What is AI ethics?
AI ethics is the study of how AI should be built, tested and used so it doesn't hurt people, and the rules that put those answers into practice. UNESCO frames it around four values: human dignity and human rights, just and peaceful societies, diversity and inclusiveness, and a flourishing environment. The five themes below turn up in almost every framework.

- Fairness: the system shouldn't treat people worse because of who they are.
- Transparency: people should know when AI is involved and how it reached a result.
- Privacy: personal data is protected across the whole life of the system.
- Accountability: a named person or company answers for harm, and the system can be audited.
- Safety: the system is tested for harm and secured against attack.
No body agrees on one fixed list. UNESCO names ten principles, the OECD names five, and the EU skips principles and writes law. That is why the table in the next section matters more than any single list.
What are the 5 ethics of AI?
The five most cited are fairness, transparency, privacy, accountability and safety, and each one appears under its own name in the major texts. Some pages swap in explainability for safety. That drops a principle every lab now writes policy about.
Here is where each one lives in the primary sources, as of November 2026:
- Fairness: UNESCO lists "Fairness and non-discrimination". The OECD puts it under "Human rights and democratic values, including fairness and privacy". The EU AI Act bans social scoring outright.
- Transparency: UNESCO and the OECD both name "Transparency and explainability". The EU AI Act requires chatbot users to be told they are talking to AI.
- Privacy: UNESCO lists "Right to privacy and data protection". OpenAI's Model Spec has a rule to protect people's privacy.
- Accountability: UNESCO's "Responsibility and accountability" says systems should be auditable and traceable. The OECD lists "Accountability" on its own.
- Safety: UNESCO has "Safety and security", the OECD has "Robustness, security and safety", and Anthropic's Responsible Scaling Policy ties deployment to catastrophic risk tests.
Two corrections to a common version of this list. First, accountability sits with the people who deploy a system as well as the people who build it, and the EU AI Act gives deployers their own duties. Second, no framework promises that an AI system will never cause harm. They ask for risk assessment and mitigation, which is a lower and more honest bar.
How do the major AI ethics frameworks compare?
Seven frameworks cover most of what people mean by AI ethics, and they differ most on who can enforce them. The table below states what each one says on its own page, as of November 2026. A cell that the page leaves silent says not stated.
| Framework | Issued by and when | Covers | Enforced by |
|---|---|---|---|
| Recommendation on the Ethics of AI | UNESCO, November 2021, 193 member states | 4 values and 10 principles, including proportionality and do no harm, human oversight and sustainability | Not stated on its page |
| AI Principles | OECD, adopted 2019, revised 2024, 47 adherents | 5 principles: inclusive growth, human rights and fairness, transparency, robustness and safety, accountability | Not stated on its page |
| AI Act | European Union, in force 1 August 2024 | 4 risk tiers, 9 groups of banned practices, duties for high risk systems and general purpose models | Fines up to 35 million euros or 7% of worldwide turnover |
| AI Risk Management Framework | NIST (United States), 26 January 2023, Generative AI Profile 26 July 2024 | 4 functions: Govern, Map, Measure, Manage | Voluntary use |
| AI Principles | Bold innovation, responsible development and deployment, collaborative progress | Google itself | |
| Model Spec | OpenAI, draft 8 May 2024, updated 12 February 2025 | Rules such as protect privacy, comply with laws, respect creators | OpenAI itself |
| Responsible Scaling Policy | Anthropic, 19 September 2023 | Safety levels ASL-1 to ASL-3, no deployment of ASL-3 models with meaningful catastrophic misuse risk | Anthropic itself |
The pattern is plain. UNESCO and the OECD set shared values for governments. Only the EU attaches fines. NIST offers a method companies can choose to follow. Labs write their own rules and mark their own homework, so a lab policy tells you what a company intends, not what a regulator checked.
UNESCO is the best primary source on values. Its page points to 11 policy action areas but doesn't list them in the text, so this page summarizes the values and principles only.
What are UNESCO's ten principles of AI ethics?
UNESCO's Recommendation lists ten core principles under a human rights approach, in this order on its page:
- Proportionality and do no harm
- Safety and security
- Right to privacy and data protection
- Multi-stakeholder and adaptive governance and collaboration
- Responsibility and accountability
- Transparency and explainability
- Human oversight and determination
- Sustainability
- Awareness and literacy
- Fairness and non-discrimination
Two of these rarely make the short lists. Human oversight and determination says a person, not the system, keeps the final call. Sustainability ties AI to its environmental cost, which fits the fourth UNESCO value, the flourishing of the environment and ecosystems. UNESCO also runs the Global AI Ethics and Governance Observatory, which it describes as a resource for policymakers, regulators, academics, the private sector and civil society, showing how ready countries are to adopt AI responsibly.
How do companies put AI ethics into practice?
They turn principles into process: a risk assessment before launch, records of what the system does, and a person who can overrule it. NIST's framework organizes that work into four functions, Govern, Map, Measure and Manage, and says it is for voluntary use. For generative AI, NIST added a Generative AI Profile on 26 July 2024.
The EU makes a similar list mandatory for high risk systems. The European Commission names risk management, dataset quality, logging, documentation and human oversight as the strict obligations. Most of this is dull paperwork, and that is the point: ethics that nobody can audit stays a slogan.
What do real cases of unethical AI look like?
Three well documented cases show the principles failing, each in a different way.
- Fairness, COMPAS (2016). ProPublica analyzed risk scores for more than 7,000 people arrested in Broward County, Florida, in 2013 and 2014. It reported that 44.9% of Black defendants who did not re-offend were labeled higher risk, against 23.5% of white defendants. It also found that only 20 percent of the people predicted to commit violent crimes did so.
- Accountability, Dutch childcare benefits. Amnesty International reported that the Dutch tax authorities' algorithm used whether an applicant had Dutch nationality as a risk factor, and that tens of thousands of parents and caregivers, mostly from low income families, were falsely accused of fraud. A parliamentary inquiry reported in December 2020 that the parents had been left powerless to fight back, and the Dutch government resigned on 15 January 2021, as DutchNews reported at the time.
- Transparency and truth, Mata v. Avianca (2023). Two lawyers filed a brief that cited fake cases made up by ChatGPT. On 22 June 2023 Judge P. Kevin Castel of the Southern District of New York ordered the lawyers and their firm to pay 5,000 dollars, finding both acted in bad faith through conscious avoidance and false and misleading statements to the court, as the ABA Journal reported. The case turned a hallucination into a sanction.
None of the three was a robot uprising. Each was an ordinary system used without anyone checking it. The AI hallucination statistics page covers how often models invent facts, and how to fact check AI answers covers the check that would have caught the third case.
Who is responsible when AI causes harm?
The people and companies that build, sell and deploy the system are responsible, and the frameworks say so in different words. UNESCO says AI systems should not displace ultimate human responsibility and accountability. The EU AI Act splits the duty: providers of high risk systems must keep a risk management system, technical documentation and logs, while deployers carry a smaller set of duties.
The cases match that split. In Mata v. Avianca the court sanctioned the lawyers who filed the motion, not the chatbot. In the Dutch case the state's tax authority ran the system, and the government resigned in the aftermath. A model can't be fined or resign. A person or an organization has to be able to answer.
For you as a user, the practical rule is short. If you put your name on AI output, you own it. Check it before you send it, file it or grade with it.
What does the EU AI Act require now?
The EU AI Act is the only framework in the table with fines, and it phases in over several years. The European Commission says it sorts AI into four tiers: unacceptable risk, which is banned, high risk, transparency risk and minimal risk. Chatbots must tell people they are talking to AI, and deepfakes must be labelled.
Per the European Commission's AI Act page, as of November 2026: the Act entered into force on 1 August 2024, prohibited practices and AI literacy duties applied from 2 February 2025, general purpose model rules from 2 August 2025, and general application from 2 August 2026, with some exceptions. The Commission's page says its AI Omnibus, in force since 27 July 2026, moved the high risk duties for Annex III uses such as biometrics, education and employment to 2 December 2027, and for AI built into regulated products to 2 August 2028. It also added a ninth banned practice, covering AI that generates non consensual sexual content, from December 2026.
Article 99 of Regulation (EU) 2024/1689 sets the fines. Banned practices cost up to 35 million euros or, for a company, up to 7% of worldwide annual turnover, whichever is higher. Breaches of the other listed operator duties cost up to 15 million euros or 3%, and for small firms and start-ups the lower of the two figures applies. The dates have already moved once, so check the Commission page before you rely on one.
What should you ask an AI about ethics?
Ask a model to apply a framework to a case, not to tell you what is right. These five prompts work in Claude, ChatGPT, Gemini or Whizi, and each one asks for sources you can open.
- "Summarize the [UNESCO or OECD or EU AI Act] position on [topic, such as facial recognition] in five sentences. Quote the exact principle name and give me the link to the official page."
- "Here is how my company uses AI: [describe]. Which EU AI Act risk tier is it, and which duties apply? Say which parts you are unsure of."
- "Argue the strongest case for and against [using AI to screen job applicants] and cite one real incident for each side."
- "Review this prompt and output for bias: [paste]. Rewrite the prompt with the gender, age and country swapped and tell me every difference in the answers."
- "I am writing a school policy on AI use for [grade]. Draft one page from the five principles in [framework], and mark which parts are my choice, not the framework's."
A model can't certify that something is ethical. It can lay out the arguments and the rules, and it can get them wrong, so open every source it names.
Where does Whizi fit?
Whizi is one app for Claude, GPT and Gemini, so you can put the same ethics question to two labs and read where they differ. Its privacy facts are short. Whizi states that it does not use your prompts, files or conversations to train Whizi owned models, and that it does not sell that content as training data.
It also says what it can't promise. Routed AI providers may temporarily cache part of a prompt or reply under their own terms, and uploaded files expire after up to 30 days. Whizi doesn't claim zero retention at every provider. You can delete a single conversation without deleting your account.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card, and Pro's first week is $1.99. See chatbot privacy compared for how other apps handle the same question.
What can AI ethics frameworks not do?
Most of them can't force anyone to act. Of the seven in the table, only the EU AI Act attaches fines, and the lab policies are the labs' own promises. A framework also can't settle a real trade off, such as privacy against safety, because it names both and leaves the choice to people.
They date quickly, too. OpenAI updated its Model Spec within a year of its first draft, and the EU has moved its high risk dates since the law passed. Treat any list of principles as a snapshot and check the date on the page.
For the plain pros and cons of using AI day to day, read is AI bad. For a short safety answer, see is AI safe to use.
- Read the training line in a tool's privacy policy before pasting anything private
- Check for a named company and a way to report a bad output
- Swap a name, age or country in a prompt to test for bias
- Open the official page behind every rule or fine an AI quotes
- Check the law for your use case, especially hiring, credit and school marks
Frequently asked questions
What are the 5 ethics of AI?
Fairness, transparency, privacy, accountability and safety are the five most cited. UNESCO and the OECD both name transparency and explainability, accountability and safety or security, and both cover fairness and privacy. No framework fixes the list at exactly five.
Why is AI ethics important?
Because AI systems already decide things about people, and the cases show real harm. ProPublica reported higher false flag rates for Black defendants in 2016, and tens of thousands of Dutch parents were falsely accused of fraud, according to Amnesty International. Ethics frameworks give regulators and companies shared tests for that harm.
Is the EU AI Act the same as AI ethics?
No. The EU AI Act is a law with fines up to 35 million euros or 7% of turnover. AI ethics is the wider set of principles, and the Act turns some of them into legal duties.
Do AI companies publish their own ethics rules?
Yes. OpenAI publishes its Model Spec, Anthropic its Responsible Scaling Policy and Google its AI Principles, as of November 2026. They are company promises, not law, so no regulator checks that the company follows them.