AI Ethics in Saudi Arabia

AI ethics becomes practical when a business must choose between outcomes that serve different legitimate interests. A recommendation can be technically accurate while leaving an important question unresolved: which objective should govern the decision, who carries its burden, and whether the proposed means are justified. The organisation needs a reasoned answer that people can examine and challenge.

For Saudi businesses, ethical deliberation sits alongside applicable law, contractual obligations and the SDAIA AI Ethics Principles. This guide focuses on difficult decisions and the evidence needed to resolve them. Use the Saudi AI authority hub for the wider framework and responsible AI design for translating a decision into a working service.

Recognise an ethical choice before assigning a technical score

Questions about whose request receives priority, how much information to collect, what explanation to provide or when a person should intervene can contain competing values. A team may favour efficiency, consistency, privacy, access and commercial sustainability at the same time. Writing all of them into a policy does not determine how a specific tension should be resolved.

First distinguish an unresolved value choice from a breach of an applicable requirement. Ethical discussion cannot authorise an unlawful act or erase a contractual obligation. Establish the constraints with the appropriate owner, then identify the choices that remain within them.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence recognises that values and principles can be in tension and calls for contextual assessment. The method developed below is a business decision exercise. It is not a claim that one formula produces an official ethical approval or a universally correct allocation.

A useful opening question is: “Which decision would still be difficult if the model’s prediction were perfectly accurate?” The answer often reveals the policy choice being hidden inside an optimisation target. Resolving that choice should precede fine-tuning the system’s preference for one outcome.

Separate verified facts from persuasive descriptions

Prepare a fact sheet with four categories: verified, reported but unconfirmed, inferred and unknown. Include the source and the decision that depends on it. This prevents urgency, customer importance or model confidence from becoming a substitute for evidence.

A salesperson’s statement that a customer “must have everything today” may be accurate, exaggerated or simply incomplete. The organisation may need to know the minimum usable quantity, the real deadline and the consequence of partial fulfilment. Those facts can change the decision without requiring an intrusive account of the customer’s whole business.

Identify assumptions the system makes from historical data. Frequent purchasers may appear to have greater need because their records are more complete. A customer with fewer interactions may simply have less evidence in the system. Treat those differences as questions for assessment rather than automatic measures of deservingness.

Keep uncertainty visible. When a fact could materially change the choice, decide whether to obtain it, make a reversible interim decision or defer the relevant part. Do not let the model fill the gap with a confident narrative that cannot be traced to a source.

Worked example: allocating a scarce business supply

A fictional materials supplier has 100 units available and requests totalling 130 units. For this exercise, 60 units are already committed under verified arrangements that the authorised team has treated as a fixed constraint. The remaining requests comprise 45 units for repair work and 25 units for a new project. There are 40 units left to allocate against those 70 requested units.

The commercial team proposes prioritising the new project because it has a higher expected margin. The operations team points to repair customers whose work may be delayed. An AI recommendation ranks the project first using the selected margin objective. Its ranking does not resolve whether that objective should govern the remaining allocation.

Further checks establish that 35 of the 45 repair units are associated with verified time-sensitive work under the exercise assumptions. The remaining 10 repair units and the 25 project units also have legitimate commercial uses. There is no assumption of an immediate safety emergency. The ethical question concerns allocation within the stated constraints, not permission to disregard a binding obligation.

The team considers several feasible options. All allocate the same 40 available units. Each leaves 30 requested units unfulfilled, but distributes that shortfall differently. The example illustrates deliberation; it does not prescribe a standard policy for real supply contracts.

Three feasible allocations with different reasons

Option	Repair units	Project units	Reason and issue to examine
Commercial-priority allocation	15	25	Serves the full project request; leaves 30 repair units unmet
Time-sensitive allocation	35	5	Addresses verified time-sensitive repair demand; limits the project contribution
Approximately proportionate allocation	26	14	Shares the available units roughly by request size; whole-unit rounding is explicit

Do not label an option fair merely because it is easy to calculate. A partly supplied project may be unusable; a repair request may have a minimum workable quantity. Check those conditions before comparing outcomes. A feasible stock allocation can still fail the practical purpose of the recipients.

Ask what each option protects and what it sacrifices

Describe the strongest legitimate reason for each option before choosing. Commercial sustainability can support continued service and employment. Responding to verified urgency can prevent avoidable disruption. Proportionate sharing can reduce the impression that one relationship determines access. These considerations can matter without all receiving the same weight in every circumstance.

Then describe the burden. Who receives less, what consequence follows and whether an alternative reduces that consequence. A delayed delivery, substitute material or staged order may change the available choices. The ethical review should examine such alternatives instead of treating the initial allocation table as the entire solution space.

Avoid a single weighted “ethics score” unless its interpretation is genuinely justified. Assigning numbers to incomparable considerations can conceal the very judgment that needs scrutiny. A written reason, its evidence and the rejected alternatives may be more useful than a total that appears objective but depends on unexplained weights.

The final decision-maker should explain why the chosen reason is appropriate for this case and what would change the answer. That condition makes the decision open to evidence rather than a fixed preference dressed in technical language.

Examine power, incentives and conflicts of interest

Consider who can influence the process. A large customer may have direct access to senior management while smaller customers use a standard queue. A salesperson may be rewarded for margin, while service teams carry the cost of complaints. These incentives do not make people’s views irrelevant, but they affect which evidence and perspectives reach the decision.

Record relevant conflicts and provide a route for a view that challenges the preferred outcome. An employee should be able to point out an unsupported assumption without having to accuse a colleague of bad intent. Ask for the missing fact or unexamined consequence in concrete terms.

Where affected people cannot participate directly, identify how their interests will be represented. This might involve reviewing appropriate service records, asking an authorised representative or examining the actual consequences of partial fulfilment. Avoid assuming that the person with the most complete data or the loudest escalation represents the greatest need.

Apply the same scrutiny to an AI-generated recommendation. The system may reproduce a priority built into its objective or source material. It does not provide an independent ethical viewpoint merely because the recommendation is presented in neutral language.

Collect enough evidence without expanding the intrusion

Ethical assessment sometimes creates pressure to collect more information about affected people. Ask whether each requested fact can change the decision and whether a less intrusive way of establishing it is available. Verifying a delivery deadline may be sufficient without obtaining a customer’s complete project file.

For personal data, resolve the applicable questions through the PDPL and AI compliance guide and the responsible privacy owner. Ethical usefulness does not establish a lawful processing basis. Conversely, possessing information for one permitted purpose does not automatically justify using it to rank people or infer their circumstances for another purpose.

In the supply example, record the relevant urgency confirmation, its source and any limitations. Restrict access to commercial details that other participants do not need. A decision group can often consider the consequence and confidence of a fact without seeing every underlying document.

Decide how long the evidence is needed and how corrections will be handled. A temporary urgency assessment should not silently become a permanent label attached to a customer. Future decisions need current facts, not an inherited moral judgment.

Explain the decision without misleading or exposing others

An explanation should identify the material reason in terms the recipient can understand. It should not invent a neutral-sounding capacity problem if the actual decision followed another criterion. Equally, explanation does not require disclosing another customer’s confidential records or commercially sensitive information.

Prepare an explanation for the affected party and a fuller internal record for authorised review. The first should communicate what was decided, what relevant constraint or criterion applied and how a factual error can be raised. The internal record should show the evidence, alternatives and authority behind the choice.

Test whether the explanation remains true when read without the internal discussion. A statement such as “the system decided” obscures the organisation’s role in selecting the objective and using the output. Describe the accountable process instead, including the person or role that can examine a challenge.

Use clear Arabic and English where needed. Translation should preserve the distinction between a confirmed fact, a reported need and an organisation’s judgment. A smoother phrase that removes uncertainty can materially change the justification offered to a customer.

Use a short deliberation record for consequential choices

A decision record should help a later reviewer reconstruct the reasoning. Keep it proportionate to the consequence and uncertainty. For a material allocation choice, a useful record includes the following fields.

• The decision, available options and constraints already established.
• The affected parties and consequences considered.
• The facts relied on, assumptions and unresolved questions.
• The competing reasons and any relevant conflict of interest.
• The chosen option, accountable owner and explanation of why it was selected.
• The conditions for correction, reconsideration or expiry.

Retain a meaningful dissent where it identifies a relevant issue. Agreement in the room does not establish that all consequences were understood. A dissenting view can help design a later test or identify the condition that should trigger reconsideration.

The record should not become a way to legitimise a decision already fixed by commercial pressure. Establish who can ask for further evidence and who has authority to pause the proposed action. The human oversight guide explains how to make those interventions usable in the operating workflow.

Test consistency with counterfactuals and boundary cases

Ask whether the reasoning would change if irrelevant details changed. Would the same verified need receive a different outcome because the account manager is more influential? Would a newly acquired customer with equivalent evidence receive the same consideration? These questions expose assumptions that a general fairness statement may miss.

Also change a relevant fact. If the project requires a minimum delivery that cannot be met by partial allocation, the comparison may change. If the urgency evidence is withdrawn, an urgency-based reason needs reconsideration. Consistency means applying the stated reasoning to the facts, not forcing identical outcomes across different circumstances.

Use these cases to test the implementation after the decision method is agreed. Check the objective, data fields, business rules, generated explanation and human intervention together. An ethical decision recorded in a workshop can be lost when a developer simplifies it into a ranking feature.

Document unresolved cases instead of hiding them behind a model threshold. The risk-management framework guide helps organise consequences and uncertainty, while the final value judgment remains with the authorised organisation.

Teach staff to reason through a case

Training should give staff practice separating facts, constraints and preferences. Present a case with missing information and ask what they would need before deciding. Then introduce a new fact and examine whether their reasoning changes appropriately. The aim is thoughtful judgment within authority, not memorisation of a preferred answer.

Use roles with different perspectives: a customer representative, commercial manager, operations owner and privacy specialist may identify different omissions. Keep the discussion anchored in the actual decision. Broad philosophical debate is less useful if nobody can say which field, rule or approval should change.

The AI literacy guide for leaders provides a foundation for understanding what the system can and cannot establish. Ethical deliberation adds another layer: deciding what the organisation should ask it to do and what evidence is needed before acting on the result.

Assess the reasoning as well as the answer. A participant who reaches the expected outcome by inventing a fact has not demonstrated sound judgment. A different conclusion with a clear, lawful basis may reveal a policy choice the organisation has not yet resolved.

Review the effect of the decision, including the burdens

After implementation, examine whether the consequences match the assumptions. In the supply example, a partial delivery might prove unusable, or an apparently urgent request might have had a workable alternative. These observations can inform future criteria without automatically showing that the earlier decision was unreasonable on the evidence then available.

Look for repeated concentration of burdens. If the same group of requests repeatedly carries the shortfall, assess whether the underlying service policy, capacity or objective needs revision. Do not respond only by rewriting each individual explanation.

Review complaints and successful corrections alongside commercial results. A decision method should be maintainable under real workload and capable of recognising its own limits. Keep the basis for material changes visible so that staff understand why the organisation’s approach has evolved.

For a concrete case requiring structured discussion, consulting can help frame the decision, while governance setup addresses the authority and evidence route around it. Bring the real tension with confidential details removed; a specific case makes the required work easier to define.