AI ETHICS · ENGLISH DAILY OVERVIEW

Who bears responsibility when AI prepares the work?

A plausible draft is not yet a responsible decision. What matters is whether a responsible role can understand the case, change or stop the proposal with practical effect, and later explain and correct the decision.

R-184 is wholly fictional: an exercise using invented tickets. There are no real customer data, no external communication and no message is sent.

Read the full article

1. Responsibility requires three connections

A human click is not enough. Responsibility arises only from case-specific knowledge, a practical right to intervene, and an attributable decision with an accessible route to correction.

Knowledge

The role sees the permitted case file, missing facts and applicable rule, rather than only the model text or a status indicator.

Intervention

It may change, reject or stop the draft, and has the time, cover and clear escalation destination needed to do so.

Accountability

A trail connects case material, change, rule version, decision and later feedback. It makes the decision reviewable; it does not automatically make it correct.

2. R-184 starts with an open fact

Fictional complaint

A workshop reports damaged packaging for a component with possible safety relevance. Installation is scheduled. Whether the box was opened and what condition the part is in remain open.

AI draft

The model proposes requesting photos and announcing a decision within five working days. Smooth wording replaces neither fact-checking nor an authorised communication decision.

The working question is: With what information, rights and evidence may a service worker confirm an AI recommendation for a transport-damage complaint?

3. F-B: Three hypotheses instead of assigning blame

F-B sets competing explanations for a loss of responsibility alongside each other. The non-AI option remains a professional checklist with text modules.

B1 · Knowledge gap

The service worker does not fully see the ticket, delivery status, photo or missing information.

B2 · Intervention gap

A formal change is possible, but time, cover or authority to ask follow-up questions are absent in practice.

B3 · Accountability gap

The click is stored, but the basis, exception, rule version and responsible correction role are not.

F-D and F-C: The unclear box opening and the safety marker trigger a stop. A service worker confirms only when all required case information is present, no stop trigger applies, and the draft remains within an approved rule; otherwise, they hand the case on.

4. K governance: roles need knowledge, rights and evidence

K changes the matrix: exploration separates sources and gaps; reflection makes the interests of workshop, service, warehouse and management visible; decision assigns rights; and feedback considers stopped, changed or escalated cases.

Role	Must know	May and must intervene	Evidence
Service worker	Original ticket, draft, delivery status, rule and missing facts.	Change or reject the draft; do not close when a trigger applies.	Case ID, decision, short reason and missing evidence.
Warehouse specialist	Photos, packaging and shipping data, product identification and review criteria.	Mark the extent of damage as unclear; initiate inspection or holding the part.	Finding, source, time and, where relevant, hold marker.
Service management	Complete case file, earlier interventions, consequences and deployment boundary.	Stop the draft; decide on communication only on the basis of the finding; refer onward where needed.	Decision-maker, reason, rule or exception, and intended communication.
Business management	Patterns from escalations, workload, open risks and resource position.	Narrow, pause or end the deployment class; secure time, training and cover.	Rule version, stop decision, review date and resource decision.

Another case role can perform a cross-check; a deputy replaces a role, but does not cross-check that role’s own initial decision.

5. Stop and hand over with a trace

Immediate stop

Do not close when there is possible safety relevance, an unclear extent of damage, a commitment or rejection without evidence, a rule deviation, or a draft that cannot be explained.

Minimum decision trail

Record: case ID, purpose and boundary, permitted inputs, model/prompt/rule version, records, change or rejection, escalation, outcome and feedback.

For R-184, this means that the draft remains stopped and unsent. The warehouse specialist examines the facts; service management decides on intended communication only from a more complete basis.

6. R: A narrow specialist-review request, not a legal stamp

R is not completed in full for every ticket. For R-184, the intake begins in a structured way because safety relevance, replacement costs and claims may be involved. Specialist review is required only when the decision clearly depends on a contract or warranty, compensation, cross-border delivery, new data sharing or goodwill not covered by a rule.

Handover

Case material, open facts, relevant records where available, jurisdiction, data flow and decision deadline are handed over as a bounded request.

Return

The specialist role clarifies the concrete question or requests records. It returns review priority, open facts and, where applicable, an option for action, not a general compliance seal.

7. Three proportional thresholds

Threshold	Question for the next level
1 · Exercise	Is the role and escalation logic with invented material only understandable and complete enough to make its gaps visible?
2 · Later bounded deployment	Have purpose, roles, decision rights, data flows, route to correction, comparison standard and required specialist reviews been documented and reviewed separately?
3 · External effects	Is there an independent basis for considering limited external effects or broader operation at all?

A positive result at one threshold does not replace review at the next. The thresholds measure neither speed nor error rate, and they do not authorise real deployment.

8. 90 minutes: “The click has no name”

Time	Task	Visible outcome
0–10	Read ticket and draft separately.	Mark established and missing facts, model text and possible effect.
10–25	F-A and F-B: clarify terms, create three hypotheses and a non-AI option.	Testable gaps in knowledge, intervention and accountability.
25–40	F-D and F-C with box opening, safety relevance and scheduling consequences.	Stop, follow-up question and narrower rule without inventing a fact.
40–60	K exploration and reflection as a role round.	Legitimate interests, missing information and possible actions for each role.
60–73	K decision: complete the matrix, cover and two stop triggers.	Knowledge, intervention right and accountability assigned with evidence.
73–83	Write a narrow R specialist-review request.	Concrete question, missing records and route back, without a legal opinion.
83–90	Peer review with the question: Can the service worker really stop?	Documented revision; a gap returns to F-D or K.

9. Sources and limits of application

NIST AI RMF 1.0 is voluntary guidance for roles, responsibilities, communication routes and feedback. It does not prescribe a specific role matrix or a legal consequence for R-184.
The UNESCO Recommendation on the Ethics of AI and the OECD Recommendation on AI offer normative guidance on attributable responsibility, review, correction and override. They are not individual legal advice.
Santoni de Sio and van den Hoven provide the philosophical concepts of tracking and tracing. These concepts are not an empirical effectiveness study for this fictional business.
The studies on automation assistance cited in the article offer bounded findings from their respective simulations. They provide no error rate for generative AI, customer service or R-184.

This overview is a learning path. The matrix and thresholds are teaching, editorial constructions. They certify neither legal compliance nor the safety or effectiveness of a real AI deployment.