In short
- Definition: The degree to which a system respects moral principles such as fairness, transparency, non-discrimination and respect for users’ rights.
- Goal: Prevent the system from causing harm or treating people unfairly.
- Example S.M.A.R.T. requirement: An ethics impact assessment approved before launch, and no group with an outcome rate below 80% of the best-treated group.
1. Definition
As a non-functional requirement, ethics describes how well a system respects moral principles such as fairness, non-discrimination, transparency, respect for privacy and respect for users’ rights and autonomy [1], [2].
Ethical questions do not disappear when nobody writes them down. They come back as complaints, headlines or lawsuits.Key takeaway
Ethical requirements matter most for systems that make or support decisions about people: credit, hiring, insurance, health, policing, content moderation, and anything using artificial intelligence. The OECD AI Principles, for example, call for human-centred values and fairness, transparency and explainability, and accountability [3].
Ethics also covers design choices, such as avoiding deceptive patterns that push users into choices they would not make freely.
2. Goal
The goal is to prevent the system from causing harm or treating people unfairly, and to keep the trust of users and society.
Unexamined systems can discriminate at scale, manipulate users or make decisions nobody can explain, with legal, financial and reputational consequences.
3. How to Measure It (S.M.A.R.T.)
| Metric | Unit | Typical target | How to measure |
|---|---|---|---|
| Ethics impact assessment | Completed and approved | Before launch and after major changes | Assessment report signed by the review board |
| Outcome disparity between groups | Ratio of approval rates | ≥ 0.8 of the best-treated group | Bias test on representative data |
| Explanations available | % of automated decisions with an explanation | 100% for decisions affecting individuals | Functional testing |
| Deceptive patterns | Count found in review | 0 | UX review against a checklist |
The 0.8 ratio, often called the four-fifths rule, is one common indicator of disparate impact [4]; the right fairness measure depends on the context and should be chosen with experts.
4. Example Requirements
Each requirement below is ready to copy into a specification: specific, measurable, achievable, relevant and time-bound.
- Automated decisions: “Before launch, the loan pre-approval model shall pass a bias test showing no protected group with an approval rate below 80% of the best-treated group.”
- Governance: “An ethics impact assessment shall be approved by the review board before go-live and repeated after any change to the decision model.”
- Design: “The subscription flow shall contain no pre-checked consent boxes and offer cancellation in no more steps than sign-up.”
5. Questions to Ask Stakeholders
- Does the system make or influence decisions about people?
- Which groups could be affected unfairly?
- Which values and principles does the organization commit to?
- Can people understand and contest decisions that affect them?
- Who reviews ethical risks, and when?
- Are there AI or anti-discrimination laws that apply?
6. Common Pitfalls
- Treating ethics as a one-time sign-off. Data and models change. Re-assess regularly.
- Removing sensitive attributes and stopping there. Other data can act as proxies. Test outcomes, not only inputs.
- No route to contest decisions. Provide human review and an appeal path for decisions that matter.
7. Related NFRs
- Categories of Non-Functional Requirements — the full checklist of NFR categories.
- Transparency — explanations are a core ethical safeguard.
- Privacy — respecting personal data is an ethical duty as well as a legal one.
- Compliance — AI and anti-discrimination laws turn some ethical principles into obligations.
- Safety — avoiding harm to people is shared by safety and ethics.
8. FAQ
Parts of it are: bias tests, completion of impact assessments, availability of explanations. Others need human judgment and review.
A rule of thumb that flags possible discrimination when a group’s selection rate is below 80% of the most favored group’s rate.
No. Any system can raise ethical questions, for example through manipulative design or unfair rules.
9. Acronyms, Initialisms & Terms
Acronyms are abbreviations pronounced as words (like NASA); initialisms are spelled out letter by letter (like FBI). The classification below follows common pronunciation.
Acronyms
| Abbreviation | Stands for | Definition | Ref. |
|---|---|---|---|
| SMART | Specific, Measurable, Achievable, Relevant, Time-bound | Criteria for writing a goal or requirement so that anyone can check whether it has been met. | [5] |
Initialisms
| Abbreviation | Stands for | Definition | Ref. |
|---|---|---|---|
| AI | Artificial intelligence | Computer systems that perform tasks usually needing human intelligence, such as recognizing images or making predictions. | [6] |
| FAQ | Frequently asked questions | List of common questions with their answers. | [7] |
| NFR | Non-functional requirement | Requirement that describes how well a system must work (its qualities, such as speed or security) rather than what it does. | [2] |
| OECD | Organisation for Economic Co-operation and Development | International organization of 38 countries that develops policy standards, including the OECD AI Principles. | [8] |
| UX | User experience | Everything a person experiences when using a product: ease, emotions, and the whole journey. | [9] |
Terms
Plain-language definitions of technical and project-management terms used in this article.
| Term | Plain-language definition | Ref. |
|---|---|---|
| Governance | Rules, roles and decision processes that steer how IT is managed. | [10] |
| Personal data | Any information that relates to an identifiable person. | [11] |
| Stakeholder | Person or group with an interest in the system: users, sponsors, operators, regulators. | [12] |
10. Sources and Further Reading
- [1] “Algorithmic bias.” Wikipedia. https://en.wikipedia.org/wiki/Algorithmic_bias (accessed Nov. 18, 2026).
- [2] “Non-functional requirement.” Wikipedia. https://en.wikipedia.org/wiki/Non-functional_requirement (accessed Nov. 18, 2026).
- [3] OECD. “Recommendation of the Council on Artificial Intelligence (OECD AI Principles).” OECD Legal Instruments. https://legalinstruments.oecd.org/en/instruments/oecd-legal-0449 (accessed Nov. 18, 2026).
- [4] “Disparate impact.” Wikipedia. https://en.wikipedia.org/wiki/Disparate_impact (accessed Nov. 18, 2026).
- [5] “SMART criteria.” Wikipedia. https://en.wikipedia.org/wiki/SMART_criteria (accessed Nov. 19, 2026).
- [6] “Artificial intelligence.” Wikipedia. https://en.wikipedia.org/wiki/Artificial_intelligence (accessed Nov. 19, 2026).
- [7] “FAQ.” Wikipedia. https://en.wikipedia.org/wiki/FAQ (accessed Nov. 19, 2026).
- [8] “OECD.” Wikipedia. https://en.wikipedia.org/wiki/OECD (accessed Nov. 19, 2026).
- [9] “User experience.” Wikipedia. https://en.wikipedia.org/wiki/User_experience (accessed Nov. 19, 2026).
- [10] “Corporate governance of information technology.” Wikipedia. https://en.wikipedia.org/wiki/Corporate_governance_of_information_technology (accessed Nov. 19, 2026).
- [11] “Personal data.” Wikipedia. https://en.wikipedia.org/wiki/Personal_data (accessed Nov. 19, 2026).
- [12] “Project stakeholder.” Wikipedia. https://en.wikipedia.org/wiki/Project_stakeholder (accessed Nov. 19, 2026).
Written by Philippe Huard · LinkedIn · Last updated: November 18, 2026 · Header photo by @tingeyinjurylawfirm; inline photo by @sasun1990, both on Unsplash.








