Do not collapse different tasks
A delayed delivery, a complaint and a change of address can demand different handling. A delivery-status question may call for reliable system data, not a language model.
AI ETHICS · ENGLISH TOPIC OVERVIEW
AI ethics begins before the first prompt
Before selecting a model or writing an instruction, decide what problem is being addressed, whose interests and burdens count, and who may review, stop or revise the proposed action.
The spare-parts business used here is entirely fictional. This overview is an educational aid, not legal advice and not approval to deploy an AI system.
Read the complete articleStart with the decision
“Which decision is to be prepared or made here, by whom, and with what information?”
Make the goal contestable
“Faster” can mean an acknowledgement, routing to the right team, substantive resolution or being taken seriously. Choose measures and a harm threshold before testing.
Keep four questions distinct
Moral judgement
What seems wrong or at stake for someone affected?
Ethics
Which interests conflict, what reasons support each option, and what would an opposing view change?
Law
Which concrete duties and rights apply in the relevant jurisdiction?
Governance
Who may decide, review, stop and improve? A nominal human check is ineffective without time, information and authority.
F: revise the question before testing
A
Clarify
Separate goals and expose assumptions.
B
Alternatives
Keep competing explanations and a non-AI option open.
D
Errors and gaps
Name uncertainty, missing evidence and possible refutation.
C
Revised question
Set an alternative, role, observable result and boundary.
Opening request
“How can we have AI answer all customer enquiries immediately?”
It already assumes automated external sending.
Revised question
“Can a bounded trial using invented messages show whether AI sorts only non-critical enquiries better for internal review than an improved form, without sending messages itself?”
Connect C to GROW: Goal, Reality, Options, Way forward. This is a plan for a next action, not an approval for real data or external communication.
K: change the plan when affected people raise an objection
Explore affected perspectives
Customers, customer service, the specialist team and management can experience different consequences. Also ask who can safely object and where a reported error can go.
Document a change
If customer service shows that a complaint can look like a status question but require a binding commitment, suspected complaints move to a separate manual class. Participation must alter the plan, or it is merely a list of attendees.
K is not R. K is the consultation model used here to explore perspectives and potentially change the test boundary. The separate five-phase advisory model R structures the specialist review path once a concrete productive data flow is being considered. Neither is an approval mechanism.
Three thresholds keep a demonstration from becoming deployment
| Threshold | Question |
|---|---|
| 1 · Learn from artificial cases | Is the problem definition clear enough to create invented cases and learn from them? |
| 2 · Prepare a real pilot | Have purpose, roles, data flows, affected-persons pathways, comparison standard, missing facts and needed specialist reviews been documented for R Phase 1? |
| 3 · Consider external effects | Do actual observations justify limited external effects or broader operation? |
The first “yes” does not replace either later “yes.” Company size alone does not remove the need for context and risk assessment.
A bounded decision card
| Decision | Internal pre-sorting only; automatic sending lies outside the test. |
|---|---|
| Comparison | An improved input form and clearer responsibility rules without a language model. |
| Test material | Only newly invented messages; no real customer data and no external messages. |
| Roles | Management decides on the bounded test; specialists assess routing; customer service records exceptions; one named person has stop authority. |
| Stop signal | A message leaves the test without authorisation, a real message enters it, or a consequential exception is treated as routine. |
| Open matters | Baseline, data and contractual questions, participation of affected people and any approval for a later pilot. |
A mini-trial can teach without pretending to prove everyday performance
Before the comparison
Create twelve wholly invented messages: four clear status questions, four with missing information, and four plausible complaint or exception cases. Agree on the expected specialist class and treatment first. Mark disagreement as open.
During the comparison
Team A uses the improved form. Team B uses internal, reviewed AI pre-sorting. Record the proposed class, human correction, time spent and reason for escalation.
Boundary: An accidental sending or real data set pauses the exercise immediately. One exercise error does not prove every AI approach unsuitable; examine its cause before revising. Twelve invented messages do not represent a real customer population.
What the result can and cannot say
Empirical
Does one variant route more specified cases correctly, and what rework results?
Normative
Can a gain in clear cases justify a possible disadvantage for people with rare concerns?
Institutional
Who approves, objects, acts after an error and has the real authority to do so?
Permitted outcome
“No AI for now” or “continue examining the non-AI solution” can be well-reasoned results.
Sources and scope
- NIST AI RMF 1.0 (2023): voluntary risk-management framework; this decision card is an independent educational derivation.
- UNESCO Recommendation on the Ethics of AI (2021) and the OECD Recommendation on AI: normative guidance, not case-specific legal assessment.
- Elish (2019), “Moral Crumple Zones” and Selbst et al. (2019), “Fairness and Abstraction in Sociotechnical Systems”: conceptual sources, not measured evidence for this fictional SME.