What is AI ethics

Skewed data makes a skewed AI

AI learns from data. Show it many examples and it makes judgments by following the patterns in them. But if the examples lean to one side, the AI learns that lean as well. An AI that has seen lots of one kind of example will be poor at recognizing other kinds.
Skew does not come only from uneven numbers. If past human decisions contained prejudice, an AI that learns from those records will repeat it. An AI trained on old hiring records, for example, may end up passing over one group of applicants.

Press the circle or triangle button to add examples to one side only. When the data leans to one side, the boundary is pushed over and the smaller side keeps getting misjudged.

Adding examples to one side pushed the boundary over. The side the AI saw a lot it gets right; the side it saw little it keeps getting wrong. That is not because the AI is bad but because what it learned from was skewed. This is called bias.
So when building an AI, people need to check that the data was gathered evenly and that old prejudice is not hiding in it. After it is built, its accuracy is measured separately for different groups to see whether it treats them alike.

Confidently making things up

We learned that an LLM continues text by picking the next word by probability. Its aim is to continue plausibly, so it does not separately check whether what it says is true. Ask it something it does not know, and instead of saying it does not know, it may produce a plausible-sounding sentence.
The tone is confident, but the content is wrong. This is called hallucination.

Tap the questions the AI can answer and the ones it cannot. Even for things no one can know, like tomorrow's lottery numbers, it does not stop but makes up a plausible answer. That is hallucination.

For questions it knew, the right answer came out, but for questions no one could answer it made up an answer in exactly the same confident tone. On top of that, an LLM only knows material up to when it was trained, so it is shaky on anything that happened afterward.
So an AI's answers always need checking. Numbers, sources and recent events in particular are safest to look up again in trustworthy sources. Some systems now show search results alongside their answers to make that checking easier.

Same technology, for good or for harm

AI can make pictures, imitate voices and write fluently. These abilities are very useful. They can restore a blurry old photo, give people who lost their voice to illness their own voice back, and explain hard texts in plain words.
But the very same abilities can make fake faces or voices that look and sound real and be used to deceive people. Fakes like these are called deepfakes. One tool can be put to opposite uses.

Pick a technology and look at its helpful use and its harmful use side by side. You can see that how a tool is used decides whether it does good or harm.

The same technology helped on one side and harmed on the other. The technology itself is neither good nor bad; what decides is the choice of the person using it. That is why the more powerful a tool, the more we need agreements on how it is used.
Labeling AI-made content as AI-made, not using someone's face or voice without permission, and not helping yourself to other people's work or personal data are agreements of that kind. A number of countries have started writing such rules into law.

People check, transparency, fairness

So what does using AI responsibly involve? Three things are worth remembering. First, people check: do not take an AI's answer as given, and have a person review important decisions.
Second, be transparent: do not hide what data was used, how the system was built or where AI was used. Third, watch for fairness: make sure no particular group of people is the one that loses out.

Tap the three checks in order to pass them. Only when the human check, transparency and fairness are all on is the AI in a state you can use with confidence.

With all three on, the AI was finally safe to use. Looking back, everything we learned comes together here. AI learns from data, stacks neurons and layers to catch patterns, and an LLM continues text by probability.
Its limits are as clear as its power. So the last square always belongs to people: checking, disclosing and using it fairly.

Wrapping up

In short, AI is powerful but not perfect. Skewed data skews its judgments (bias), it can confidently make up what it does not know (hallucination), and the same technology can be used to harm people (misuse).
So the responsibility to check, to be open and to be fair must always come along. The smarter AI becomes, the more it matters to use that ability responsibly.

Tap the AI journey in order to wrap it up. (learns from data → patterns from neurons and layers → an LLM speaks by probability → limits are bias, hallucination and misuse → use it responsibly)

We have come a long way. Starting from data, we looked at a single neuron, networks stacked in layers, CNNs that see images, embeddings that give words coordinates, attention that reads context, LLMs that continue text by probability, and generative AI that makes something new.
At the end we gained one more thing: not only how to use AI cleverly but how to use it responsibly. Weighing what an AI says and deciding where and how to use it is a judgment that stays with people to the end.