AI ethics
AI ethics covers the principles and trade-offs that determine whether AI systems are built and used responsibly, including in areas where no law provides an answer.
Ethics starts where the law is silent or not yet in force.
AI ethics is applied ethics. It asks which uses of AI are defensible: whether a system treats people fairly, whether those affected know AI is involved, who can stop a result, and who is accountable for harm. Companies often call it responsible AI and mean the practice: rules, ownership, and reviews.
In April 2019, the European Commission's High-Level Expert Group on AI published Ethics Guidelines for Trustworthy AI. The OECD adopted its AI Principles in 2019 and updated them in May 2024. In November 2021, UNESCO's 193 member states adopted a Recommendation on the Ethics of Artificial Intelligence. None of them binds a company legally.
The law is binding, in the EU above all the AI Act and the GDPR. Recital 27 of the AI Act refers explicitly to the expert group's seven principles, and several of them reappear as obligations. Anything beyond that is each company's call, and it belongs to senior management because it affects reputation, liability, and trust.
Chapter 1 of 7
Three frameworks set the benchmark.
The EU expert group's Ethics Guidelines of April 8, 2019, name three components of trustworthy AI: it should be lawful, ethical, and robust. Seven requirements follow: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination, and fairness; societal and environmental well-being; and accountability.
The OECD AI Principles list five values-based principles, including transparency and explainability and accountability, plus five recommendations for governments. The May 2024 revision reflects developments since 2019.
UNESCO calls its Recommendation of November 2021 the first global standard on AI ethics. It rests on four core values, including human rights and human dignity, and ten principles such as proportionality and human oversight. Eleven policy areas describe how states should apply it.
Chapter 2 of 7
The AI Act turns some ethical principles into law.
Law sets a floor for everyone; ethics asks what a company wants to do beyond it. Article 5 of the EU AI Act, Regulation (EU) 2024/1689, bans certain practices, such as inferring emotions in the workplace. For high-risk systems it requires transparency (Article 13) and effective human oversight (Article 14). Article 95 provides for voluntary codes of conduct, which can include elements of the EU ethics guidelines.
Under amending Regulation (EU) 2026/1744, the obligations for high-risk systems listed in Annex III apply from December 2, 2027. These include AI that screens job applications or evaluates candidates. Article 4 now requires providers and deployers to support their staff's AI literacy, without guaranteeing a particular level (as of November 2026).
Ethics starts where the law is silent or not yet in force. One example is whether a company wants to evaluate its employees with AI at all.
Chapter 3 of 7
Responsible AI turns principles into processes.
In its AI Risk Management Framework from January 2023, the US National Institute of Standards and Technology (NIST) calls risk management a key component of responsible development and use of AI. It lists the characteristics of trustworthy AI as valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. The framework is voluntary.
In practice, it comes down to a few building blocks. Each AI system has a named owner who decides on deployment and shutdown. A pre-launch review clarifies who the result affects, what an error would cause, and whether any group could be disadvantaged. For consequential decisions about people, a person can override the result. Those affected have a way to complain. The decisions are documented.
Chapter 4 of 7
How ethics becomes auditable.
A principle like fairness can only be checked once it is translated into a concrete question with a measure. The EU expert group published the Assessment List for Trustworthy AI (ALTAI) in July 2020, a self-assessment list. It turns the seven requirements into questions a team answers for its system.
A short translation per principle is usually enough. For fairness, results are compared across groups, for example the share of applications passed on. Transparency means people learn AI was involved and can get a plain explanation on request. For oversight, it is settled who may stop a result. Accountability means reviews, approvals, and changes are recorded with a date and a name.
Auditable also means outsiders can verify it: internal audit or an external auditor compares practice with the company's own rules and recognized frameworks.
Chapter 5 of 7
Three misconceptions come up often.
A published code of conduct alone isn't AI ethics. In 2019, philosopher Luciano Floridi described five risks in putting digital ethics into practice, among them ethics bluewashing: superficial claims of ethical behavior that create a better image than the practice supports. The remedy is evidence. Stated principles should come with the review behind each one.
Ethics does not replace a legal review. Voluntary principles come on top of the obligations under the AI Act, the GDPR, and anti-discrimination law.
Ethics does not deliver clear-cut answers, either. Principles sometimes pull against each other: a more accurate model can be harder to explain, and more transparency about decisions can clash with protecting personal data. That calls for a reasoned trade-off that someone owns.
Smaller companies face the same trade-offs when they use AI in hiring or customer service, just with fewer systems.
Chapter 6 of 7
Who decides and how much effort it takes.
Responsibility sits with senior management because it decides on use, risk, and reputation. It can delegate: a small committee from the business unit, IT, data protection, and employee representatives prepares policy questions, and each system's owner carries them out.
Effort depends on the number and kind of systems. A writing assistant needs a short rule and a training session. Letting AI help decide on applications, loans, or benefits, though, calls for a documented review per system, regular analysis, and a complaints channel. There is no standard market price for consulting or audits.
A useful first step is a list of all AI systems in use, with purpose, affected people, and owner. It shows where a short note will do and where a full review is needed. Permissions, approvals, and logs are a matter for AI governance.
Chapter 7 of 7
Where the term stops working
AI ethics answers what is defensible. Compliance checks adherence to the AI Act and the GDPR; AI governance sets permissions, approvals, and logs. AI bias is a specific failure that AI ethics handles under fairness, and AI security protects systems against attacks and malfunctions. Machine ethics, whether machines can act morally themselves, is a separate research field.
Frequently asked questions
What is AI ethics?
AI ethics is the study of which uses of AI are defensible and the practice of applying that inside an organization. It deals with fairness, transparency, human oversight, privacy, and accountability, even where no law applies.
What is responsible AI?
Responsible AI means putting ethical principles into processes, ownership, and reviews. NIST calls risk management a key component of responsible development and use of AI. It is often a synonym for AI ethics.
Is AI ethics required by law?
No, guidelines from the EU expert group, the OECD, or UNESCO don't bind companies legally. The EU AI Act and the GDPR are binding, and some ethical principles reappear in them as obligations, such as human oversight of high-risk systems.
What is the difference between AI ethics and AI governance?
AI ethics defines what a company considers defensible; AI governance sets who may do what and how that is checked. Governance is how ethical principles and legal obligations hold up day to day.
What are the principles of AI ethics?
The 2019 EU guidelines name seven requirements, including human oversight, transparency, fairness, and accountability. The OECD has five principles and UNESCO ten. Each company decides which ones it adopts and how it checks them.
Who is responsible for AI ethics in a company?
Senior management holds overall responsibility because it decides on use and risk. Each AI system should have a named owner who decides on operation and shutdown.
Sources
- European Commission, High-Level Expert Group on AI: Ethics Guidelines for Trustworthy AI, April 8, 2019, with ALTAI from July 2020
- OECD: AI Principles, 2019, updated May 2024
- UNESCO: Recommendation on the Ethics of Artificial Intelligence, November 2021
- NIST: Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023
- European Union: Regulation (EU) 2024/1689 (Artificial Intelligence Act), Recital 27, Articles 5, 13, 14, and 95, and Annex III, 2024
- European Union: Regulation (EU) 2026/1744 amending the AI Act (Digital Omnibus on AI), new Article 4 and high-risk application dates, 2026
- European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22, 2016
- Floridi: Translating Principles into Practices of Digital Ethics: Five Risks of Being Unethical, Philosophy & Technology 32, 2019