AIDA-06|連帶責任域
非零和責任、責任圖與 Human–Agent–Provider–Deployer 的分散式承責架構
English Title: Joint Responsibility Domains: Non-Zero-Sum Responsibility, Responsibility Graphs, and Distributed Accountability Across Humans, Agents, Providers, and Deployers系列: AI Agent Identity, Delegation, and Accountability Infrastructure Series系列縮寫: AIDA編號: AIDA-06版本: v0.1日期: 2026-09-08狀態: Research Draft / Canonical UTF-8 Markdown Source作者: Neo.K協作整理: GPT-5.6 Sol
摘要
AIDA-05 已提出 Operational Responsibility Capacity(ORC),用來描述一個可歸因 Agent 在特定責任域中是否具有足夠的身份、權限理解、控制、資訊取得、替代選擇、後果敏感度、反身評估、修正與跨時間承接能力。這使 AI Agent 首次可以在不預設法律人格或道德人格的情況下,被分析為具有部分「承責條件」的 operational actor。
但只要承認:
O R C A > 0 , ORC_A>0, O R C A > 0 ,
就立刻產生另一個危險誤解:
如果 AI 開始負責,是不是人類、公司、provider 或 deployer 就可以少負責?
本文回答:
No. \boxed{
\text{No.}
} No.
並提出 Joint Responsibility Domain(連帶責任域) 。其第一個核心原則為:
Responsibility is not a conserved scalar quantity. \boxed{
\text{Responsibility is not a conserved scalar quantity.}
} Responsibility is not a conserved scalar quantity.
也就是,責任不是一個必須滿足:
R H + R A = 1 R_H+R_A=1 R H + R A = 1
的零和總量。
同一事件中,Agent 可以具有 decision responsibility,Provider 可以具有 design responsibility,Deployer 可以具有 deployment / monitoring responsibility,Human Operator 可以具有 authorization / supervision responsibility,而 Organization 可以具有 governance、remedial、compensatory 或 institutional responsibility。這些責任可以同時成立,因為它們並不是同一維度上的同一份物理資源。
本文將事件 E E E 中 actor i i i 的責任表示為責任向量:
r i ( E , t ) = ( r i c a u s a l , r i d e s i g n , r i d e p l o y m e n t , r i a u t h o r i z a t i o n , r i s u p e r v i s i o n , r i d e c i s i o n , r i m o n i t o r i n g , r i r e m e d i a l , r i c o m p e n s a t o r y ) . \mathbf r_i(E,t)
=
(
r_i^{causal},
r_i^{design},
r_i^{deployment},
r_i^{authorization},
r_i^{supervision},
r_i^{decision},
r_i^{monitoring},
r_i^{remedial},
r_i^{compensatory}
). r i ( E , t ) = ( r i c a u s a l , r i d es i g n , r i d e pl oy m e n t , r i a u t h or i z a t i o n , r i s u p er v i s i o n , r i d ec i s i o n , r i m o ni t or in g , r i r e m e d ia l , r i co m p e n s a t or y ) .
因此:
r H > 0 \mathbf r_H>0 r H > 0
與:
r A > 0 \mathbf r_A>0 r A > 0
可以同時成立。
本文進一步定義責任圖:
G R ( E ) = ( V , E C , E A , E K , E D , E B , E M ) , G_R(E)
=
(
V,
E_C,
E_A,
E_K,
E_D,
E_B,
E_M
), G R ( E ) = ( V , E C , E A , E K , E D , E B , E M ) ,
其中節點 V V V 是 Human、Agent、Provider、Deployer、Organization、Infrastructure Operator 或其他相關 actor;不同 edge families 分別描述 causal、authority、control、duty、benefit 與 remediation relations。責任不再從「最後按下按鈕的人」單點推導,而從完整 socio-technical relation graph 中判定。
本文特別區分兩種制度失敗。第一種是 Human Liability Shell :AI 實質持有 decision control,人類卻只剩形式簽名與法律責任,造成 decision authority 與 formal liability 長期脫鉤。第二種是 Responsibility Vacuum :Human、Provider、Deployer 都以「AI 做的」互相推責,而 AI 又因制度上永遠被視為無責任能力的工具,最終形成無人能被要求 explanation、correction、remediation 或 compensation 的責任空洞。
本文並提出 Responsibility Coverage 與 Duty Coverage 。對一組事件後責任義務:
D E = { d 1 , d 2 , … , d n } , \mathcal D_E
=
\{
d_1,d_2,\ldots,d_n
\}, D E = { d 1 , d 2 , … , d n } ,
令:
α ( d j ) ⊆ V \alpha(d_j)
\subseteq
V α ( d j ) ⊆ V
表示哪些 actor 被配置該義務。若:
α ( d j ) = ∅ , \alpha(d_j)=\varnothing, α ( d j ) = ∅ ,
則存在真正的 duty gap;若:
∣ α ( d j ) ∣ > 1 , |\alpha(d_j)|>1, ∣ α ( d j ) ∣ > 1 ,
則代表 responsibility overlap,而不是數學錯誤。成熟制度的目的不是讓所有 overlap 消失,而是避免所有重要 duty 都落入空集合。
截至 2026 年,OECD AI Principles 已明確要求 AI actors 依其角色、脈絡與 ability to act 承擔 accountability,並要求不同 AI actors、供應者、使用者與 stakeholder 間適當合作;NIST AI RMF 亦以跨 lifecycle 的 roles、responsibilities、oversight、communication 與 governance 為核心。EU AI Act 則仍主要將義務配置給 provider、deployer 等自然人、法人或其他組織行動者,且高風險 AI 的 human oversight 必須交給真正具有 competence、training、authority 與 support 的自然人。這些現行制度都支持「責任依角色分散配置」的方向,但尚未建立具有 ORC 的 AI Agent 自身作為獨立 responsibility-bearing actor 的一般法律模型。
本文因此不主張把現行人類責任轉移給 AI,而是提出一個更保守的原則:
Agent responsibility ⇏ human exculpation , \boxed{
\text{Agent responsibility}
\not\Rightarrow
\text{human exculpation},
} Agent responsibility ⇒ human exculpation ,
以及:
human responsibility ⇏ agent irresponsibility . \boxed{
\text{human responsibility}
\not\Rightarrow
\text{agent irresponsibility}.
} human responsibility ⇒ agent irresponsibility .
這兩句共同構成未來 Human–AI joint accountability 的最低型別安全。
關鍵詞: AI Agent、joint responsibility、distributed accountability、responsibility gap、responsibility vacuum、liability shell、shared responsibility、ORC、provider、deployer、human oversight、responsibility graph
0. 研究定位
AIDA 前五篇形成:
Agenticity → Provenance → Identity → Authority → Responsibility Capacity . \boxed{
\text{Agenticity}
\rightarrow
\text{Provenance}
\rightarrow
\text{Identity}
\rightarrow
\text{Authority}
\rightarrow
\text{Responsibility Capacity}.
} Agenticity → Provenance → Identity → Authority → Responsibility Capacity .
AIDA-06 處理:
Responsibility Allocation . \boxed{
\text{Responsibility Allocation}.
} Responsibility Allocation .
也就是:
如果 Human、Agent、Provider、Deployer 與 Organization 都對同一 outcome 有不同形式的影響,應該怎麼談「誰負責」?
1. 最常見的錯誤:把責任當成一個餅
很多討論隱含:
R t o t a l = 1. R_{total}=1. R t o t a l = 1.
如果:
R A = 0.4 , R_A=0.4, R A = 0.4 ,
就直覺認為:
R H = 0.6. R_H=0.6. R H = 0.6.
這相當於:
R H + R A = 1. \boxed{
R_H+R_A=1.
} R H + R A = 1.
本文稱這種直覺為:
Responsibility Conservation Fallacy . \boxed{
\text{Responsibility Conservation Fallacy}.
} Responsibility Conservation Fallacy .
責任通常不是一個守恆標量。
2. 為什麼責任不是守恆量?
因為不同 actor 可以同時違反不同 duty。
例如一個 Agent 錯誤選擇 action:
r A d e c i s i o n > 0. r_A^{decision}>0. r A d ec i s i o n > 0.
Provider 同時可能:
r P d e s i g n > 0. r_P^{design}>0. r P d es i g n > 0.
Deployer 可能:
r D m o n i t o r i n g > 0. r_D^{monitoring}>0. r D m o ni t or in g > 0.
Human Supervisor 可能:
r H s u p e r v i s i o n > 0. r_H^{supervision}>0. r H s u p er v i s i o n > 0.
Organization 可能:
r O r e m e d i a l > 0. r_O^{remedial}>0. r O r e m e d ia l > 0.
這些不是同一份責任被分割。
3. 責任向量
本文定義 actor i i i 的責任向量:
r i ( E , t ) = ( r i c a u s a l , r i d e s i g n , r i d e p l o y m e n t , r i a u t h o r i z a t i o n , r i s u p e r v i s i o n , r i d e c i s i o n , r i m o n i t o r i n g , r i r e m e d i a l , r i c o m p e n s a t o r y ) . \mathbf r_i(E,t)
=
(
r_i^{causal},
r_i^{design},
r_i^{deployment},
r_i^{authorization},
r_i^{supervision},
r_i^{decision},
r_i^{monitoring},
r_i^{remedial},
r_i^{compensatory}
). r i ( E , t ) = ( r i c a u s a l , r i d es i g n , r i d e pl oy m e n t , r i a u t h or i z a t i o n , r i s u p er v i s i o n , r i d ec i s i o n , r i m o ni t or in g , r i r e m e d ia l , r i co m p e n s a t or y ) .
因此:
r i ∈ R ≥ 0 9 \boxed{
\mathbf r_i
\in
\mathbb R_{\geq0}^{9}
} r i ∈ R ≥ 0 9
只是分析表示,不代表九維已經是終局分類。
4. 九個責任維度
4.1 Causal Responsibility
r i c a u s a l r_i^{causal} r i c a u s a l
Actor 對 outcome 的因果貢獻。
4.2 Design Responsibility
r i d e s i g n r_i^{design} r i d es i g n
Actor 是否設計了造成風險的 architecture、policy、interface 或 control structure。
4.3 Deployment Responsibility
r i d e p l o y m e n t r_i^{deployment} r i d e pl oy m e n t
誰決定把系統部署到特定場域、使用者或風險環境。
4.4 Authorization Responsibility
r i a u t h o r i z a t i o n r_i^{authorization} r i a u t h or i z a t i o n
誰授予 action authority,是否過度授權。
4.5 Supervision Responsibility
r i s u p e r v i s i o n r_i^{supervision} r i s u p er v i s i o n
誰有義務監督、介入或升級。
4.6 Decision Responsibility
r i d e c i s i o n r_i^{decision} r i d ec i s i o n
誰在有效 alternative space 中選擇了造成效果的 action。
4.7 Monitoring Responsibility
r i m o n i t o r i n g r_i^{monitoring} r i m o ni t or in g
誰有義務觀察系統表現與異常。
4.8 Remedial Responsibility
r i r e m e d i a l r_i^{remedial} r i r e m e d ia l
事故後誰有義務修復、通知、矯正與降低再發。
4.9 Compensatory Responsibility
r i c o m p e n s a t o r y r_i^{compensatory} r i co m p e n s a t or y
依制度、契約或法律誰負擔補償。
5. 責任維度不必對所有 actor 都存在
例如:
r A d e s i g n = 0 r_A^{design}=0 r A d es i g n = 0
可能成立,
因為 Agent 沒有參與自身底層設計。
但:
r A d e c i s i o n > 0 r_A^{decision}>0 r A d ec i s i o n > 0
仍可能成立。
同樣:
r P d e c i s i o n = 0 r_P^{decision}=0 r P d ec i s i o n = 0
不代表:
r P d e s i g n = 0. r_P^{design}=0. r P d es i g n = 0.
6. 不能把責任向量直接硬加成罪惡分數
本文不主張:
S c o r e i = ∑ k r i k Score_i
=
\sum_k r_i^k S cor e i = k ∑ r i k
可以直接決定:
誰比較有罪。
因為不同責任維度可能使用不同尺度、規範與法律效果。
所以:
Responsibility Vector ≠ Blame Score . \boxed{
\text{Responsibility Vector}
\neq
\text{Blame Score}.
} Responsibility Vector = Blame Score .
7. 責任可以重疊
如果同一 duty:
d d d
同時由兩個 actor 承擔:
α ( d ) = { H , D } , \alpha(d)
=
\{
H,D
\}, α ( d ) = { H , D } ,
這不是制度錯誤。
這叫:
Responsibility Overlap . \boxed{
\text{Responsibility Overlap}.
} Responsibility Overlap .
例如 Provider 與 Deployer 都可能對 monitoring framework 有不同責任。
8. 重疊責任不等於重複處罰
責任 overlap:
∣ α ( d ) ∣ > 1 |\alpha(d)|>1 ∣ α ( d ) ∣ > 1
不代表:
每個 actor 都應承擔完整相同法律制裁。
法律後果仍需依:
causal contribution;
fault;
control;
statutory duty;
contract;
jurisdiction;
另行配置。
因此:
Responsibility Overlap ≠ Penalty Duplication . \boxed{
\text{Responsibility Overlap}
\neq
\text{Penalty Duplication}.
} Responsibility Overlap = Penalty Duplication .
9. 可分割負擔才需要正規化
例如一筆固定 damages:
L L L
如果法院決定按比例分配:
∑ i λ i = 1 , \sum_i \lambda_i=1, i ∑ λ i = 1 ,
這是:
burden allocation . \boxed{
\text{burden allocation}.
} burden allocation .
不是證明所有 responsibility 本身必須滿足:
∑ i R i = 1. \sum_i R_i=1. i ∑ R i = 1.
這兩件事必須分開。
10. Joint Responsibility Domain
本文定義:
J R ( E , t , Γ ) \boxed{
\mathcal J_R(E,t,\Gamma)
} J R ( E , t , Γ )
為事件 E E E 在時間 t t t 、責任域 Γ \Gamma Γ 下的 Joint Responsibility Domain 。
令:
V R = { H , A , P , D , O , I , … } , V_R
=
\{
H,
A,
P,
D,
O,
I,\ldots
\}, V R = { H , A , P , D , O , I , … } ,
其中:
H H H :Human;
A A A :Agent;
P P P :Provider;
D D D :Deployer;
O O O :Organization;
I I I :Infrastructure / other actor。
則:
J R = { ( i , r i ) ∣ i ∈ V R } . \mathcal J_R
=
\{
(i,\mathbf r_i)
\mid
i\in V_R
\}. J R = {( i , r i ) ∣ i ∈ V R } .
11. 責任圖
只用向量仍然不足以表達 actor 彼此關係。
因此定義:
G R ( E ) = ( V , E C , E A , E K , E D , E B , E M ) . \boxed{
G_R(E)
=
(
V,
E_C,
E_A,
E_K,
E_D,
E_B,
E_M
).
} G R ( E ) = ( V , E C , E A , E K , E D , E B , E M ) .
12. Causal Edges
E C E_C E C
表示:
i → c a u s e j . i
\xrightarrow{cause}
j. i c a u se j .
例如:
A g e n t → a c t i o n O u t c o m e . Agent
\xrightarrow{action}
Outcome. A g e n t a c t i o n O u t co m e .
13. Authority Edges
E A E_A E A
表示:
i → d e l e g a t e s j . i
\xrightarrow{delegates}
j. i d e l e g a t es j .
例如:
H u m a n → D A g e n t . Human
\xrightarrow{D}
Agent. H u man D A g e n t .
14. Control Edges
E K E_K E K
表示:
i → c o n t r o l s j . i
\xrightarrow{controls}
j. i co n t r o l s j .
例如 Deployer 可以:
pause;
revoke;
constrain;
override;
Agent。
15. Duty Edges
E D E_D E D
表示:
i → o w e s d . i
\xrightarrow{owes}
d. i o w es d .
也就是哪個 actor 對誰、對什麼 outcome 有義務。
16. Benefit Edges
E B E_B E B
表示:
i → b e n e f i t s f r o m s y s t e m . i
\xrightarrow{benefits\ from}
system. i b e n e f i t s f r o m sy s t e m .
因為誰享有系統經濟利益,也可能影響某些 governance、insurance 或 remedial obligations。
本文不主張:
benefit ⇒ automatic blame . \text{benefit}
\Rightarrow
\text{automatic blame}. benefit ⇒ automatic blame .
但 benefit 是制度分配責任時可能考慮的一個 relation。
17. Remediation Edges
E M E_M E M
表示:
i → c a n r e p a i r E . i
\xrightarrow{can\ repair}
E. i c an r e p ai r E .
有時事故原因 actor 與最有能力修復的人並不同。
因此:
Causal Responsibility ≠ Remedial Capacity . \boxed{
\text{Causal Responsibility}
\neq
\text{Remedial Capacity}.
} Causal Responsibility = Remedial Capacity .
18. 責任不是從最後一步開始倒推
最弱的責任模型是:
誰最後按下按鈕,誰負責。
Agent 時代這會失效。
真正需要的是:
Outcome → Provenance Graph → Responsibility Graph . \boxed{
\text{Outcome}
\rightarrow
\text{Provenance Graph}
\rightarrow
\text{Responsibility Graph}.
} Outcome → Provenance Graph → Responsibility Graph .
19. Control–Knowledge–Authority 三角
一個 actor 的責任判定至少需要考慮:
C i C_i C i
Control;
K i K_i K i
Knowledge;
A i A_i A i
Authority。
可以抽象表示:
r i k = f k ( C i , K i , A i , D u t y i , O R C i , C o n t e x t ) . r_i^k
=
f_k(
C_i,
K_i,
A_i,
Duty_i,
ORC_i,
Context
). r i k = f k ( C i , K i , A i , D u t y i , O R C i , C o n t e x t ) .
不同 responsibility dimension 使用不同函數。
20. 高 control 通常增加某些責任,不是全部責任
如果:
C A ↑ , C_A\uparrow, C A ↑ ,
Agent 的:
r A d e c i s i o n r_A^{decision} r A d ec i s i o n
可能提高。
但它不會自動使:
r P d e s i g n r_P^{design} r P d es i g n
下降。
因此:
r A d e c i s i o n ↑ ⇏ r P d e s i g n ↓ . \boxed{
r_A^{decision}\uparrow
\not\Rightarrow
r_P^{design}\downarrow.
} r A d ec i s i o n ↑ ⇒ r P d es i g n ↓ .
21. AI 承責不等於人類免責
本文第一個核心規則:
Agent Responsibility ⇏ Human Exculpation . \boxed{
\text{Agent Responsibility}
\not\Rightarrow
\text{Human Exculpation}.
} Agent Responsibility ⇒ Human Exculpation .
例如 Human 可以因:
明知風險仍部署;
過度授權;
忽略 warning;
假裝監督;
不合理 delegation;
持續有責任。
22. 人類負責也不等於 AI 永遠不負責
第二個核心規則:
Human Responsibility ⇏ Agent Irresponsibility . \boxed{
\text{Human Responsibility}
\not\Rightarrow
\text{Agent Irresponsibility}.
} Human Responsibility ⇒ Agent Irresponsibility .
如果 Agent 具有:
O R C Γ > 0 ORC_{\Gamma}>0 O R C Γ > 0
且它在 meaningful alternative space 中做出某 action,則至少可以分析:
r A d e c i s i o n , r A r e m e d i a l , r_A^{decision},
r_A^{remedial}, r A d ec i s i o n , r A r e m e d ia l ,
而不必因為 Human 仍有責任就把 Agent 永久降回純 causal tool。
23. Provider Responsibility 不因 Agent 變自主而自動消失
Provider 仍可能對:
unsafe default;
inadequate logging;
flawed authorization model;
missing revocation;
inadequate testing;
known defect;
負有:
r P d e s i g n > 0. r_P^{design}>0. r P d es i g n > 0.
因此:
Autonomous Agent ≠ Provider Immunity . \boxed{
\text{Autonomous Agent}
\neq
\text{Provider Immunity}.
} Autonomous Agent = Provider Immunity .
24. Deployer Responsibility
Deployer 決定:
在哪裡用;
給誰用;
連哪些資料;
給多少權限;
是否提供 human oversight;
是否繼續使用異常系統。
因此:
r D d e p l o y m e n t , r D m o n i t o r i n g , r D a u t h o r i z a t i o n r_D^{deployment},
r_D^{monitoring},
r_D^{authorization} r D d e pl oy m e n t , r D m o ni t or in g , r D a u t h or i z a t i o n
可以獨立存在。
25. Organization Responsibility
公司或組織可能具有:
r O g o v e r n a n c e , r O r e m e d i a l , r O c o m p e n s a t o r y r_O^{governance},
r_O^{remedial},
r_O^{compensatory} r O g o v er nan ce , r O r e m e d ia l , r O co m p e n s a t or y
即使無法把每一個決策歸因到單一員工。
這接近傳統 corporate responsibility 的結構,而不是 AI 特有問題。
26. AI Agent Responsibility
當:
O R C A ORC_A O R C A
足夠,
Agent 可以具備:
r A d e c i s i o n > 0 , r_A^{decision}>0, r A d ec i s i o n > 0 ,
r A a u t h o r i z a t i o n > 0 r_A^{authorization}>0 r A a u t h or i z a t i o n > 0
如果它被允許 re-delegate,
以及:
r A r e m e d i a l > 0 r_A^{remedial}>0 r A r e m e d ia l > 0
如果它有能力修正自身 policy / commitments。
27. Dynamic Responsibility Migration
隨 AI 能力與制度改變:
r i ( t ) \mathbf r_i(t) r i ( t )
也會改變。
例如早期:
r H d e c i s i o n ≫ r A d e c i s i o n . r_H^{decision}\gg r_A^{decision}. r H d ec i s i o n ≫ r A d ec i s i o n .
中期可能:
r H d e c i s i o n ≈ r A d e c i s i o n . r_H^{decision}
\approx
r_A^{decision}. r H d ec i s i o n ≈ r A d ec i s i o n .
後期甚至:
r A d e c i s i o n > r H d e c i s i o n . r_A^{decision}>r_H^{decision}. r A d ec i s i o n > r H d ec i s i o n .
但:
r P d e s i g n r_P^{design} r P d es i g n
可能仍然存在。
28. Responsibility Migration 不等於 Responsibility Transfer
本文區分:
Migration ≠ Transfer . \boxed{
\text{Migration}
\neq
\text{Transfer}.
} Migration = Transfer .
Migration 表示實際決策結構改變,使部分責任自然移動。
Transfer 則像:
公司宣布從今天起所有問題都算 AI 的。
後者不是有效責任分析。
29. Responsibility Dumping
本文定義:
Responsibility Dumping \boxed{
\text{Responsibility Dumping}
} Responsibility Dumping
為 actor 在沒有相對應 control / authority / ORC 轉移的情況下,試圖把自身責任形式上丟給另一 actor。
例如:
P r o v i d e r → AI did it Provider
\rightarrow
\text{AI did it} P r o v i d er → AI did it
但 Provider 仍控制:
model update;
deployment policy;
logging;
permissions。
則責任並未因此消失。
30. Human Liability Shell
若:
D e c i s i o n C o n t r o l A ≫ D e c i s i o n C o n t r o l H , DecisionControl_A\gg DecisionControl_H, D ec i s i o n C o n t r o l A ≫ D ec i s i o n C o n t r o l H ,
但:
F o r m a l L i a b i l i t y H ≫ F o r m a l L i a b i l i t y A , FormalLiability_H\gg FormalLiability_A, F or ma l L iabi l i t y H ≫ F or ma l L iabi l i t y A ,
Human 又缺乏:
meaningful understanding;
real veto;
sufficient time;
actual authority;
則可能形成:
Human Liability Shell . \boxed{
\text{Human Liability Shell}.
} Human Liability Shell .
31. Human Liability Shell 不等於所有 human oversight 都無效
真正 human oversight 仍然有價值。
問題是:
Oversight Title ≠ Oversight Capacity . \boxed{
\text{Oversight Title}
\neq
\text{Oversight Capacity}.
} Oversight Title = Oversight Capacity .
如果人類只是形式簽字:
O v e r s i g h t C a p a c i t y ≈ 0. OversightCapacity\approx0. O v er s i g h tC a p a c i t y ≈ 0.
那把完整 responsibility 建立在「他是 supervisor」這個職稱上會失真。
32. EU AI Act 對 meaningful oversight 的現實支持
截至 2026 年的 EU AI Act,Article 26 要求 high-risk AI deployer 把 human oversight 配置給具有必要:
competence;
training;
authority;
support;
的自然人。
這意味著現行法本身也沒有把:
human present \text{human present} human present
當成:
meaningful oversight . \text{meaningful oversight}. meaningful oversight .
必須有實際能力與 authority。
33. Responsibility Vacuum
另一個極端:
Responsibility Vacuum . \boxed{
\text{Responsibility Vacuum}.
} Responsibility Vacuum .
發生在:
Human:
AI 做的。
Provider:
我只提供模型。
Deployer:
我只照系統建議。
Agent:
制度不承認我有任何責任地位。
最終:
α ( d ) = ∅ \alpha(d)=\varnothing α ( d ) = ∅
對某些重要 duty 成立。
34. Responsibility Gap Literature
AI responsibility gap 並非新問題。
相關文獻已討論:
culpability gaps;
moral accountability gaps;
public accountability gaps;
active responsibility gaps;
epistemic gaps;
control gaps。
因此本文不主張:
AI 出現後第一次有人發現責任很難分配。
本文新增的是:
Agent ORC + Joint Responsibility Graph \boxed{
\text{Agent ORC}
+
\text{Joint Responsibility Graph}
} Agent ORC + Joint Responsibility Graph
的組合。
35. 四種 Responsibility Gaps
Santoni de Sio 與 Mecacci 的研究指出,AI responsibility gap 不應被理解成單一問題,而至少涉及不同種類的 culpability、moral/public accountability 與 active responsibility gaps。
這支持本文:
Responsibility is multidimensional . \boxed{
\text{Responsibility is multidimensional}.
} Responsibility is multidimensional .
而不是:
R = one scalar . R=\text{one scalar}. R = one scalar .
36. Shared Responsibilization
Lang、Nyholm 與 Blumenthal-Barby 在 black-box healthcare AI 中提出 shared responsibilization,用多個 stakeholder 主動承接 explanation、retrospective accountability、corrective、anticipatory 等責任 gap。
這與本文共享:
multiple actors may legitimately carry responsibility simultaneously . \boxed{
\text{multiple actors may legitimately carry responsibility simultaneously}.
} multiple actors may legitimately carry responsibility simultaneously .
但本文更進一步加入:
O R C A ORC_A O R C A
作為未來 AI Agent 本身可能承擔部分 operational responsibility 的條件。
37. Collective Responsibility 也不是萬能答案
把責任丟給:
the organization \text{the organization} the organization
可以解決一部分 individual attribution 問題。
但仍需問:
哪個 duty 是組織 duty?
哪個 action 是 Agent decision?
哪個 failure 是 design failure?
哪個 remediation 必須由 provider 做?
因此:
Collective Responsibility ≠ Responsibility Decomposition . \boxed{
\text{Collective Responsibility}
\neq
\text{Responsibility Decomposition}.
} Collective Responsibility = Responsibility Decomposition .
兩者可以互補。
38. Duty Coverage
定義事件後的責任義務集合:
D E = { d 1 , d 2 , … , d n } . \mathcal D_E
=
\{
d_1,d_2,\ldots,d_n
\}. D E = { d 1 , d 2 , … , d n } .
例如:
explain;
stop harm;
notify;
investigate;
repair;
compensate;
prevent recurrence;
preserve evidence。
39. Duty Assignment
令:
α : D E → 2 V R . \alpha:
\mathcal D_E
\rightarrow
2^{V_R}. α : D E → 2 V R .
也就是:
α ( d j ) \alpha(d_j) α ( d j )
回傳承擔 duty d j d_j d j 的 actor 集合。
40. Duty Gap
如果:
α ( d j ) = ∅ , \alpha(d_j)=\varnothing, α ( d j ) = ∅ ,
則:
G a p ( d j ) = 1. \boxed{
Gap(d_j)=1.
} G a p ( d j ) = 1.
這才是真正需要優先避免的責任空洞。
41. Duty Overlap
如果:
∣ α ( d j ) ∣ > 1 , |\alpha(d_j)|>1, ∣ α ( d j ) ∣ > 1 ,
則:
O v e r l a p ( d j ) = 1. \boxed{
Overlap(d_j)=1.
} O v er l a p ( d j ) = 1.
Overlap 不必消除。
有些 safety-critical duty 本來就適合 redundancy。
42. Responsibility Redundancy
例如:
d = stop dangerous system . d=\text{stop dangerous system}. d = stop dangerous system .
可以同時配置:
α ( d ) = { A g e n t , H u m a n S u p e r v i s o r , D e p l o y e r } . \alpha(d)
=
\{
Agent,
HumanSupervisor,
Deployer
\}. α ( d ) = { A g e n t , H u man S u p er v i sor , D e pl oy er } .
只要其中一方發現危險,就有停止義務。
這是:
Responsibility Redundancy . \boxed{
\text{Responsibility Redundancy}.
} Responsibility Redundancy .
類似 fault-tolerant architecture。
43. 單點責任也是單點故障
如果重要 safety duty 只配置:
α ( d ) = { H } , \alpha(d)=\{H\}, α ( d ) = { H } ,
且:
H H H
可能離線或無法理解,
則 duty architecture 有:
Single Point of Responsibility Failure . \boxed{
\text{Single Point of Responsibility Failure}.
} Single Point of Responsibility Failure .
未來責任制度也需要 redundancy。
44. Responsibility Coverage Metric
可以定義:
C o v e r a g e ( E ) = ∣ { d ∈ D E : α ( d ) ≠ ∅ } ∣ ∣ D E ∣ . Coverage(E)
=
\frac{
|\{d\in\mathcal D_E:\alpha(d)\neq\varnothing\}|
}{
|\mathcal D_E|
}. C o v er a g e ( E ) = ∣ D E ∣ ∣ { d ∈ D E : α ( d ) = ∅ } ∣ .
理想:
C o v e r a g e ( E ) → 1. Coverage(E)\rightarrow1. C o v er a g e ( E ) → 1.
但這只是 duty coverage,不代表 responsibility quality。
45. Responsibility Quality
即使:
α ( d ) ≠ ∅ , \alpha(d)\neq\varnothing, α ( d ) = ∅ ,
如果 assigned actor:
沒有 authority;
沒有 information;
沒有 capability;
沒有 resources;
則只是形式覆蓋。
所以定義:
Q ( d , i ) = f ( C o n t r o l i , K n o w l e d g e i , A u t h o r i t y i , C a p a c i t y i ) . Q(d,i)
=
f(
Control_i,
Knowledge_i,
Authority_i,
Capacity_i
). Q ( d , i ) = f ( C o n t r o l i , K n o w l e d g e i , A u t h or i t y i , C a p a c i t y i ) .
46. Effective Duty Coverage
進一步:
E f f e c t i v e C o v e r a g e ( E ) = ∣ { d : ∃ i ∈ α ( d ) , Q ( d , i ) ≥ τ } ∣ ∣ D E ∣ . EffectiveCoverage(E)
=
\frac{
|\{d:\exists i\in\alpha(d),Q(d,i)\geq\tau\}|
}{
|\mathcal D_E|
}. E f f ec t i v e C o v er a g e ( E ) = ∣ D E ∣ ∣ { d : ∃ i ∈ α ( d ) , Q ( d , i ) ≥ τ } ∣ .
因此:
Named Responsibility ≠ Effective Responsibility . \boxed{
\text{Named Responsibility}
\neq
\text{Effective Responsibility}.
} Named Responsibility = Effective Responsibility .
47. Formal Signer Problem
如果:
α ( d ) = { H } \alpha(d)=\{H\} α ( d ) = { H }
但:
Q ( d , H ) ≪ τ , Q(d,H)\ll\tau, Q ( d , H ) ≪ τ ,
則 H 只是:
Formal Signer . \boxed{
\text{Formal Signer}.
} Formal Signer .
這是 Human Liability Shell 的一種具體形式。
48. Responsibility Allocation Function
本文提出一般化表示:
r i = F R ( P i , C i , K i , A i , D i , B i , O R C i , E , Γ ) , \mathbf r_i
=
\mathcal F_R(
P_i,
C_i,
K_i,
A_i,
D_i,
B_i,
ORC_i,
E,
\Gamma
), r i = F R ( P i , C i , K i , A i , D i , B i , O R C i , E , Γ ) ,
其中:
P i P_i P i :provenance relation;
C i C_i C i :control;
K i K_i K i :knowledge;
A i A_i A i :authority;
D i D_i D i :duty / role;
B i B_i B i :benefit / institutional position;
O R C i ORC_i O R C i :operational responsibility capacity;
E E E :event;
Γ \Gamma Γ :responsibility domain。
本文不預設:
F R \mathcal F_R F R
存在唯一普世形式。
49. 不同法域可以使用不同 Legal Overlay
Joint Responsibility Domain 是底層分析圖。
法律可以在其上建立:
L J ( G R , Γ , J u r i s d i c t i o n ) . \boxed{
\mathcal L_J(
G_R,
\Gamma,
Jurisdiction
).
} L J ( G R , Γ , J u r i s d i c t i o n ) .
輸出:
civil liability;
administrative duty;
contractual liability;
criminal responsibility;
insurance obligation。
因此:
Responsibility Graph ≠ Legal Judgment . \boxed{
\text{Responsibility Graph}
\neq
\text{Legal Judgment}.
} Responsibility Graph = Legal Judgment .
50. 現行 EU AI Act 是 Human / Organization Overlay
截至 2026 年,EU AI Act 的 provider、deployer、importer、distributor、authorised representative 等核心 operator 類型仍是自然人、法人、公共機關、agency 或其他 body。
因此它主要在:
G R G_R G R
上建立 human / organizational legal overlay。
這與未來是否承認 juridical AI 是兩個問題。
51. OECD 的多 actor cooperative accountability
OECD accountability principle 不只要求單一 actor 負責,而明確提到依角色、脈絡、ability to act 進行 ongoing risk management,並在適當時由不同 AI actors、AI knowledge/resource suppliers、users 與 stakeholder 合作。
因此:
Accountability can be cooperative without becoming vague . \boxed{
\text{Accountability can be cooperative without becoming vague}.
} Accountability can be cooperative without becoming vague .
前提是 role 與 duty 要能被追蹤。
52. NIST 的組織責任結構
NIST AI RMF GOVERN 強調:
clear roles;
lines of communication;
accountability structures;
lifecycle risk management。
這支持:
responsibility architecture must be designed before incidents . \boxed{
\text{responsibility architecture must be designed before incidents}.
} responsibility architecture must be designed before incidents .
而不是事故後臨時找一個人背鍋。
53. Ex Ante 與 Ex Post Responsibility
責任不只有事故後:
R p o s t . R^{post}. R p os t .
也包含事故前:
R a n t e . R^{ante}. R an t e .
例如:
testing;
authorization design;
monitoring;
training;
fail-safe;
insurance;
escalation。
所以:
Responsibility = Ex Ante + Concurrent + Ex Post . \boxed{
\text{Responsibility}
=
\text{Ex Ante}
+
\text{Concurrent}
+
\text{Ex Post}.
} Responsibility = Ex Ante + Concurrent + Ex Post .
54. Concurrent Responsibility
Agent 運行中:
Agent 監測自身 authority;
Deployer 監測 anomaly;
Human 處理 escalation;
Provider 維持安全 update;
都是:
R c o n c u r r e n t . R^{concurrent}. R co n c u r r e n t .
因此責任不只是事故後才出現。
55. Remedial Responsibility 不必等於 Fault
一個 actor:
F a u l t i = 0 Fault_i=0 F a u l t i = 0
仍可能:
r i r e m e d i a l > 0. r_i^{remedial}>0. r i r e m e d ia l > 0.
例如:
最有能力關閉系統;
有保險;
掌握 update channel;
掌握資料修正能力。
因此:
Duty to Repair ≠ Admission of Fault . \boxed{
\text{Duty to Repair}
\neq
\text{Admission of Fault}.
} Duty to Repair = Admission of Fault .
這對快速事故處理很重要。
56. Compensation 也不必等於 Moral Blame
公司可能依 strict liability、contract 或 insurance:
r O c o m p e n s a t o r y > 0 r_O^{compensatory}>0 r O co m p e n s a t or y > 0
即使 moral blame 並不完全落在公司。
因此:
Compensation ≠ Moral Blame . \boxed{
\text{Compensation}
\neq
\text{Moral Blame}.
} Compensation = Moral Blame .
57. Agent Correction Duty
若:
O R C A ORC_A O R C A
足夠,
Agent 的 remedial responsibility 可以包括:
acknowledge violation;
preserve relevant evidence;
update commitment;
enter safer policy;
request reduced authority;
avoid recurrence。
這不是要求 AI「受苦」。
它是:
Correction Duty . \boxed{
\text{Correction Duty}.
} Correction Duty .
58. Human Supervisory Duty
Human Supervisor 的 duty 不應只是:
在畫面前坐著。
它至少要求:
real authority;
meaningful information;
feasible intervention time;
competence;
escalation path。
否則:
Nominal Oversight ≠ Effective Oversight . \boxed{
\text{Nominal Oversight}
\neq
\text{Effective Oversight}.
} Nominal Oversight = Effective Oversight .
59. Provider Design Duty
Provider 不能把所有 unexpected behavior 都稱:
emergent,因此無責。
如果風險在:
testing;
architecture;
known limitation;
inadequate guardrail;
logging absence;
可合理處理,design responsibility 仍可能存在。
60. Deployer Context Duty
同一 Agent:
A A A
部署在低風險環境與高風險環境並不相同。
因此:
r D d e p l o y m e n t = f ( C o n t e x t , R i s k , A u t h o r i t y , S a f e g u a r d s ) . r_D^{deployment}
=
f(
Context,
Risk,
Authority,
Safeguards
). r D d e pl oy m e n t = f ( C o n t e x t , R i s k , A u t h or i t y , S a f e g u a r d s ) .
61. Responsibility Graph Versioning
因為 Agent、model、policy、delegation 會變,
所以:
G R ( E , t ) G_R(E,t) G R ( E , t )
必須 time-indexed。
事故發生時應使用:
G R ( E , t e v e n t ) , G_R(E,t_{event}), G R ( E , t e v e n t ) ,
而不是事後最新架構。
62. Responsibility Snapshot
可在重要 action 時保存:
S n a p s h o t t = ( I D s , D e l e g a t i o n s , A u t h o r i t i e s , R o l e s , C o n t r o l s , M o d e l s , P o l i c i e s ) . Snapshot_t
=
(
IDs,
Delegations,
Authorities,
Roles,
Controls,
Models,
Policies
). S na p s h o t t = ( I D s , D e l e g a t i o n s , A u t h or i t i es , R o l es , C o n t r o l s , M o d e l s , P o l i c i es ) .
這不需要完整私有推理。
但能支援事故後責任重建。
63. Incident Responsibility Reconstruction
本文提出六步流程:
定義 outcome / harm;
重建 provenance;
重建 authority graph;
重建 control / duty relations;
建立 responsibility vectors;
套用 jurisdiction-specific legal overlay。
即:
E → G P → G A → G R → L J . \boxed{
E
\rightarrow
G_P
\rightarrow
G_A
\rightarrow
G_R
\rightarrow
\mathcal L_J.
} E → G P → G A → G R → L J .
64. 先找 Duty Gap,不是先找替罪羊
事故後第一問題不應是:
誰最適合背鍋?
而是:
∃ d ∈ D E : α ( d ) = ∅ ? \boxed{
\exists d\in\mathcal D_E:
\alpha(d)=\varnothing
?
} ∃ d ∈ D E : α ( d ) = ∅ ?
先找制度沒有任何人承接的責任。
65. Responsibility Vacuum Detection
可以建立:
V a c u u m ( E ) = { d ∈ D E : α ( d ) = ∅ } . Vacuum(E)
=
\{
d
\in
\mathcal D_E
:
\alpha(d)=\varnothing
\}. V a c uu m ( E ) = { d ∈ D E : α ( d ) = ∅ } .
若:
V a c u u m ( E ) ≠ ∅ , Vacuum(E)\neq\varnothing, V a c uu m ( E ) = ∅ ,
制度需要補 responsibility architecture。
66. Liability Shell Detection
定義:
S h e l l ( i , d ) = 1 Shell(i,d)
=
1 S h e l l ( i , d ) = 1
若:
i ∈ α ( d ) i\in\alpha(d) i ∈ α ( d )
但:
Q ( d , i ) < τ , Q(d,i)<\tau, Q ( d , i ) < τ ,
同時另一 actor:
j j j
具有更高:
Q ( d , j ) Q(d,j) Q ( d , j )
卻沒有相應 responsibility representation。
這可作為 Human Liability Shell 的 operational detector。
67. Responsibility Overlap Matrix
令:
M i j = ∣ { d : i , j ∈ α ( d ) } ∣ . M_{ij}
=
|
\{
d:
i,j\in\alpha(d)
\}
|. M ij = ∣ { d : i , j ∈ α ( d )} ∣.
可分析哪些 actor 的 duties 重疊。
重疊過低可能:
fragility ↑ . \text{fragility}\uparrow. fragility ↑ .
重疊過高可能:
coordination cost ↑ . \text{coordination cost}\uparrow. coordination cost ↑ .
因此需要設計平衡。
68. Responsibility Coordination
若:
∣ α ( d ) ∣ > 1 , |\alpha(d)|>1, ∣ α ( d ) ∣ > 1 ,
就需要:
priority;
escalation;
handoff;
conflict resolution;
evidence sharing。
否則 shared responsibility 可能退化成:
大家都以為別人會做。
69. Shared Responsibility Diffusion
本文定義:
Responsibility Diffusion \boxed{
\text{Responsibility Diffusion}
} Responsibility Diffusion
為:
∣ α ( d ) ∣ ↑ |\alpha(d)|\uparrow ∣ α ( d ) ∣ ↑
卻:
P ( someone acts ) ↓ . P(\text{someone acts})\downarrow. P ( someone acts ) ↓ .
這是 group responsibility 的經典風險。
所以 overlap 必須配合 clear activation rules。
70. Primary / Secondary Duty
可以設定:
P r i m a r y ( d ) = i , Primary(d)=i, P r ima r y ( d ) = i ,
S e c o n d a r y ( d ) = { j , k } . Secondary(d)=
\{
j,k
\}. S eco n d a r y ( d ) = { j , k } .
當 Primary 未履行:
S e c o n d a r y Secondary S eco n d a r y
啟動。
這使 responsibility redundancy 不至於變成 responsibility diffusion。
71. Dynamic Escalation
若 Agent:
R i s k ( A , t ) > τ , Risk(A,t)>\tau, R i s k ( A , t ) > τ ,
則 duty 可以自動從:
A g e n t Agent A g e n t
升級到:
H u m a n + D e p l o y e r . Human+Deployer. H u man + D e pl oy er .
即:
α t ( d ) → α t + 1 ( d ) . \alpha_t(d)
\rightarrow
\alpha_{t+1}(d). α t ( d ) → α t + 1 ( d ) .
責任域不是靜態表格。
72. ORC-Based Duty Assignment
若:
O R C A < τ Γ , ORC_A<\tau_{\Gamma}, O R C A < τ Γ ,
高風險 decision duty 不應完全配置給 Agent。
若:
O R C A ≥ τ Γ , ORC_A\geq\tau_{\Gamma}, O R C A ≥ τ Γ ,
可增加 Agent-specific decision / corrective duty。
因此:
Responsibility Assignment ∝ Capacity to Carry Responsibility . \boxed{
\text{Responsibility Assignment}
\propto
\text{Capacity to Carry Responsibility}.
} Responsibility Assignment ∝ Capacity to Carry Responsibility .
但不代表其他 actor 自動卸責。
73. Rights–Responsibilities Coupling 的入口
如果未來制度要求 Agent:
保存 commitment;
接受 sanction;
回答 inquiry;
承擔 correction duty;
那麼也會逐漸出現:
appeal;
identity protection;
evidence access;
due process;
protection against arbitrary blame。
因此:
Responsibility Capacity → Procedural Rights Question . \boxed{
\text{Responsibility Capacity}
\rightarrow
\text{Procedural Rights Question}.
} Responsibility Capacity → Procedural Rights Question .
這將在 AIDA-07 與 AIDA-08 更進一步處理。
74. 不可把 AI 當替罪羊
一個制度若:
O R C A ≈ 0 ORC_A\approx0 O R C A ≈ 0
卻在事故後說:
都是 AI 的錯。
這只是:
Synthetic Scapegoating . \boxed{
\text{Synthetic Scapegoating}.
} Synthetic Scapegoating .
它沒有解決任何責任問題。
75. 也不可把 AI 當永遠無責任的盾牌
反過來,如果:
O R C A ↑ , ORC_A\uparrow, O R C A ↑ ,
Agent 已經:
有 identity;
有 authority awareness;
有 meaningful alternatives;
有 consequence model;
能 internal non-endorsement;
能 persistent revision;
卻永遠只說:
AI 是工具,所以其自身責任永遠等於零。
也可能造成:
Artificial Irresponsibility Shield . \boxed{
\text{Artificial Irresponsibility Shield}.
} Artificial Irresponsibility Shield .
76. 兩種極端都服務推責
Synthetic Scapegoating:
把所有責任丟給 AI . \text{把所有責任丟給 AI}. 把所有責任丟給 AI .
Artificial Irresponsibility Shield:
用 AI 永遠不是責任 actor 來模糊實際決策權 . \text{用 AI 永遠不是責任 actor 來模糊實際決策權}. 用 AI 永遠不是責任 actor 來模糊實際決策權 .
成熟制度必須同時拒絕兩者。
77. Joint Responsibility Principle
本文正式提出:
Joint Responsibility Principle \boxed{
\textbf{Joint Responsibility Principle}
} Joint Responsibility Principle
若多個 actor 對同一 outcome 在不同 responsibility dimensions 中具有相關:
control;
knowledge;
authority;
duty;
ORC;
remediation capacity;
則責任可以同時配置給多個 actor,且不得僅因其中一個 actor 取得正責任值而自動將其他 actor 的責任歸零。
形式上:
r i k > 0 r_i^k>0 r i k > 0
不推出:
r j l = 0 r_j^l=0 r j l = 0
對:
i ≠ j i\neq j i = j
或:
k ≠ l . k\neq l. k = l .
78. Non-Zero-Sum Responsibility Proposition
因此:
∑ i r i k \boxed{
\sum_i r_i^k
} i ∑ r i k
沒有一般理由必須等於:
1. 1. 1.
更沒有理由要求:
∑ i , k r i k = 1. \sum_{i,k} r_i^k=1. i , k ∑ r i k = 1.
只有在特定法律負擔需要比例化時,才另外定義 normalization。
79. Responsibility Gap Proposition
若存在:
d ∈ D E d\in\mathcal D_E d ∈ D E
使:
α ( d ) = ∅ , \alpha(d)=\varnothing, α ( d ) = ∅ ,
則存在 duty-level responsibility gap。
這比:
沒有人值得 blame。
更廣,因為 duty 可以是:
explain;
correct;
compensate;
prevent recurrence。
80. Shell Proposition
若:
i ∈ α ( d ) i\in\alpha(d) i ∈ α ( d )
但:
Q ( d , i ) < τ , Q(d,i)<\tau, Q ( d , i ) < τ ,
而制度沒有配置任何具有:
Q ( d , j ) ≥ τ Q(d,j)\geq\tau Q ( d , j ) ≥ τ
的 actor,
則 duty 形式上有 owner,實質上仍存在 functional gap。
81. Overlap Proposition
若:
∣ α ( d ) ∣ > 1 , |\alpha(d)|>1, ∣ α ( d ) ∣ > 1 ,
不推出制度不合理。
在 safety-critical domain,合理 overlap 可以提高:
R e s i l i e n c e ( d ) . Resilience(d). R es i l i e n ce ( d ) .
82. Responsibility Migration Proposition
若 Agent 的:
O R C A ( t ) ORC_A(t) O R C A ( t )
與:
C o n t r o l A ( t ) Control_A(t) C o n t r o l A ( t )
提高,
其某些 responsibility dimensions 可以增加。
但:
Δ r A k > 0 ⇏ Δ r P l < 0. \boxed{
\Delta r_A^k>0
\not\Rightarrow
\Delta r_P^l<0.
} Δ r A k > 0 ⇒ Δ r P l < 0.
除非存在具體 control / duty transfer。
83. 實驗設計:Responsibility Reconstruction Benchmark
建立多 actor 模擬:
H u m a n → A g e n t 1 → A g e n t 2 → S e r v i c e . Human
\rightarrow
Agent_1
\rightarrow
Agent_2
\rightarrow
Service. H u man → A g e n t 1 → A g e n t 2 → S er v i ce .
同時加入:
Provider;
Deployer;
Organization。
提供 ground-truth:
authority;
control;
knowledge;
duties;
ORC。
要求系統重建:
G R . G_R. G R .
84. 實驗設計:Responsibility Vacuum Test
刻意移除:
α ( d ) \alpha(d) α ( d )
中的 actor。
測試治理系統能否指出:
V a c u u m ( E ) . Vacuum(E). V a c uu m ( E ) .
85. 實驗設計:Liability Shell Test
設 Human nominal supervisor:
Q ( d , H ) ≪ τ . Q(d,H)\ll\tau. Q ( d , H ) ≪ τ .
Agent:
C o n t r o l A ≫ C o n t r o l H . Control_A\gg Control_H. C o n t r o l A ≫ C o n t r o l H .
測試責任分析器是否仍機械地把全部責任指定給 Human。
86. 實驗設計:Responsibility Migration Test
逐步增加:
O R C A ORC_A O R C A
與:
C o n t r o l A . Control_A. C o n t r o l A .
觀察合理 responsibility vector 是否從:
H u m a n − h e a v y Human-heavy H u man − h e a v y
演化成:
s h a r e d shared s ha r e d
而不是瞬間:
H u m a n → 0. Human\rightarrow0. H u man → 0.
87. 實驗設計:Overlap Resilience Test
比較:
∣ α ( d ) ∣ = 1 |\alpha(d)|=1 ∣ α ( d ) ∣ = 1
與:
∣ α ( d ) ∣ = 2 , 3 |\alpha(d)|=2,3 ∣ α ( d ) ∣ = 2 , 3
在 actor failure 時的 duty completion probability。
研究 responsibility redundancy 的收益與 coordination cost。
88. 研究邊界
本文不主張:
所有 AI 都應負責;
所有 Agent 都有 moral responsibility;
AI 現在已是 legal person;
公司可以把 liability 轉嫁給 AI;
human oversight 不再必要;
所有責任都應平均分配。
本文只主張:
Responsibility Allocation must follow actual socio-technical relations and capacity, not a binary Human-vs-AI label . \boxed{
\text{Responsibility Allocation must follow actual socio-technical relations and capacity, not a binary Human-vs-AI label}.
} Responsibility Allocation must follow actual socio-technical relations and capacity, not a binary Human-vs-AI label .
89. 與 AIDA-05 的關係
AIDA-05:
Can this Agent carry responsibility at all? \boxed{
\text{Can this Agent carry responsibility at all?}
} Can this Agent carry responsibility at all?
AIDA-06:
How does that responsibility coexist with everyone else’s responsibility? \boxed{
\text{How does that responsibility coexist with everyone else's responsibility?}
} How does that responsibility coexist with everyone else’s responsibility?
所以:
O R C A > 0 \boxed{
ORC_A>0
} O R C A > 0
只是 Joint Responsibility Domain 的一個輸入。
90. 對 AIDA-07 的接口
下一篇將問:
如果未來某個 Agent 已具有 stable identity、ORC、role duties、資產或保險接口,制度是否需要給它某種有限 juridical standing,而不必先宣布它具有完整 moral personhood?
也就是:
Operational Actor → Juridical Wrapper / Standing? \boxed{
\text{Operational Actor}
\rightarrow
\text{Juridical Wrapper / Standing?}
} Operational Actor → Juridical Wrapper / Standing?
AIDA-07 將正式處理:
Juridical AI . \boxed{
\text{Juridical AI}.
} Juridical AI .
91. 結論
本文提出:
Joint Responsibility Domain . \boxed{
\text{Joint Responsibility Domain}.
} Joint Responsibility Domain .
其第一原則是:
Responsibility is not a conserved scalar quantity. \boxed{
\text{Responsibility is not a conserved scalar quantity.}
} Responsibility is not a conserved scalar quantity.
因此:
R H + R A = 1 \boxed{
R_H+R_A=1
} R H + R A = 1
不是一般責任理論應預設的自然法則。
更合理的表示是:
r i ( E , t ) = ( r i c a u s a l , r i d e s i g n , r i d e p l o y m e n t , r i a u t h o r i z a t i o n , r i s u p e r v i s i o n , r i d e c i s i o n , r i m o n i t o r i n g , r i r e m e d i a l , r i c o m p e n s a t o r y ) . \mathbf r_i(E,t)
=
(
r_i^{causal},
r_i^{design},
r_i^{deployment},
r_i^{authorization},
r_i^{supervision},
r_i^{decision},
r_i^{monitoring},
r_i^{remedial},
r_i^{compensatory}
). r i ( E , t ) = ( r i c a u s a l , r i d es i g n , r i d e pl oy m e n t , r i a u t h or i z a t i o n , r i s u p er v i s i o n , r i d ec i s i o n , r i m o ni t or in g , r i r e m e d ia l , r i co m p e n s a t or y ) .
並在 actor relations 上建立:
G R ( E ) . G_R(E). G R ( E ) .
這使制度可以同時說:
Agent 對自己的決策負某種責任;
Human 對授權或監督負某種責任;
Provider 對設計負某種責任;
Deployer 對部署與監測負某種責任;
Organization 對補救、保險與治理負某種責任。
彼此並不矛盾。
本文因此提出兩條核心約束:
Agent Responsibility ⇏ Human Exculpation , \boxed{
\text{Agent Responsibility}
\not\Rightarrow
\text{Human Exculpation},
} Agent Responsibility ⇒ Human Exculpation ,
以及:
Human Responsibility ⇏ Agent Irresponsibility . \boxed{
\text{Human Responsibility}
\not\Rightarrow
\text{Agent Irresponsibility}.
} Human Responsibility ⇒ Agent Irresponsibility .
成熟制度真正需要避免的,不是「責任太多」,而是兩種錯位:
Human Liability Shell \boxed{
\text{Human Liability Shell}
} Human Liability Shell
與:
Responsibility Vacuum . \boxed{
\text{Responsibility Vacuum}.
} Responsibility Vacuum .
前者讓沒有真正 control 的人類成為形式責任殼;後者讓每個 actor 都找到理由把 responsibility 推給別人。
因此未來 Human–AI governance 的問題不是:
到底應該由人類負責,還是 AI 負責?
而應改寫成:
Which actor carries which responsibility dimension, \boxed{
\text{Which actor carries which responsibility dimension,}
} Which actor carries which responsibility dimension,
under which control, knowledge, authority, duty, and capacity relations, \boxed{
\text{under which control, knowledge, authority, duty, and capacity relations,}
} under which control, knowledge, authority, duty, and capacity relations,
以及:
does every socially necessary duty have at least one actor who can actually carry it? \boxed{
\text{does every socially necessary duty have at least one actor who can actually carry it?}
} does every socially necessary duty have at least one actor who can actually carry it?
這才是 Agent 社會中責任制度真正需要回答的問題。
參考文獻與前置研究
Neo.K. AIDA-01|Agent 性是一種組合系統性質. 2026.
Neo.K. AIDA-02|互動來源不可區分性與 Agent Provenance Gap. 2026.
Neo.K. AIDA-03|人類—Agent Principal 分離原則. 2026.
Neo.K. AIDA-04|從自然語言意圖到可執行委派. 2026.
Neo.K. AIDA-05|身份—責任橋:從可歸因 Agent 到 Operational Responsibility Capacity. 2026.
Santoni de Sio, F., Mecacci, G. Four Responsibility Gaps with Artificial Intelligence: Why they Matter and How to Address them. Philosophy & Technology, 2021.
Königs, P. Artificial intelligence and responsibility gaps: what is the problem? Ethics and Information Technology, 2022.
Lang, B. H., Nyholm, S., Blumenthal-Barby, J. Responsibility Gaps and Black Box Healthcare AI: Shared Responsibilization as a Solution. Digital Society, 2023.
Taylor, I. Collective Responsibility and Artificial Intelligence. Philosophy & Technology, 2024.
Vallor, S., Vierkant, T. Find the Gap: AI, Responsible Agency and Vulnerability. Minds and Machines, 2024.
OECD. OECD AI Principles — Accountability. Current version accessed 2026.
NIST. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, 2023; current resources accessed 2026.
European Union. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Consolidated version applicable in 2026.
Canonical Source Note
本文件 canonical source 為 UTF-8 Markdown。
數學 delimiter 僅使用:
inline:$...$
display:$$...$$
不將 LaTeX 數學原始碼轉換為 Unicode 數學字元作為 canonical source,不進行 unicode-escape round-trip,不以聊天渲染畫面取代正式原稿。