An ethical dilemma in AI arises when two widely accepted principles, such as fairness and accuracy or privacy and transparency, cannot both be fully met in the same system. Most documents that list those principles are voluntary, including UNESCO's Recommendation on the Ethics of Artificial Intelligence, adopted on November 23, 2021, and the OECD AI Principles. Binding rules exist mainly in the European Union's AI Act, which has banned certain practices since February 2, 2025, and in a Council of Europe treaty that Canada signed on February 11, 2025.
Key takeaways
- UNESCO's Recommendation, adopted by 193 member states, lists ten principles without ranking them, so conflicts between them are left to the people building and using AI.
- A 2016 mathematical result by Kleinberg, Mullainathan, and Raghavan showed that three common fairness conditions cannot all hold at once outside special cases, which makes some trade-offs unavoidable.
- The OECD principles (2019, updated 2024) and the NIST AI Risk Management Framework (January 26, 2023) are voluntary.
- The EU AI Act bans nine practices, including social scoring and emotion recognition in workplaces and schools, and its high-risk rules were pushed back to December 2, 2027 and August 2, 2028.
- Canada signed the Council of Europe AI treaty on February 11, 2025, and the treaty takes effect only after five ratifications.
Why principles collide
The UNESCO Recommendation was adopted by the General Conference in Paris on November 23, 2021, and member states are asked to apply it on a voluntary basis. It names four values, among them human dignity and human rights, and ten principles: proportionality and do no harm, safety and security, privacy and data protection, multi-stakeholder governance, responsibility and accountability, transparency and explainability, human oversight, sustainability, awareness and literacy, and fairness and non-discrimination. The text does not set an order among them.
That gap is where dilemmas live. A system that explains every decision may expose personal data, and a system tuned to treat groups equally on one measure may treat them unequally on another. The mathematics behind the second case is exact. In the paper "Inherent Trade-Offs in the Fair Determination of Risk Scores", first posted on September 19, 2016, the authors formalize three fairness conditions for probability scores and show that, outside highly constrained special cases, no method can satisfy all three together.
Four common conflicts
The pairs below are illustrative readings of the UNESCO list, not findings from a study.
| Dilemma | Principles in tension | Typical question |
|---|---|---|
| Explaining versus protecting data | Transparency and explainability; privacy and data protection | How much of the training data or a person's record can be disclosed to justify a result? |
| Equal treatment versus equal outcomes | Fairness and non-discrimination; proportionality | Which fairness measure applies when the measures cannot all be met? |
| Speed versus review | Human oversight; safety and security | Which decisions need a person to approve them before they take effect? |
| Benefit versus harm | Proportionality and do no harm; responsibility | Is the use justified by its aim, and who answers if it fails? |
Voluntary frameworks
The OECD AI Principles were first adopted in May 2019 and updated in 2024. They rest on five principles: inclusive growth and well-being, human rights and democratic values including fairness and privacy, transparency and explainability, reliability, security and safety, and accountability. The OECD lists 47 adherents.
The NIST AI Risk Management Framework 1.0 was released on January 26, 2023 and is intended for voluntary use. NIST's page states that the framework is being revised as part of the White House AI Action Plan, with no completion date given. A Generative AI Profile followed on July 26, 2024. None of these documents carries a penalty. Their weight comes from procurement rules, contracts, and professional expectations that point to them.
Binding rules
The European Commission's AI Act page reports that the act entered into force on August 1, 2024. Prohibited practices and AI literacy duties applied from February 2, 2025. The nine prohibitions include harmful manipulation, social scoring, untargeted scraping of face images to build recognition databases, and emotion recognition in workplaces and education institutions. A 2026 amendment known as the AI Omnibus entered into force on July 27, 2026. It moved the high-risk rules for listed uses such as employment and education to December 2, 2027, and those for AI built into regulated products to August 2, 2028. It also adds a ban on systems that generate non-consensual sexual content, effective December 2026.
The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law was opened for signature in Vilnius on September 5, 2024. Global Affairs Canada announced Canada's signature on February 11, 2025, and says the convention takes effect only after five states, at least three of them Council of Europe members, consent to be bound. The Canadian announcement does not report a ratification. The Council of Europe Treaty Office holds the current count of ratifications.
A worked example: screening job applicants
Suppose a company in Toronto tests a scoring tool for résumés. This is an illustrative case, not a reported one. The team must pick a fairness measure, and by the 2016 result it cannot satisfy every common one. It must decide how much to tell applicants about how scores are produced without exposing other candidates' data. It must decide whether a person reviews each rejection. Each choice favors one principle over another, and none of the voluntary documents says which. Where the tool is sold into the EU, the Act's treatment of employment as a high-risk use is scheduled for December 2, 2027, and any attempt to infer applicants' emotions would fall under the workplace ban already in force.
A checklist for teams
- List the principles at stake for each AI use and mark where two of them pull apart.
- Write down which measure or trade-off was chosen, and why.
- Name the person who answers for the result and the person who can override it.
- Check which binding rules apply in each place the system is used, and note their start dates.
- Revisit the record when a framework, treaty, or law is revised.
The treaty, regulation, and dates named above are descriptions, not legal advice, and they can change. A team facing a real compliance decision should read the original texts or ask a lawyer.
FAQ
What is the main ethical dilemma in AI?
There is no single one. The recurring problem is that goals such as fairness, privacy, transparency, and safety can conflict, and no document ranks them.
Is the UNESCO AI ethics Recommendation legally binding?
No. It asks member states to apply its provisions on a voluntary basis.
Has Canada ratified the Council of Europe AI treaty?
The Canadian announcement of February 11, 2025 reports a signature only, and the Council of Europe Treaty Office holds the current ratification count. The treaty needs five ratifications to take effect.
Does the EU AI Act ban emotion recognition?
It prohibits emotion recognition in workplaces and education institutions, and that prohibition has applied since February 2, 2025.
Related reading
- the WHO view on AI ethics
- AI and copyright law
- ensuring morality in AI systems
- AI and Christian ethics
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Recommendation on the Ethics of Artificial Intelligence, UNESCO, November 23, 2021
- Inherent Trade-Offs in the Fair Determination of Risk Scores, Kleinberg, Mullainathan, and Raghavan, arXiv, 2016
- OECD AI Principles, OECD AI, 2019 (updated 2024)
- AI Risk Management Framework, National Institute of Standards and Technology, 2023
- AI Act regulatory framework page, European Commission, 2026
- Canada signs the Council of Europe Framework Convention on Artificial Intelligence, Global Affairs Canada, February 11, 2025