English:Artificial Intelligence, Bias, and Ethics
Artificial Intelligence, Bias, and Ethics
Introduction
Artificial intelligence (AI) is already part of everyday life. It can recommend videos, translate text, detect patterns in medical images, support hiring, generate writing and images, and help people search large amounts of information. These systems can be useful, but they can also make mistakes, repeat unfair patterns, or affect people in ways that are hard to see.
In this aiMOOC, you will learn how artificial intelligence works at a basic level, how algorithmic bias can arise, and how AI ethics helps people ask responsible questions. The course is designed for Grades 9–10. You do not need programming experience. You will practice reading evidence, questioning automated decisions, discussing different viewpoints, and communicating your conclusions clearly in English.
Guiding question: How can we use AI in ways that are useful, fair, transparent, safe, and accountable?
Learning Goals
By the end of the course, you should be able to explain the difference between AI, machine learning, data, and an algorithm; identify several sources of bias in an AI system; distinguish ordinary model error from unfair patterns of error; explain why fairness depends on context; apply ethical principles such as transparency, privacy, accountability, safety, and human oversight; evaluate evidence about an AI system; and propose practical ways to reduce harm.
How Artificial Intelligence Learns
AI is a broad term for computer systems that perform tasks associated with abilities such as recognizing patterns, generating language, planning, or making predictions. Many modern AI systems use machine learning. Instead of being given a separate rule for every possible situation, a machine-learning model learns patterns from examples or other data.
A simple supervised learning process often includes these steps: people define a task, collect and label data, split the data into training and testing sets, train a model, test the model, and then decide whether it is good enough for the intended use. The final decision is not automatically trustworthy just because a model has a high overall accuracy score.
Data, Features, Labels, and Models
Data are the examples or observations used by a system. A feature is information the system uses as input. A label is the target answer in supervised learning. A model is the learned mathematical structure that maps inputs to outputs.
Imagine a school creates a model to predict which students may need extra tutoring. Features might include attendance, assignment completion, or previous assessment results. The label might be whether a student later needed additional support. Ethical questions begin immediately: Are the records accurate? Do the features measure what the school really cares about? Could a feature act as a proxy for income, disability, language background, or access to technology? Who can challenge an incorrect prediction?
Predictions Are Not Facts
A model output is a prediction, score, ranking, classification, or generated response. It is not automatically a fact. Models can be wrong because their data are incomplete, the world changes, the task is poorly defined, or the model is used outside the setting for which it was tested.
For a low-stakes task, such as suggesting a song, an error may be annoying. For a high-stakes task, such as screening job applicants or supporting a medical decision, an error can have serious consequences. Ethical evaluation therefore depends on context, impact, and who carries the risk.
What Is Bias?
The word bias can mean different things. In statistics and machine learning, bias can describe a systematic modeling error. In discussions of AI ethics, algorithmic bias usually refers to systematic patterns that create unfair or unequal outcomes for people or groups. These meanings can overlap, but they are not identical.
A system can be biased even if no developer intended to discriminate. Bias can enter through choices made across the entire AI lifecycle: what problem is selected, who is represented in the data, how categories are defined, what outcome is optimized, how the model is tested, and how people use its outputs.
Common Sources of Algorithmic Bias
Historical bias appears when data reflect unfair patterns already present in society. If past decisions were discriminatory, a model trained to imitate them may learn those patterns.
Representation bias appears when some groups or situations are underrepresented in the data. A model may perform well for common cases but poorly for groups it rarely encountered.
Measurement bias appears when a chosen measurement does not capture the concept that really matters. For example, using the number of online messages as a measure of student engagement may overlook students who participate in other ways.
Label bias can occur when the target labels are subjective, inconsistent, or shaped by earlier human judgments.
Proxy bias occurs when a seemingly neutral feature closely tracks a sensitive characteristic. A postal code, for example, can sometimes act as a proxy for socioeconomic conditions.
Objective bias can appear when a system optimizes the wrong goal. Maximizing clicks may reward sensational content even if the deeper goal is reliable information.
Feedback loops can amplify earlier decisions. If a system sends more attention to places it already predicts are risky, it may collect more data from those places and then treat that extra data as evidence that its first prediction was correct.
Bias Is Not Only a Data Problem
Improving a dataset can help, but ethical AI requires more than balanced data. The chosen task, model, threshold, user interface, institutional rules, and human decision process can all create harm. A technically accurate model can still be unfair if it is used for an inappropriate purpose or if people have no meaningful way to appeal its decisions.
The graph above comes from research on ORES, a machine-learning system used to help classify Wikipedia edits. It shows why subgroup performance matters: an overall score can hide different error rates for different groups. When you inspect an AI system, ask not only “How accurate is it?” but also “Accurate for whom, under what conditions, and with what consequences?”
Case Study: Facial Analysis
Facial analysis is a useful case for studying bias because performance can differ across demographic groups and because errors can affect privacy, identity, access, and surveillance.
In the 2018 study Gender Shades, Joy Buolamwini and Timnit Gebru evaluated three commercial gender-classification systems and found large differences in error rates across groups defined by skin type and gender presentation. In that historical evaluation, darker-skinned women were the most misclassified group, with error rates reported as high as 34.7%, while the maximum error rate for lighter-skinned men was 0.8%. These figures describe the systems tested at that time; they should not be treated as performance claims about every current facial-analysis system. The study is important because it demonstrated how an average score can hide serious subgroup disparities. You can read the original research at Gender Shades.
Questions to Ask About a Case Study
When you evaluate an AI case, identify the stakeholders: the people who build, buy, operate, regulate, or are affected by the system. Then ask what benefits are expected, what harms are possible, who receives the benefits, who carries the risks, what evidence supports the claims, and what happens when the system is wrong.
A good ethical analysis does not stop at saying that technology is “good” or “bad.” It compares alternatives. Sometimes the right question is not “How can we make this AI more accurate?” but “Should AI be used for this decision at all?”
Fairness: More Than One Definition
Fairness sounds simple, but people can reasonably disagree about what it requires. One approach is to compare error rates across groups. Another is to ask whether people with similar relevant circumstances receive similar treatment. A third is to focus on whether a system reduces or increases existing inequality.
These goals can conflict. For example, two groups may have different underlying rates of an outcome in a dataset. A system may not be able to equalize every statistical measure of fairness at the same time while also preserving the same calibration. This is why fairness cannot be reduced to a single number. The choice of a fairness measure is partly a technical decision and partly an ethical and social decision.
Fairness Questions for Learners
When you see a claim that an AI system is “fair,” ask: Which groups were tested? Which metric was used? What types of errors matter most? Were affected people involved in the design? Is there a process for appeal or correction? Does the system improve on the realistic alternative, including human decision-making?
Core Principles of Responsible AI
International organizations and technical standards use different wording, but several ideas appear repeatedly. UNESCO’s Recommendation on the Ethics of Artificial Intelligence emphasizes human rights and dignity, fairness and non-discrimination, transparency and explainability, privacy, accountability, safety, sustainability, and human oversight. The U.S. National Institute of Standards and Technology, through the AI Risk Management Framework, also describes trustworthy AI in terms that include validity, reliability, safety, security, accountability, transparency, explainability, privacy, and managing harmful bias.
Fairness and Non-Discrimination
A responsible system should be tested for harmful differences in performance or impact. Developers may need more representative data, better measurements, subgroup testing, accessibility work, or changes to the decision process. Fairness also requires listening to people who may be affected.
Transparency and Explainability
Transparency means providing meaningful information about what a system is, what it is designed to do, what data or processes are relevant, and what its limits are. Explainability means helping people understand why a particular output or decision was produced, at a level appropriate to the context.
An explanation that is useful to a software engineer may not be useful to a student, patient, or job applicant. Good explanations are understandable to the people who need them.
Accountability and Human Oversight
Accountability means that people and organizations remain responsible for how AI is designed, deployed, monitored, and corrected. “The computer decided” is not a sufficient answer when a system affects someone’s rights or opportunities.
Human oversight is most useful when the human has enough time, information, authority, and training to notice problems and act. A person who only clicks “approve” without meaningful review is not strong oversight.
Privacy and Data Protection
AI systems can use large amounts of personal information. Ethical questions include whether data were collected lawfully and fairly, whether people understand the use, whether only necessary data are collected, how long data are kept, who can access them, and whether sensitive information can be inferred from other data.
Safety, Security, and Reliability
A system should be tested for expected use, foreseeable misuse, failures, and changing conditions. Security matters because an unsafe or manipulated model can cause harm. Reliability also requires monitoring after deployment, not just before release.
Sustainability
AI has environmental costs because training and operating digital systems require hardware, electricity, cooling, and material resources. Ethical evaluation can therefore include whether the expected benefit justifies the resource use and whether more efficient alternatives are available.
A Practical AI Ethics Checklist
You can analyze an AI system with the following questions.
- Purpose: What problem is the system supposed to solve, and is AI an appropriate tool for that problem?
- Data: Where did the data come from, who is represented, and what important cases may be missing?
- Measurement: Do the features and labels measure the concepts that actually matter?
- Fairness: Which groups could experience different error rates, access, or outcomes?
- Transparency: Do users know AI is involved, and can affected people understand important decisions?
- Privacy: Is personal information necessary, protected, and used in a way people can reasonably expect?
- Accountability: Who is responsible for monitoring, correcting, and responding to harm?
- Human oversight: Can a person meaningfully review, override, or appeal an important decision?
- Safety: What happens if the model is wrong, manipulated, or used outside its intended context?
- Sustainability: What resources are required, and are there less resource-intensive ways to reach the goal?
Evaluating Claims About AI
AI debates often include dramatic claims. Strong evaluation requires evidence. Look for the source of a claim, the date, the population studied, the definition of success, the comparison group, the size of the test, and whether the results were independently checked.
Be careful with a single accuracy number. Suppose a classifier is 95% accurate overall. That number does not tell you whether one group has much higher false-positive or false-negative rates. It also does not tell you how serious each error is. In a medical screening system, a false negative may have different consequences from a false positive. In another setting, the priorities may differ.
Correlation, Causation, and Proxies
AI systems are excellent at finding correlations. A correlation means two things vary together; it does not automatically show that one causes the other. A model can make useful predictions using a proxy variable without understanding the real reason behind the pattern. This can create ethical problems if the proxy connects to sensitive characteristics or if decision-makers treat prediction as explanation.
Generative AI and Bias
Generative AI can produce text, images, audio, or code. Its outputs can reflect patterns, stereotypes, gaps, and errors in training data and feedback. Because generative systems can produce fluent answers, users may overestimate their reliability. Responsible use includes checking important claims, protecting private information, identifying generated content when appropriate, and avoiding the use of AI output as the sole basis for high-stakes decisions about people.
Generative AI also raises questions about authorship, copyright, labor, misinformation, environmental impact, and whose languages and cultures are represented. Ethical judgment therefore involves both the output and the larger system around it.
How Bias Can Be Reduced
There is no single fix for algorithmic bias, but several practices can reduce risk. Teams can define the purpose carefully, include diverse stakeholders, document data sources, improve representation, inspect labels, test subgroup performance, compare several fairness measures, use privacy and security protections, conduct impact assessments, provide meaningful explanations, enable appeals, monitor systems after deployment, and stop or redesign systems when harms cannot be reduced enough.
The goal is not to claim that an AI system is perfectly unbiased. A more responsible goal is to identify risks, measure them where possible, make value choices visible, involve affected people, and keep improving or withdrawing the system when evidence shows unacceptable harm.
Evidence and Standards
The following sources can help you check claims and explore the topic further.
- UNESCO Recommendation on the Ethics of Artificial Intelligence: A global framework centered on human rights, dignity, fairness, transparency, accountability, privacy, oversight, and sustainability.
- NIST AI Risk Management Framework: A practical framework for identifying and managing AI risks and characteristics of trustworthy AI.
- OECD AI Principles: International principles covering human-centered values, fairness, transparency, robustness, safety, and accountability.
- Gender Shades: A peer-reviewed 2018 study showing why intersectional subgroup testing matters in facial analysis.
- Ethics of artificial intelligence: An overview of major ethical questions in AI.
- Algorithmic bias: Background on how systematic unfair outcomes can arise in algorithmic systems.
Interactive Tasks
Quiz: Test Your Knowledge
What best describes algorithmic bias in an ethics discussion? (A systematic pattern that can create unfair outcomes) (!Any random mistake made by a computer) (!A computer running more slowly than expected) (!A preference for one programming language)
Why can high overall accuracy be misleading? (It can hide very different error rates across groups) (!It proves the model is always fair) (!It means the training data were perfectly balanced) (!It removes the need for human review)
What is representation bias? (Some groups or situations are underrepresented in the data) (!The model uses too much electricity) (!The user interface has too many buttons) (!The model gives every group identical results)
Which question is mainly about transparency? (Do people know AI is involved and understand its role) (!How fast can the computer processor run) (!How many colors appear on the screen) (!How long is the source code)
What is a proxy variable? (A feature that can indirectly track another characteristic) (!A guaranteed explanation of cause) (!A type of computer virus) (!A replacement for testing)
Why is human oversight important in high-stakes AI? (People must be able to review and act on important problems) (!Humans always make perfect decisions) (!AI cannot produce any useful prediction) (!Oversight automatically removes all bias)
What is the best response to a claim that an AI system is fair? (Ask which groups and fairness measures were evaluated) (!Accept the claim if the model is new) (!Check only the overall accuracy) (!Assume fairness means equal predictions for everyone)
What can create a feedback loop? (A system collects more data where it already focuses attention) (!A model is tested once before deployment) (!A dataset is stored in two folders) (!A user closes an application)
Which action best supports accountability? (Naming who must monitor and correct the system) (!Allowing no one to challenge a decision) (!Hiding the model purpose from users) (!Treating every output as a fact)
What is a responsible way to use generative AI for school research? (Check important claims against reliable sources) (!Assume fluent writing is always correct) (!Enter private data whenever requested) (!Use generated text as the only evidence)
Memory Game
| Algorithmic bias | Systematic patterns that can create unfair outcomes |
| Transparency | Meaningful information about how an AI system is used and limited |
| Accountability | Responsibility for decisions, monitoring, and correction |
| Proxy | A feature that indirectly tracks another characteristic |
| Oversight | Human ability to review and intervene in an AI-supported process |
| Audit | A structured examination of evidence about a system and its impacts |
Drag and Drop
| Match the correct terms. | Topic |
|---|---|
| Representation bias | Missing or underrepresented groups in training data |
| Measurement bias | A chosen measure does not capture the real concept well |
| Feedback loop | Earlier outputs influence later data and reinforce a pattern |
| Explainability | Helping people understand why an output was produced |
| Privacy | Protecting personal information and limiting unnecessary data use |
...
Crossword Puzzle
| Fairness | What principle asks whether people and groups are treated justly? |
| Privacy | What principle protects personal information? |
| Dataset | What collection of examples can be used to train or test a model? |
| Proxy | What feature can indirectly track another characteristic? |
| Audit | What structured review examines a system and its evidence? |
| Oversight | What term describes meaningful human review and intervention? |
LearningApps
Cloze Text
Open-Ended Tasks
Easy
- AI in everyday life: Find three examples of AI you encounter in one day. For each example, write what the system probably uses as input, what it produces, and one question you would ask about fairness or privacy.
- Bias vocabulary poster: Create a one-page visual that explains representation bias, measurement bias, proxy bias, and feedback loops in your own words with a simple example for each term.
- Claim checking: Choose one public claim about AI and check it against two reliable sources. Write a short paragraph explaining what the evidence supports and what remains uncertain.
- Ethical question interview: Interview a classmate or family member about one AI-supported decision they would trust and one they would not trust. Summarize the reasons without including private information.
Standard
- School AI policy: Draft a short class policy for responsible generative AI use that covers learning goals, fact-checking, privacy, attribution, and situations in which AI should not replace a student’s own work.
- Fairness audit: Design a fictional test for an AI classifier. Decide which groups you would compare, which errors you would count, and what evidence would make you pause deployment.
- AI ethics infographic: Produce an infographic linking fairness, transparency, accountability, privacy, safety, and human oversight to one real-world AI application.
- Structured debate: Prepare and record a three-minute argument for or against using AI to rank applicants for a fictional school opportunity. Include one benefit, two risks, one counterargument, and one safeguard.
Advanced
- Dataset investigation: Build a small, non-sensitive dataset of objects, words, or public information and examine whether some categories are underrepresented. Explain how the imbalance could affect a model trained on it.
- Algorithmic impact assessment: Write a two-page impact assessment for a fictional AI system used in education, hiring, healthcare, or public services. Identify stakeholders, possible harms, evidence needed, safeguards, and an appeal process.
- Bias experiment: Using a teacher-approved public AI tool, design a safe prompt-based experiment that tests whether outputs change across equivalent scenarios. Keep variables controlled, record results, and discuss limitations without making claims about protected groups from a tiny sample.
- Ethics documentary: Create a five-minute video or audio report that combines an interview, evidence from at least three reliable sources, a case study, and your own recommendation for responsible AI use.
Learning Assessment
- Case analysis: Analyze a fictional AI system that predicts student support needs. Identify at least three possible sources of bias, explain who could be harmed, and propose safeguards that address both technical and human decision processes.
- Fairness reasoning: Compare two possible fairness goals for a high-stakes classifier and explain why choosing between them requires ethical judgment as well as mathematics.
- Evidence evaluation: Given an overall accuracy score and subgroup error rates, decide whether the overall score is sufficient evidence for deployment and justify your answer using the consequences of false positives and false negatives.
- Responsible design: Redesign an AI-supported decision process so that transparency, privacy, accountability, and meaningful human oversight are built into the workflow.
- Transfer task: Choose a new setting not discussed in detail in the course, such as sports, transportation, environmental monitoring, or online moderation, and apply the AI ethics checklist to recommend whether and how AI should be used.
- Source critique: Compare a news report, a company claim, and a research or standards source about the same AI topic. Judge which claims are strongest and explain how date, evidence, incentives, and missing context affect your confidence.
Evidence of Learning
Strong evidence of learning includes accurate knowledge of AI, machine learning, data, models, bias, and ethical principles; the ability to distinguish different sources of bias; careful interpretation of overall and subgroup performance; clear use of evidence; recognition of uncertainty and limitations; reasoned comparison of benefits and harms; responsible handling of privacy; and the ability to explain why fairness depends on context.
Useful products can include an audit plan, infographic, interview summary, policy, impact assessment, debate, dataset investigation, or short documentary. Strong transfer is shown when you can apply the same questions to a new AI system, identify stakeholders and risks, compare realistic alternatives, and recommend safeguards that match the seriousness of the decision.
OERs on the Topic
You can also explore Algorithmic bias, Fairness in machine learning, Explainable artificial intelligence, AI safety, Privacy, Machine learning, and Digital literacy.
Linked Learning Areas
This topic connects computer science with English-language discussion, evidence-based writing, media literacy, civics, philosophy, mathematics, and digital citizenship. It is especially useful when you practice explaining technical ideas clearly, comparing sources, building arguments, and discussing how technology affects people.
aiMOOC Projects
MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.