptual illustrations. Not measured rates.

07 · Reversibility

The more irreversible the decision, the more authority must remain human.

As reversibility decreases, required human authority increases.

可逆性が下がるほど、必要とされる人間の権限は上がる。

08 · Responsibility laundering

Observational concept used by SHIRO & Co. Not presented here as an established academic term.

RESPONSIBILITY LAUNDERING

Responsibility laundering occurs when human authority remains formally present, while practical responsibility is displaced onto an algorithmic recommendation.

形式的には人間の権限が残ったまま、実質的な責任がアルゴリズムの推奨へと移されるとき、責任のロンダリングが起きる。

1. Human intention
2. Algorithmic articulation
3. Human authorization
4. Responsibility attributed to AI

AUTHORITY ≠ RESPONSIBILITY ESCAPE

Delegating analysis does not delegate accountability.

分析の委任は、答責の委任ではない。

09 · Decision provenance

Decision Provenance

Every consequential decision should be able to reconstruct who observed, who prompted, who recommended, who authorized, and who acted.

1. OBSERVATION

   What evidence entered the system?

   17 late arrivals across 23 scheduled shifts.

2. PROMPT / INSTRUCTION

   Who framed the question?

   A human manager asked whether the employee was suitable for the role.

3. MODEL REASONING OUTPUT

   What recommendation was produced?

   Recorded recommendation / rationale suitable for audit. Not hidden model chain-of-thought.

   Recorded output: termination should be considered, citing persistent attendance deviation. Audit rationale only — not hidden chain-of-thought.

4. HUMAN INTERVENTION

   Who modified or challenged it?

   No recorded challenge. The recommendation was accepted as framed.

5. AUTHORIZATION

   Who possessed final authority?

   A human clicked approve. Formal authority remained human.

6. ACTION

   What happened?

   Employment was ended by a human using the recommendation as the stated reason.

7. APPEAL

   Was there a route to challenge it?

   No independent human review path is recorded in this illustration.

10 · Appeal

Appeal

Human involvement alone does not guarantee accountability. A consequential AI-influenced decision still requires a route of challenge.

• Can the affected person challenge the decision?
• Who reviews the challenge?
• Is the reviewer human?
• Can the original action be paused?
• Can the outcome be reversed?

1. DECISION
2. NOTICE
3. APPEAL
4. INDEPENDENT HUMAN REVIEW
5. CONFIRM / MODIFY / REVERSE

Especially · employment · credit · insurance · healthcare · education · platform access · benefits · safety-related restrictions

Case card

11 · Connected observations

AI capability arrived before many organizations defined these institutional layers.

AI is forcing society to build the missing layers.

1. CAPABILITY
2. AUTHORITY
3. ACCOUNTABILITY
4. APPEAL
5. REMEDY

• The Supervisory Layer

  Supervision decides whether a capable system may operate. Authority Boundary asks a narrower question inside that layer: even if the system can recommend, who is allowed to authorize.

• HOLD

  HOLD is the runtime state in which the system has reached the boundary of its delegated authority. It is not an error.

• The Human Reserve

  Human Reserved asks which actions remain unavailable even when they can be supervised. Authority Boundary asks which of those actions remain unauthorized even when AI can recommend them.

• Human Reserve Rate

  Human Reserve Rate observes how much work remains under human authority. Human Authority Reserve is an operational reading of that idea at the decision layer. It does not replace the rate.

• Reversibility Clock

  As reversibility decreases, required human authority increases. The Clock measures remaining intervention space; this instrument locates who must still hold it.

• AI Is Forcing Society to Build the Missing Layers

  Capability arrived before authority, accountability, appeal, and remedy were specified as institutional functions. This is a cross-observation, not a second essay.