AI ethics and bias

In AI, bias means the system treats some people or things unfairly, often because it learned from examples that were not balanced. For example, if a face-recognition AI mostly saw light-skinned faces while learning, it may work poorly on darker-skinned faces. Bias usually comes from the data, not from the computer 'wanting' to be unfair.

AI learns patterns from the examples it is shown, so if the examples only cover one kind of situation, the AI will struggle with anything different. Using a wide variety of examples helps the AI work well for many people and cases, reducing unfair mistakes. Variety in training data is one of the best ways to make AI fairer.

Humans decide what data to collect, how to label it, and which examples to include or leave out. If those choices favor one group or reflect existing unfairness, the AI copies that unfairness. So bias can enter the AI through the everyday decisions people make while building it.

Fairness means an AI makes decisions without unfairly favoring or harming people because of things like their gender, skin color, or where they live. A fair AI gives people similar outcomes when their relevant situations are similar. Achieving fairness often requires carefully checking the AI's results across different groups.

The AI may learn to favor patterns common in men's resumes and unfairly rank women lower, even if they are equally qualified. This is an example of bias coming from unbalanced training data. The fix involves using more balanced data and testing the AI on different groups before using it.

Representative data includes examples from all the groups and situations the AI will actually face, in fair proportions. It matters because an AI only learns about the groups it sees in its data. If some groups are missing or rare, the AI is likely to make more mistakes for them.

Privacy is a person's right to control their own personal information, like their location, photos, or messages. AI builders should care because many AI systems learn from personal data, and misusing it can harm people or break their trust. Respecting privacy means collecting only what is needed and protecting it carefully.

People have a right to know and decide how their personal information is used. Collecting data without permission can lead to misuse, embarrassment, or harm, and it breaks trust. Ethical AI builders ask for clear consent and explain how the data will be used.

Transparency means people can understand how an AI works and why it made a certain decision. It is valuable because if we can see how an AI reaches its conclusions, we can spot mistakes, unfairness, or errors and fix them. Hidden, unexplained decisions are harder to trust or correct.

Accountability means there are people or organizations responsible for what an AI does and for fixing any harm it causes. Because AI itself cannot be blamed like a person, humans who build and use it must take responsibility. Accountability encourages careful testing and a way to correct mistakes.

This is an example of stereotype bias, where the AI repeats common social stereotypes it learned from biased data. Doctors and nurses can be any gender, so the AI's assumption is unfair and inaccurate. It shows how AI can absorb and spread unfair ideas present in its training examples.

A feedback loop happens when an AI's biased decisions create new biased data, which is then used to train the AI again, making the bias stronger. For example, if an AI sends police mostly to one neighborhood, more arrests there create data suggesting that area needs even more policing. Breaking these loops requires careful monitoring and correction.

Overall accuracy measures average performance, so an AI can score high by working well for the majority while failing for a small group. The small group's errors get hidden in the average. This is why we should test accuracy separately for each group, not just look at the total.

You can measure the AI's accuracy or error rate separately for each group, such as different age ranges or regions, and compare them. If one group has much worse results than others, that signals possible unfairness. This group-by-group testing helps reveal bias the overall score might hide.

# Compare accuracy per group
groups = {"A": [1,1,0,1], "B": [0,0,1,0]}  # 1=correct
for name, results in groups.items():
    acc = sum(results) / len(results)
    print(name, "accuracy:", acc)

Misinformation is false or misleading information presented as if it were true. AI can spread it when it confidently gives a wrong answer or generates fake-looking text, images, or videos. Because AI sounds convincing, people may believe and share its errors, so we should always check important AI claims.

A deepfake is a realistic but fake image, video, or audio made by AI, often showing someone saying or doing something they never did. It raises concerns because it can be used to trick people, spread lies, or harm someone's reputation. Knowing deepfakes exist reminds us to be careful about trusting media without checking the source.

AI can make mistakes or be biased, and high-stakes decisions can seriously affect people's lives, so a human should check the results. This 'human in the loop' approach catches errors the AI misses and adds responsibility and judgment. It keeps people, not machines, in charge of decisions that matter most.

It means that if you train an AI with poor-quality or biased data ('garbage in'), it will produce poor or biased results ('garbage out'). The AI cannot magically fix bad data; it learns whatever patterns are in it. So good, fair data is essential for a fair AI.

If a group is missing from the training data, the AI never learns about them and is likely to make more mistakes when it encounters them. For example, a voice assistant trained without certain accents may not understand people who have those accents. This 'exclusion' leads to unfair, worse service for the missing group.

Consent means a person agrees, after understanding what will happen, to let their information or photos be used. It matters because people have the right to decide how their images are used, and using them without permission can be unfair or harmful. Ethical AI projects get clear consent before using personal photos.

This shows gender bias, where the AI assumes a profession belongs to one gender based on patterns in its training data. Engineers can be any gender, so the assumption is inaccurate and unfair. It demonstrates how language AIs can carry hidden stereotypes from the text they learned from.

If an AI's decision cannot be explained, we cannot tell whether it was fair, correct, or based on the right reasons. This makes it hard to catch mistakes or unfair treatment and hard for affected people to challenge the decision. Explainability helps build trust and allows errors to be found and fixed.

Accuracy is about how often the AI gets the right answer overall, while fairness is about whether it treats different groups of people justly. An AI can be accurate on average but still unfair to a particular group. Good AI design aims for both accuracy and fairness, not just one.

A recommendation AI shows you more of what you already like, so over time you mostly see one type of content and miss other viewpoints. This narrow stream of information is called a filter bubble. It can limit what you learn and make it harder to see different ideas fairly.

Testing helps find mistakes, unfair results, and harmful behavior before they affect real people. Without testing, biased or dangerous problems could spread widely and cause harm. Responsible builders test on diverse cases and fix problems first, which is part of being accountable.

Data labeling is when people tag examples, like marking a photo as 'cat' or an email as 'spam', so the AI can learn from them. Bias can creep in if the labelers make unfair or inconsistent choices, or apply their own stereotypes. Since the AI learns from these labels, biased labels lead to a biased AI.

The AI might unfairly label students based on biased past data, such as their background or neighborhood, and these labels could affect their opportunities. There are also privacy risks in collecting student data. The school should check the AI for fairness, protect privacy, and keep human teachers in charge of important decisions.

Inclusive design means building AI so it works well and fairly for many kinds of people, including those who are often left out. It involves considering different abilities, languages, ages, and backgrounds from the start. Inclusive design helps prevent bias and makes technology useful for more people.

AI can present wrong answers in a confident, convincing way, since it does not truly 'know' when it is mistaken. Confidence in wording is not proof of correctness. That is why we should double-check important AI answers using reliable sources rather than trusting the tone alone.

People from different backgrounds notice different problems and blind spots, so a diverse team is more likely to spot unfairness others might miss. They can question assumptions and represent more groups affected by the AI. This variety of viewpoints helps build fairer and more inclusive systems.

Algorithmic bias is when the rules or learned patterns of an AI system produce systematically unfair results for certain people. It often comes from biased data or design choices rather than a single mistake. Recognizing it is the first step toward testing for and fixing unfair outcomes.

It is safer to have a human review final grades because the AI might be biased or make errors that unfairly hurt a student. Grades affect students' futures, so human judgment and accountability are important. The AI can help speed up scoring, but people should make or confirm the final decision.

If a face-recognition camera was trained mostly on certain faces, it may misidentify people from underrepresented groups more often. This is serious because a wrong match could lead to someone being falsely accused or treated unfairly. High-stakes systems like this need careful fairness testing and human checking.

You can collect more examples from the underrepresented group so it appears in fairer proportion to the others. Sometimes builders also carefully add or reweight examples so the AI pays enough attention to that group. The goal is for the data to better represent everyone the AI will serve.

People in different countries may speak differently, dress differently, or have different customs and environments, which the AI never learned about. With no examples from other places, the AI is likely to make more mistakes there. This shows why globally used AI needs data from many regions.

Using AI responsibly means being honest about when you used it, checking its answers for accuracy and fairness, and not using it to cheat or harm others. It also means protecting your own and others' privacy. Responsible use treats AI as a helpful tool while you stay the thoughtful decision-maker.

Biased AI used for jobs, loans, or school admissions can unfairly reject qualified people from certain groups, limiting their chances. Because these decisions shape careers and finances, the harm can be large and long-lasting. This is why fairness testing matters most for AI that affects opportunities.

A proxy is a piece of data that stands in for something else, like a postal code that secretly reflects people's race or income. Even if an AI is told not to use a sensitive trait directly, it can still be unfair by using a proxy linked to that trait. Spotting proxies is important to catch hidden bias.

Even without the gender label, the AI can find other clues, called proxies, that are linked to gender and still make biased decisions. So unfairness can sneak back in through related data. True fairness requires testing the AI's outcomes across groups, not just hiding one feature.

A good question is 'What data was this AI trained on, and was it tested for fairness across different groups?' This helps you judge whether the decision might be biased or unreliable. Asking about the data and testing encourages transparency and accountability.