Core Ethical Principles: Fairness, Transparency, Privacy, Safety and Accountability

The first two pages gave you a method: ask who is affected, weigh benefits and harms, compare options, name a person, set a date. A method needs values to steer it. When you say "this seems unfair" or "this feels too invasive", you are using a principle, even if you have not named it. This page names the common ones, shows what each looks like in a real design choice, and shows how two international documents have grouped them.

The principles here are not a checklist that, once ticked, makes a tool ethical. They are questions that keep your attention on the right things. They also sometimes pull in different directions; the next page deals with that. Here we learn what each one asks for.

What you will learn

• What human dignity, fairness, avoiding harm and balancing risks mean as practical design questions, not slogans.
• How to explain transparency to a non-technical user, and why accountability is not the same as blaming one person.
• How privacy and data governance, accessibility and inclusion become concrete requirements.
• What the OECD AI Principles and UNESCO's Recommendation on the Ethics of AI are, who issued them, and that they are voluntary.
• How to run a short principles check on an AI use and write down what each principle asks of you.

One running scenario

We will test each principle against the same invented example. Read it once now and come back to it as you go.

Invented scenario:

Amara leads digital services for a city council. Residents ask about permits, bills and
benefit applications. The council adds an AI assistant to its website and phone app. It answers
questions in text, helps fill in forms, and sorts requests into "standard" or "needs a person".
Standard requests get an automatic reply. The council wants shorter queues at its front desk.
Residents include older people, people who do not read well, people who speak another language
at home, people using only a basic phone, and people using screen readers.

Human dignity and human rights as design considerations

Human dignity is the idea that every person has worth that does not depend on how useful, rich or easy to handle they are. Human rights are the basic protections that international and national texts describe, such as freedom from discrimination, privacy, and the ability to take part in public life. You do not need a law degree to use these ideas as design questions: does this design treat people as people, and could it remove something they are entitled to?

Dignity is easy to say and hard to measure. It helps to ask a very simple thing: if you were the resident, would you feel respected by this process at its worst moment, not at its best?

Fairness is not identical treatment

Many people think fairness means giving everyone exactly the same thing. Sometimes that is right: the same rules for everyone applying for the same permit. Sometimes it is wrong. If one candidate in an exam has a reading difficulty, giving everyone the same 60 minutes does not give them the same chance. Giving that candidate extra time is not favouritism; it removes a barrier that has nothing to do with what the exam measures. Fairness asks whether people are treated in ways that are justified by relevant differences, and whether irrelevant differences, such as background, language or disability, are not holding them back.

Be careful in the other direction too. Different treatment must be justified by a relevant reason and be open to question; it must not be a polite way of excluding people. Page 6 of this tutorial, Fairness and Its Trade-Offs: Inclusion, Accessibility and Language, takes the measurement side further. For now, remember that "same for everyone" and "fair" can differ, and that you should be able to say which relevant difference justifies a change.

Where reasonable people disagree

Should simple language and large text be the default for everyone, or offered on request?

Both views want residents to be served well. They differ on whether to guess what people need or to let them say. A council could combine them: plain language by default, with a visible option for more detail or a different format.

Avoiding unnecessary harm, and balancing benefits against foreseeable risks

The word unnecessary is important. Almost every useful tool carries some risk; the question is whether a particular harm could be avoided without losing the benefit. Foreseeable means that a reasonable person who thought about it for a short while could have expected it. Together these give two practical steps: first, remove harms you can remove cheaply; second, for those that remain, weigh them against the benefit, as step 3 and step 6 of the worksheet in Everyday AI Ethics: Who Benefits and Who Could Be Harmed ask you to do.

Weighing benefits against risks is not a calculation with one correct answer. It is a judgement made with evidence, by a named person, and open to being checked later. The more serious and less reversible the harm, and the more people who could not protect themselves, the more you need to see before you go ahead.

Transparency for a non-technical user

Transparency means people can find out that an AI is involved, what it is used for, and what it cannot do, in words they understand. It does not mean publishing source code or technical papers for a resident who needs a permit. Think of it as what a reasonable person would want to know to decide whether to trust, question or avoid the system.

Example notice:

You are talking to an AI assistant, not a person.
It can answer questions about permits and bills and help you fill in forms.
It can make mistakes. Check important dates on the official page we link to.
Your messages are kept for 30 days to fix problems, then deleted.
To speak to a person at any time, type "person" or call the front desk.
If you think an answer or a decision about you is wrong, tell us here: [contact].

Check the notice against five questions: Does it say that AI is involved? What it is for? Its limits? What happens to my data? Who can I talk to? A notice that answers these in plain language is better than a long policy that hides them. Notice also what transparency does not do: telling people about the AI does not by itself make the AI fair or accurate. It lets them make their own choices and object.

Accountability is not blaming an individual

Accountability means that for every AI use there is someone who can explain it, answer for its outcomes, and put it right. That is different from finding a person to blame after something goes wrong. Blame looks backwards and often falls on the nearest and least powerful person, such as a front-desk worker who relayed a wrong answer. Accountability looks forwards: who decided to use this tool, who checks it, who receives complaints, who can pause it?

This does not mean nobody is ever at fault. Careless or dishonest conduct still has consequences. It means the first question is how to make the next failure less likely, and that someone with authority answers for the answer, not only the person who happened to be nearest.

Privacy and data governance as design principles

Privacy is a person's reasonable control over information about themselves. Data governance is how an organisation decides what data it collects, why, who can see it, how long it is kept, and how it is protected. As design principles they push you to make choices before launch, instead of apologising after a leak. Four habits are enough for a beginner: collect only what you need; say why; keep it only as long as needed; limit who can see it.

The habit of keeping personal data out of chats came up in Prompt Engineering for AI: Beginner. The same idea applies to designers: not every piece of data that you could collect is data you should collect. Pages 9 and 10 of this tutorial return to privacy in detail.

Accessibility and inclusion as practical requirements

Accessibility means that people with different abilities and devices can use the service. Inclusion means that the design takes account of the variety of people who will use it, including those who are usually left out, rather than building for an imagined "average" user. These are requirements to check, not courtesies to add if there is time.

The simplest test is to involve people from these groups before launch and watch them use it. If that is not possible, at least check a few cases yourself. No one can guess every barrier, but you can usually find the first few by trying.

Two international sets of principles: OECD and UNESCO

Many of the ideas above appear in two widely cited international documents. They are useful as shared vocabulary across countries. They are not laws.

Legal note (as of 2026-10-12; not legal advice): The OECD AI Principles were adopted in 2019 and updated in May 2024. On 2026-10-12 the OECD.AI website listed 47 adherents, including the European Union. UNESCO's Recommendation on the Ethics of Artificial Intelligence was adopted in November 2021 by UNESCO's 193 Member States. Both documents are recommendations: they are voluntary and set no legal duty by themselves. A country, regulator or employer may build parts of them into binding rules, and which rules apply depends on your country, sector and date. Read the documents on the official OECD.AI and UNESCO websites (search for the document names there), and check the current official text for your country and sector.

Below is our own reading of how this page's principles line up with the two documents. It is a learning aid, not an official mapping, and the wording of each document is on its home page.

Two things to notice. First, different bodies group the same ideas under different headings, so do not memorise names; ask what each principle asks you to check. Second, adding up how many principles a company mentions says nothing about how it behaves. The next page returns to that. To read the originals, go to the OECD.AI website and UNESCO's website and look for the document names given above. Searching for a principle's name on a news site is a weaker route than reading the document itself.

A short principles check

You can use the principles as a quick review step inside the worksheet from page 1, usually at step 3 (benefits and harms) and step 6 (decision and conditions). It takes about ten minutes for a small use.

Template:

PRINCIPLES CHECK for: [use]
DIGNITY       Can a person reach a human? Could someone lose access to something they need?
FAIRNESS      Who might be held back by a difference that is not relevant? What would help?
HARM          Which harms are avoidable? Which remain, and are they worth the benefit?
TRANSPARENCY  Would a non-technical person know AI is involved, its limits, and who to ask?
ACCOUNTABLE   Who is the named person? Who receives complaints? Who can pause the tool?
PRIVACY       What data is collected, why, for how long, who can see it?
ACCESS        Who might not be able to use it? Who did we ask or test with?
NOT SURE      List what you could not answer; it becomes your test or your reason to delay.

Where this shows up at work

Teams use these words when they write a product brief, an internal policy, a procurement question for a supplier, or an answer to a customer who asks "is your AI fair?". A person who can say "we checked access for screen-reader users and keep chats for 30 days" is far more convincing than one who says "we take privacy very seriously". Principles that turn into specific, checkable statements are the ones that survive a hard question from a customer, a manager or a regulator.

Common mistakes

• Treating fairness as "same for everyone" every time. It can leave out people who face barriers. Fix: ask which differences are relevant and which are not.
• Mistaking a notice for transparency. A long legal text nobody reads is not transparency. Fix: write what a non-technical person needs to know in six plain lines and test it with one such person.
• Confusing accountability with blame. It makes people hide errors. Fix: name one accountable owner and make reporting mistakes safe.
• Collecting data "in case it is useful". It creates risk without a purpose. Fix: for each data item, write the reason; if there is none, do not collect it.
• Treating OECD or UNESCO text as law. They are voluntary recommendations. Fix: separate what is recommended from what a law in your country requires; the next page shows how.

Try it yourself

1. Easy. Match each situation to the principle it mainly concerns (dignity, fairness, harm, transparency, accountability, privacy, accessibility): (a) the assistant never says it is an AI; (b) a resident who uses a screen reader cannot press the "send" button; (c) a complaint about a wrong answer is sent to nobody in particular; (d) the assistant stores residents' ID numbers for two years with no stated reason.
2. Medium. Return to Mateo's applicant-screening tool from Decisions, Responsibility and Dilemmas in Everyday AI Use. Fill in the short principles check from this page for it. For each of the seven lines, write one sentence, or "not sure" and what you would do to find out.
3. Hard. Choose one principle from the OECD AI Principles and one from the UNESCO Recommendation that you think a school's AI tutoring tool should follow most strictly. Explain, in five to eight sentences, how each would change one design choice. Then state clearly that these are voluntary documents, and say where you would look to find out whether a binding rule applies to a school in your country.

Answers

Easy. (a) Transparency: people should be told AI is involved. (b) Accessibility, and also fairness and dignity if there is no alternative route. (c) Accountability: no owner receives complaints. (d) Privacy and data governance: data is collected and kept without a purpose. Pass: all four matched and one reason given for each; mentioning a second principle where it applies is a strength. Weak: matching by guesswork without a reason.

Medium. A strong answer might say: Dignity: applicants who feel rejected can reply and a person reads their email; Fairness: short, informal or second-language emails are checked by hand, not assumed weaker; Harm: strong applicants ranked low are a foreseeable risk, so Mateo reads the "Reject" group every day; Transparency: applicants are told a tool sorted the emails; Accountable: Mateo is named, and complaints go to him; Privacy: the tool does not see dates of birth or photos, and CVs are deleted after the process; Access: applicants who cannot attach a CV may reply in the message body. "Not sure" lines should have a way to find out, for example "test the tool with ten sample CVs". Pass: seven lines, each specific and checkable, and at least one honest "not sure". Weak: generic lines such as "we will be fair" with no check.

Hard. A strong answer picks, for example, "transparency and explainability" (both documents list it) and "do no harm" or "human oversight" (UNESCO), and shows concrete design changes: a plain notice that a pupil is using an AI tutor and that hints may be wrong; a teacher who can review a sample of chats and switch the tool off. It says plainly that the OECD and UNESCO documents are voluntary recommendations, and that to find binding rules it would check the education authority, data-protection regulator and national or regional law, using official sources, and ask the school for its own policy. Pass: two principles named correctly, two concrete design changes, an explicit statement that the documents are not binding, and a sensible route for finding out which rules apply. Weak: treating the documents as law, or listing principle names with no design change.

Summary

• Dignity, fairness, harm avoidance, transparency, accountability, privacy and accessibility are practical design questions you can check on a real use.
• Fair does not always mean identical: relevant differences may justify different treatment, and irrelevant ones should not hold anyone back.
• Transparency means plain words a non-technical person can use; accountability means a named owner who can fix things, not a person to blame.
• The OECD AI Principles (adopted 2019, updated May 2024) and UNESCO's Recommendation (November 2021) are voluntary documents that give shared vocabulary, not law.
• A ten-minute principles check turns values into specific questions, and an honest "not sure" becomes your next test.

Next: Frameworks, Standards and Competing Principles: Ethics, Law and Policy compares risk-management frameworks, standards and laws, and shows what to do when two good principles pull in opposite directions.