Values Without a Built-In North

How organisations justify AI decisions when good reasons conflict

A generative AI system can produce wording about fairness, safety, or responsibility. Such an output does not yet provide a moral standard. When a system is used in a specific workflow to sort, recommend, or prioritise, goals, categories, data, and exceptions all shape the decision. Treating these specifications as “neutral output” does not make values disappear; it makes them harder to see.

For small businesses, education teams, and organisations, the task is therefore not to find an artificial moral compass. They need to decide in a traceable way which reasons matter in a given case, where a non-negotiable limit lies, who can raise objections, and when an earlier balancing of reasons is reopened. A decision can be justified even when those involved do not arrive at a single value judgement.

This article develops the distinction through a wholly fictional case: a small continuing-education provider considers whether AI may create simplified and multilingual versions of an internal learning module from a German source text. The case is a model for thinking and practice. It describes neither a real organisation nor actual learners or observed effects.

1. The simple language question that does not exist

“Can we simply have the text translated into further languages and simplified?” sounds like a clearly bounded production task. An AI could rewrite a German source text, explain technical terms, and create variants for different learning contexts. Yet the very terms “simple,” “understandable,” and “equivalent” already contain value decisions: Which statement must remain unchanged in every version? Who decides whether a version is accessible enough? And whose effort counts when each variant must be reviewed?

The fictional provider is preparing an internal learning module on responsible AI use. The team wants to give comprehensible materials to staff with different levels of language proficiency. The first draft says: “Everyone receives the same content, just in their own language.” But “the same content” is not a finished specification. A sentence can be translated literally and still miss its meaning. An explanatory addition can make understanding easier while also shifting the statement. Simplification can open access; it can also remove an important condition or doubt from the text.

“Same,” too, can mean more than one thing. A shared standard version provides a central reference and can make changes easier to trace. But it can overlook linguistic differences and different levels of prior knowledge. Individually adapted explanations can help learners situate technical terms in their own circumstances. Yet they create additional variants that must be carefully checked for shifts in meaning. Professional translation without AI remains a third option. It can secure quality where sufficient time and expertise are available; but it does not automatically resolve which terms are immutable and which examples may be adapted.

The decision therefore concerns more than style. It concerns access to knowledge, comparability of learning objectives, the authority of a source version, and the distribution of review effort. A version helps determine which terms remain visible, which examples are treated as normal, and which learners receive additional explanations. That is a structural possibility, not yet evidence that a particular language version in a particular course produces a particular effect.

A helpful first distinction is therefore: A shared wording can enable comparability; it does not by itself determine whether all learners have equally good access through it. And an adapted version is not automatically fairer if it omits conditions or is misunderstood as a binding rule. Which parts of the material must remain the same and where adaptation is defensible is a normative and editorial question. It cannot be answered by choosing a translation tool.

2. Value pluralism is not arbitrariness

When several serious reasons conflict, two convenient reactions are available. The first says that there must be a single objective ranking somewhere; if we only calculate correctly, it will become visible. The second says that values are personal preferences, so there is nothing to justify anyway. Both reactions oversimplify the problem.

The philosopher Isaiah Berlin described his position as value pluralism. In his essay “The Pursuit of the Ideal,” he argues that people can pursue several goods that are each intelligible and important, but do not always fit into a final common scale. This is Berlin’s philosophical position, not an uncontested overall theory of ethics. His point is nevertheless useful for organisational decisions: A conflict between, for example, reliability, harm avoidance, self-determination, and equal treatment need not mean that one side is merely irrational. Berlin distinguishes this view from relativism: Different value judgements are not automatically mere statements of taste; they can be understood, criticised, and defended with reasons. Berlin, “The Pursuit of the Ideal”, book pages 9–12.

This does not mean that all reasons carry equal weight. A company can reject a justification because it rests on a false factual assumption, systematically excludes an affected perspective, violates a protective boundary, or suspends its own rules only for certain customer groups. Nor does pluralism mean that every decision is settled by saying, “Those are simply our values.” Rather, it shifts the task: reasons must be named concretely, their assumptions disclosed, and their consequences made reviewable.

A distinction among four levels helps here:

1. Empirical question: Which statements, conditions, or examples change between the source text and the draft? Which comprehension questions arise in a reading sample? Which feedback is still missing? This requires text comparison and appropriate feedback; fluent wording alone is not evidence of comprehensibility.

2. Normative question: Which core statements must hold in every version? When does an explanation support access, and when does it already shift the statement? Which inequality or exclusion would be avoidable? These questions require arguments, not merely a metric.

3. Institutional question: Who determines the semantic core, approves deviations, or withdraws a flawed version? Who bears the review effort, who can object, and who must respond to an objection?

4. Open question: Which learner perspectives or language variants have not yet been considered? What has actually been checked before use, and what may not enter the material as apparent certainty?

Those who mix up these levels can turn “literally the same” into a value standard, the wish for accessibility into an unsubstantiated effect, or an editorial decision into an alleged system property. The distinction does not resolve the conflict. It shows which kind of justification is missing.

3. Where values enter an AI process

It is inaccurate to ascribe a moral compass of its own to an AI system. A system can reproduce value terms, compare perspectives, or generate a recommendation. That does not mean that it itself possesses a publicly justified standard or can assume institutional responsibility. The article does not thereby claim to provide a final metaphysical answer to whether machines could ever be moral agents. For practical use, the narrower observation is sufficient: goals, data, categories, thresholds, exceptions, and authority are set in a socio-technical context and must be justified by responsible people and organisations.

The NIST AI Risk Management Framework 1.0 describes how people’s assumptions, expectations, and decisions can influence the AI lifecycle and shape risks and biases. The framework also notes that metrics and thresholds require decisions about which performance, errors, or risks are relevant. It is a voluntary risk-management framework, not a guarantee of ethical outcomes or a universal value order. The OECD Recommendation on AI directs human-rights, rule-of-law, and democracy-related principles to the actors responsible for AI. This normative orientation is an important minimum reference point. But it is not an automatic decision in individual cases and does not anticipate every specific trade-off.

The scholarly debate warns against two opposite shortcuts. First, technical implementation is not neutral merely because a rule appears formal. In their analysis of socio-technical fairness, Selbst, boyd, Friedler, Venkatasubramanian, and Vertesi show that the choice of an abstraction—that is, what is treated as a problem, category, or measurable goal—can carry a normative position. This does not mean that formalisation is useless. It means that the chosen abstraction must be justified and checked for the contexts it leaves out. Selbst et al., “Fairness and Abstraction in Sociotechnical Systems”.

Second, the question cannot be disposed of with a list of recognised principles. In 2019, Jobin, Ienca, and Vayena examined a corpus of published AI ethics guidelines. They found overlaps concerning transparency, justice and fairness, non-maleficence, responsibility, and privacy, as well as differences in how these principles were interpreted, justified, and implemented. The study describes documents in its corpus at that time; it proves neither worldwide agreement nor the effectiveness of a guideline in an organisation. Original article and openly accessible author version. Mittelstadt further argues that high-level principles alone do not guarantee ethical practice. This is a reasoned critique of principles alone, not evidence that shared principles are useless. Jobin et al.; Mittelstadt.

For the fictional course team, this means that if a version is marked “equivalent,” it must be clear which meaning it is to preserve, who checks deviations, which passages must remain unchanged, and how learners can report a misleading passage. A model can formulate a translation in linguistically convincing terms. Convincing language does not replace a review of the underlying meaning.

4. Two passages, no worldview templates

Philosophical perspectives are easily reduced to labels: “This culture is communitarian,” “that tradition is individualistic,” “here harmony counts, there comprehensibility.” Such shorthand turns internal debates into apparently uniform group voices. A careful comparison begins on a smaller scale: with specific texts, their translations, their historical setting, and their limits.

In his teaching translation of the Analects, section 13.23 describes acting in harmony with others without seeking mere uniformity. This can open a question for the training materials: Must a shared learning intention require identical words everywhere, or can a sustaining meaning be preserved in different explanations? Robert Eno emphasises that the Analects are a collection that developed over a longer period and that his translation represents his own interpretation; individual readings are not necessarily consensus. The passage therefore represents neither “the Confucian view” nor an unchanging Chinese value order. It offers a text-bound prompt for thought, not a contemporary translation rule. Analects 13.23, translation and commentary by Robert Eno, especially p. 71.

In Qur’an 30:22, the diversity of languages—together with the diversity of colours in the English translation used here—is named as one of the signs on which people should reflect. This is a specific religious passage, not a summary of Islamic ethics and not an interchangeable argument for contemporary education policy. Its theological context is part of its meaning. For the fictional course, it can prompt the question of whether linguistic variation must be regarded as a defect or whether, with equal care for content, it can have a place in learning materials. The verse does not decide this practical question. Qur’an 30:22, “The Clear Quran” translation by Mustafa Khattab.

The two passages are not added together into a shared “Eastern” or “religious” principle. Nor do they have to give the same answer. Read alongside Berlin, they instead show why the justification of values requires genuine conversation: uniformity, comprehensibility, harmony, and linguistic diversity can each provide reasons that do not automatically align in a particular learning environment. A text alone replaces neither contemporary interpretation nor the voices of those who learn or are responsible for the material.

For the course, this means learners should not guess which culture “prefers” access. They should reconstruct concrete reasons, formulate counterarguments, and identify who has so far been allowed to speak in a debate. If an organisation draws on a tradition or religious source, it needs knowledgeable interpretation and must not make any individual or group the spokesperson for all.

5. Open a question before ordering an outcome

In the provided German version, the question-expansion model is structured as the sequence A → B → D → C. A clarifies terms and assumptions and can use different questioning techniques to do so. B develops several hypotheses, meta-questions, and alternative explanations. D keeps iteration, follow-up questions, and fallback open. Only then does C fix the revised question and the stated next step for action. This article presents a question card formulated specifically for the case within A. QFT remains unused and is not reproduced; nor does the article reproduce a complete external questioning protocol. The model is a working method, not evidence that it is empirically superior in every group. (Sakızlı, Question-Expansion Model, pp. 1–3.)

A: Surface terms, open the space for thought

The course team’s initial question is: “How can AI make the text equally understandable for all learners?” F-A first clarifies the terms. Does “equal” mean verbatim, equivalent in content, or with the same learning objectives? Does “understandable” mean fewer technical terms, shorter sentences, or different examples? Where is the boundary between a translation and an additional explanation? Is the AI only to produce a draft, or may it also shorten and reorder statements?

These questions expose the assumptions. The initial question assumes that a text can be adapted for all learners, that AI recognises the relevant differences, and that “equally understandable” is an outcome that can be defined without learners. None of these assumptions has yet been demonstrated.

The purpose-created question card now opens the space for thought without prejudging the later decision:

Which statements or terms must remain semantically the same in every language version? Where would a literal translation make the content less accessible? Which additional examples explain a term, and which already change the statement? Who checks whether a simplification removes a condition, uncertainty, or counter-position? What feedback from learners could change a formulation? Can a human-created glossary or a reviewed translation meet the need without AI? Who should be able to use a version without being treated as the representative of an entire language group?

The question card is neither a vote nor evidence of representativeness. It distinguishes what can be checked through sources and text comparison, what requires a normative decision, and whose feedback is still missing for the next version.

B: Competing hypotheses instead of a premature uniform version

F-B does not turn the list into an overall score. It formulates at least two defensible hypotheses and keeps a third option visible:

Hypothesis 1 – a shared reference version protects comparability. All language versions closely follow a reviewed source text. This makes version control easier and can prevent central conditions from diverging unnoticed. The weakness: a uniform text is not automatically equally accessible. It may assume technical terms or prior knowledge that are unfamiliar to all learners.

Hypothesis 2 – adapted explanations can reduce barriers to access. Examples, glossaries, and linguistically simplified explanations could make content easier to connect with for different learning prerequisites. This is a plausible educational hypothesis, not an effect demonstrated here. The weakness: adaptation can change the meaning. Additional variants take time and must be reviewed for shifts in meaning.

Third option – translation and adaptation without AI. Expert translation or joint editorial work can serve as a comparison. It requires resources and can delay provision. But it is not automatically worse, slower, or more expensive than an AI-assisted route; those are concrete questions for the organisation concerned.

The meta-question is: Which semantic core must be preserved in every version, and which explanation may change so that access does not come at the expense of meaning? This question links the different values without forcing them onto a common scale.

D and C: Revise the line of thought visibly

The first draft might say: “A shared source and an AI translation ensure that everyone receives the same information.” F-D counters: A shared source guarantees neither comprehensibility nor a translation with the same meaning. The term “the same information” must first be specified. At the same time, it would be unsubstantiated to describe an adapted version as more accessible before appropriate learners have reviewed it.

The objection changes the option. Instead of rewriting all materials indiscriminately, the team first separates binding core statements from explanatory examples. It retains a human-reviewed source version and treats AI outputs at most as clearly labelled drafts. Which passages must remain unchanged, which adaptation is defensible, and who approves a deviation become open decisions.

F-C refines the working question: Which parts of AI-assisted simplified or translated learning material must remain unchanged in meaning, which explanations may be adapted, and what review by experts and learners is needed before a version is used? This final question prescribes neither AI nor a particular language policy. It makes visible what the decision depends on and which feedback can trigger a new review.

6. Consultation is a decision practice, not a consensus machine

In its basic structure, the Sakızlı consultation model connects exploration, reflection and analysis, decision-making and recommendation, and feedback and evaluation. The model is broader than a compliance checklist. It addresses learning objectives, different knowledge perspectives, ethics/governance, and feedback loops. The four steps of the basic cycle must be distinguished from the more detailed five-stage case scenarios, in which implementation/pilot and final assessment are set out separately. The template itself is a methodological design; it does not demonstrate that a consultation format is automatically fair, effective, or complete. (Sakızlı, Consultation Model, pp. 7–8 and 20–33.)

In this learning-material case, exploration is not merely an invitation to put as many names as possible on a list. The team clarifies which versions are to be created, what they will be used for, what counts as a binding core statement, and who can withdraw a flawed version. Possible perspectives are the role responsible for editorial content, subject-matter teachers, people who use the materials, and reviewers proficient in the relevant language. These roles are questions for the fictional design, not a claim that any individual or group can represent all learners.

In reflection and analysis, the team compares reasons, not the supposed loudness of participants. A central reference can limit divergences between versions while overlooking linguistic barriers. Adapted explanations can open access while also changing the core. An exclusively human process can be careful while also tying up scarce resources. Here the strongest counter-position belongs on the table: small providers cannot make unlimited time, translation, and review available for every text variant. Missing or delayed material can also make access harder. Comprehensive consultation is not automatically better than a short, clearly bounded, and correctable process.

Consultation must therefore enable feedback that can change the outcome. If the team has already decided that only uniform formulations are permitted and merely records feedback, reflection is decorative. Conversely, not every piece of feedback needs veto power. The team must justify who decides, which passages may be changed, which core statement is protected, and why a remaining objection was not adopted.

In decision and recommendation, the provider can limit the task instead of immediately translating every module into multiple versions. One possible provisional rule for this fictional course is: The reviewed source text remains the reference. AI may supply only non-binding explanatory drafts for clearly identified learning material. Conditions, limitations, and central terms are not changed silently. A qualified person checks the meaning; feedback from learners can reopen a version. Human-created translations or no additional version remain equally valid options to examine. This is an educational decision, not evidence of effectiveness and not a general language policy.

Feedback and evaluation do not ask only whether texts are produced more quickly. They examine whether an explanation retains the core, whether central limitations are missing, whether learners can report an unclear passage, and whether individual versions drift apart over time. A simple comparison described in advance can check whether an example still represents the core statement accurately. Such a reading sample is not evidence of general comprehensibility. A review date or a specific objection can trigger a new F-D question.

By contrast, the separate advisory model R would come into play only if the material raised a specific legal, contractual, or subject-matter question. In that case, the passage at issue, the product, the role, and the relevant context would need to be delimited. This article does not demonstrate this examination; it only marks the possible handoff. A consultation with learners is not the same as specialist advice; specialist advice, in turn, does not decide for the team which value judgement it should represent. (Sakızlı, 5-Phase Advisory Model, pp. 1–3 and 20.)

7. The value-conflict map as a completed working form

A value-conflict map suitable for SMEs should not place yet another abstract framework on top of the work. On one page, it records what is being decided, which reasons conflict, and how a decision can be reviewed. It replaces neither participation nor research or specialist review. The following entry is a practice example, not an empirically validated template.

Fictional entry: Language versions of a learning module

The map deliberately contains no overall score. A value such as “access: 8, fidelity to the source text: 9” could create the impression that very different reasons can be compared on the same scale. A team can record concrete textual deviations or outstanding feedback. These are aids to review, not a complete standard of moral appropriateness and not empirical evidence that all learners understand a version.

Compact 90-minute course run

A course can use the map in five sections:

1. 20 minutes – F-A question phase: clarify “equal,” “understandable,” “simplified,” and “faithful to content”; sort open questions into empirical, normative, and organisational points.

2. 15 minutes – F-B: formulate at least two justifiable language options and a human non-AI alternative; name benefits, costs, and open evidence.

3. 30 minutes – K exploration and reflection: distinguish teachers, learners, and specialist/language review as concrete perspectives; examine the strongest counter-position and risks to meaning.

4. 15 minutes – K decision: record a provisional rule with an immutable core, permitted adaptation, responsibility, and objection.

5. 10 minutes – feedback and revision: identify which feedback reopens a version and who initiates the revision.

Assessment addresses the quality of the justification, the distinction between factual and value questions, a fairly formulated objection, and the capacity for revision. It does not assess whether all participants adopt the same position. A useful closing question is: Which new information or objection would genuinely change your provisional recommendation? If the answer is “none,” the process may not be open to review.

8. The strongest counter-position: small teams cannot review every version without limit

A small continuing-education provider does not have unlimited time or language expertise. Every additional variant needs maintenance. Review by several people can delay publication and tie up capacity needed for other learning content. Conversely, a rapid AI output can create false confidence: it sounds finished even though no one has checked semantic equivalence. Inviting feedback does not automatically solve the problem of representation either. Someone who takes part in a course speaks only for their own experience, not for everyone who uses the same language.

These objections go to the heart of the matter. Pluralism must not lead to material being withheld indefinitely for fear of every deviation. Nor is the answer to turn every piece of feedback into unlimited review. Review effort should be proportionate to the significance, binding force, and correctability of the passage. For an optional explanation of a term, a clearly labelled draft may be sufficient. A binding instruction, a limitation, or a normative core requires more careful review. This proportionality rule is an editorial recommendation in this article, not a verbatim NIST or OECD requirement and not an empirically validated method.

A shared frame of reference remains significant. The OECD Recommendation directs human-rights, rule-of-law, and democracy-related principles to AI actors. That limits the idea that every organisation may define fundamental rights or equality as it pleases. Yet such guidelines do not determine for each course whether an explanatory passage should be translated literally, simplified, or turned into an additional example. Principles limit the room for manoeuvre; the specific justification remains open.

Even a methodologically sound map does not guarantee a good translation or fair participation. Documentation can become ritual if no version can be changed or withdrawn. A specialist review is ineffective if it cannot review the language version. Course feedback does not automatically say anything about people who did not participate. A one-time comparison is not a general effectiveness test. These limits belong in the evaluation and in the later Agentic Package, which operationalises the series’ methods for concrete materials.

9. A justified judgement remains reviewable

In the fictional workflow, a generative AI system could describe a value conflict, organise possible reasons, or outline possible consequences. Such outputs would be drafts for review; they would neither determine bindingly which reasons count nor who sets the boundary. Responsibility for the goal, rule, boundary, and consequences does not thereby automatically shift into the system. To say that an organisation’s decision “was made by the AI” does not sufficiently identify who set the criteria, released data, accepted suggestions, or failed to make corrections.

Value pluralism does not force a team into arbitrariness. It requires that a conflict of aims not be disguised as a mere calculation gap or personal taste. Those involved can exchange reasons, formulate shared boundaries, choose a provisional option, and still document what remains unresolved. This does not give a decision infallible moral authority. It gives it a reviewable justification and a way to respond to new objections or findings.

For the fictional provider, responsible work therefore does not begin with the question of which model rewrites the text most fluently. It begins with a more precise question: Which meaning must every version preserve, where may an explanation be adapted, who can recognise a shift—and which feedback would change our decision?

An artificial north is not needed for this. What is required is a disclosed purpose, sound facts, justified value judgements, clear boundaries, and an organisation that does not take its own standard to be given by nature.

Sources and further primary texts

1. Isaiah Berlin. “The Pursuit of the Ideal.” In: The Crooked Timber of Humanity: Chapters in the History of Ideas, 2nd edition, Princeton University Press, 2013, pp. 1–19; especially book pages 9–12. Online text, Wolfson College, Oxford. Philosophical position; not a ready-made organisational decision rule.

2. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, 2023, especially pp. 11–13, 23, 40–41. Official PDF. Voluntary framework; not a guarantee of conformity or effects.

3. OECD. Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449, adopted 2019, amended 2023 and 2024, especially Principle 1.2. Official consolidated version. Normative frame of reference; concrete applicability depends on role and context.

4. Anna Jobin, Marcello Ienca, and Effy Vayena. “The global landscape of AI ethics guidelines.” Nature Machine Intelligence 1 (2019), pp. 389–399. Publisher article; openly accessible author version, abstract p. 1 and discussion pp. 14–15 of the PDF version. Document analysis of a time-bound corpus, not a measurement of implementation or worldwide agreement.

5. Brent Mittelstadt. “Principles alone cannot guarantee ethical AI.” Nature Machine Intelligence 1 (2019), pp. 501–507. https://doi.org/10.1038/s42256-019-0114-4. Scholarly perspective contribution on the limits of principle-based approaches.

6. Andrew D. Selbst, danah boyd, Sorelle A. Friedler, Suresh Venkatasubramanian, and Janet Vertesi. “Fairness and Abstraction in Sociotechnical Systems.” In: Proceedings of the Conference on Fairness, Accountability, and Transparency (FAT* 2019), pp. 59–68. Publisher article; complete author version, especially printed pp. 59–61 (PDF pp. 2–4). Conceptual contribution; not a universal empirical claim about every metric.

7. The Analects of Confucius: An Online Teaching Translation, translated and commented on by Robert Eno, version 2.21 (2015), Book XIII, section 13.23, p. 71. Open-access teaching translation from Indiana University. Eno identifies his translation as his own interpretation; it does not replace a complete textual-historical or Confucian interpretation.

8. Qur’an, Sūrah Ar-Rūm 30:22; used here with the English translation “The Clear Quran” by Mustafa Khattab. Text and translation. One translation and one verse do not represent Islamic ethics or interpretation as a whole.

9. Sakızlı. Question-Expansion Model, provided German original version, pp. 1–3. Documents the internal model structure A → B → D → C; not empirical evidence of effectiveness.

10. Sakızlı. Consultation Model, provided German original version, basic cycle pp. 7–8 and cases pp. 20–33. Documents the distinction between the four-stage basic cycle and five-stage case scenarios; not a validation study.

11. Sakızlı. 5-Phase Advisory Model, provided German original version, pp. 1–3 and 20. Not applied in this article; mentioned to distinguish consultation from case-specific specialist review.

Limits of this article

The cases, maps, and procedures developed here are editorial, fictional syntheses. The value-conflict map has not been empirically validated; the article is not a legal, contractual, or system review. An actual use of AI would require a new review of the specific workflow, data, roles, affected people, applicable requirements, and available means of correction.