English:Artificial Intelligence in Society
Artificial Intelligence in Society
Introduction
Artificial intelligence is already part of social life. Search engines rank information, recommendation systems shape what people watch, navigation tools estimate routes, banks detect unusual transactions, hospitals use decision-support systems, employers use software to organize applications, and generative AI can produce text, images, audio, code, and video. For learners in Grades 11–13, understanding AI therefore means more than learning how algorithms work. You also need to ask who designs these systems, what data they use, whose interests they serve, how their errors are distributed, and who is accountable when they affect people.
In this aiMOOC, you will study artificial intelligence as a technical and social system. You will connect machine learning with data ethics, privacy, media literacy, human rights, economics, environmental questions, and democratic decision-making. The goal is not to label AI as simply good or bad. Instead, you will learn to evaluate specific systems in context, compare benefits with risks, and make evidence-based judgments.
A machine-learning workflow typically includes data preparation, modeling, and evaluation. Social choices enter at every stage.
By the end of the course, you should be able to explain core AI concepts, identify important societal effects, analyze cases of bias and misinformation, evaluate claims about AI, and propose responsible ways to design or govern AI systems.
How Artificial Intelligence Works
AI, Machine Learning, and Models
Artificial intelligence is a broad field concerned with computer systems that perform tasks associated with abilities such as perception, language processing, prediction, planning, or problem-solving. Most AI systems used today are specialized rather than generally intelligent. They operate within a defined task or set of tasks.
Machine learning is one approach to AI. Instead of programming every decision rule directly, developers use data and optimization methods to fit a model. A model is a learned mathematical structure that maps inputs to outputs. During training, the system adjusts parameters to improve performance on an objective. During inference, the trained model processes new inputs to produce predictions, classifications, recommendations, or generated content.
A simplified artificial neural network. Real systems may contain many more layers, parameters, and processing steps.
A model's behavior depends on more than its architecture. Training data, labels, objective functions, evaluation methods, computing resources, deployment conditions, and user behavior all matter. This is why technical performance in a laboratory does not automatically guarantee good social outcomes.
Generative AI
Generative AI creates new outputs such as text, images, music, video, or code. Many modern generative systems learn statistical relationships from very large datasets. A language model, for example, predicts likely sequences of tokens from patterns learned during training. It does not automatically verify every statement against reliable evidence.
The diagram contrasts a discriminative task, such as classification, with a generative task that creates new content.
Generative AI can support brainstorming, translation, accessibility, coding, tutoring, and creative experimentation. It can also produce fabricated information, reproduce stereotypes, imitate styles, expose confidential data if users submit it carelessly, or create convincing synthetic media. A fluent answer is not the same as a correct answer. You should therefore verify important claims with independent sources and understand the rules that apply in your school, workplace, or institution.
AI Is a Socio-Technical System
An AI system is never only a model. It is embedded in an organization and a social setting. A useful analysis asks about the whole chain: Who defines the problem? Who selects the data? What objective is optimized? Who deploys the system? Who is affected? Who can challenge a decision? What happens when the system fails?
This perspective is called socio-technical because technical components and social arrangements interact. A technically accurate system may still be harmful if it is used for an unsuitable purpose, if its errors fall disproportionately on a particular group, or if people have no meaningful way to appeal. Conversely, careful design, testing, documentation, human oversight, and institutional safeguards can reduce some risks.
AI in Everyday Life and Public Institutions
Recommendation, Ranking, and Personalization
Recommendation systems select or rank content based on signals such as previous interactions, similarities between users, or predicted relevance. They can make large information spaces easier to navigate. At the same time, the ranking objective matters. A system optimized for watch time, clicks, or purchases may encourage different behavior from one optimized for diversity, quality, or public value.
Personalization also creates feedback loops. What you click influences what you are shown next, and what you are shown influences what you click. This does not mean algorithms control every decision, but it means platforms help structure attention. When analyzing a recommender system, distinguish the user's preferences from the platform's design choices and business incentives.
Education
AI tools can provide explanations, language support, adaptive exercises, feedback, and accessibility features. They may help teachers generate examples or help learners compare alternative explanations. However, educational use also raises questions about accuracy, assessment, privacy, authorship, unequal access, and over-reliance.
A productive use of AI in learning should preserve your own reasoning. You can ask a system for counterarguments, examples, or formative feedback, then verify the result and explain the final reasoning yourself. In high-stakes assessment, follow the institution's rules and disclose AI assistance when required.
Healthcare and Public Services
In healthcare, AI may support image analysis, risk estimation, workflow organization, or clinical decision support. In public services, algorithms may help allocate resources, detect patterns, or prioritize cases. These uses can be valuable, but they are high-stakes because errors can affect health, rights, access to services, or legal opportunities.
For high-stakes systems, performance should be evaluated on relevant populations and realistic conditions. It is also important to know whether a tool supports a professional judgment or effectively determines an outcome. Meaningful human oversight requires time, authority, competence, and the ability to disagree with the system rather than simply approving its recommendation.
Bias, Fairness, and Discrimination
Where Bias Can Enter
Bias can enter an AI system through historical data, unrepresentative samples, measurement choices, labels, proxy variables, objective functions, or deployment practices. If historical data reflect unequal treatment, a model can learn patterns associated with that inequality. If one group is underrepresented in a dataset, the system may perform worse for that group.
Facial recognition illustrates why performance, consent, privacy, context, and unequal error rates all matter in social evaluation.
Fairness is not one simple mathematical property. Different fairness criteria can express different moral and policy goals, and they can conflict. For example, equalizing one type of error rate may not equalize another. Therefore, a fair decision process requires both quantitative evaluation and public reasoning about which harms matter in the specific context.
From Bias to Harm
Not every statistical difference is discrimination, and not every biased system produces the same level of harm. Context matters. A small error in a music recommendation is different from a small error in a medical diagnosis, fraud investigation, or hiring process.
When you evaluate an AI system, ask which groups are affected, how serious the consequences are, whether people gave meaningful consent, whether an alternative process exists, and whether there is a way to correct mistakes. Fairness analysis becomes stronger when it combines technical evidence with legal, ethical, and social perspectives.
Privacy, Surveillance, and Data Power
AI often depends on large quantities of data. Personal data can include obvious identifiers such as names and addresses, but also location traces, browsing patterns, biometric measurements, voice recordings, or inferred interests. Even data that appear anonymous may sometimes become identifiable when combined with other information.
Privacy is not only about secrecy. It also concerns control, context, autonomy, and the ability to participate in society without constant observation. A system may be technically secure yet still collect more data than necessary. Data minimization means collecting only what is needed for a legitimate purpose and retaining it no longer than necessary.
AI-enabled surveillance can be used for safety, access control, or investigation, but it can also create chilling effects, unequal monitoring, or function creep, where data collected for one purpose are later used for another. Good governance therefore asks not only whether surveillance works, but whether it is necessary, proportionate, accountable, and open to review.
Synthetic Media, Deepfakes, and Information Integrity
Generative AI can produce realistic text, images, voices, and videos. Synthetic media can be used for art, film, education, satire, accessibility, and simulation. It can also be used for impersonation, fraud, harassment, or disinformation.
A deepfake is synthetic or manipulated media that convincingly imitates a person's appearance or voice. The social problem is broader than detecting fakes. Once people know convincing fakes exist, genuine evidence can also be dismissed as fake. This makes source verification and information provenance increasingly important.
When you encounter surprising media, use lateral reading. Check who published it, look for independent reporting, trace the earliest available source, inspect context, and compare the claim with reliable evidence. Do not assume that an image is true because it looks realistic, but do not assume it is false merely because AI tools exist.
Useful provenance systems can record information about where media came from and how it was edited, but no single technical label solves the problem. Media literacy remains essential.
Work, Economy, and Inequality
AI can automate some tasks, augment human work, create new tasks, and change the skills demanded within occupations. The effect on employment depends on technology, business decisions, labor markets, regulation, education, and how productivity gains are distributed. It is therefore misleading to treat one prediction as a certain forecast of how many jobs will disappear.
This map shows the share of job postings associated with AI in selected European countries in 2024. It does not measure total employment, job quality, or future job losses.
AI may raise productivity while also changing bargaining power between workers and employers. Monitoring software can increase coordination, but it can also intensify surveillance. Automated screening may reduce routine work, but it can also hide criteria from applicants. When evaluating workplace AI, consider transparency, worker participation, opportunities for appeal, and who receives the benefits of increased productivity.
Access to AI skills and infrastructure is also uneven. Schools, companies, regions, and countries differ in computing resources, connectivity, data availability, and expertise. These differences can widen existing inequalities unless access, education, and public policy are addressed.
Creativity, Culture, and Copyright
Generative AI changes how people create and remix text, images, music, video, and software. It can lower barriers to experimentation and help people prototype ideas quickly. At the same time, creators have raised questions about training data, attribution, consent, compensation, imitation, and market effects.
Copyright law differs across jurisdictions and continues to develop in relation to generative AI. Ethical questions may be broader than legal ones. Even when an action is legally permitted, you can still ask whether it respects creators, audiences, and cultural communities.
AI-generated culture can also influence whose languages, styles, and histories are visible. Models trained mostly on dominant languages or widely digitized material may provide weaker support for underrepresented communities. Cultural diversity therefore depends partly on who contributes data, who controls infrastructure, and which communities can shape system design.
AI, Democracy, and Public Debate
AI can support translation, accessibility, public administration, and analysis of large information collections. It can also be used to generate persuasive messages at scale, imitate public figures, micro-target audiences, or flood information spaces with low-quality content.
The democratic challenge is not simply to ban automation. It is to preserve conditions for informed public judgment: trustworthy institutions, independent journalism, transparent political communication, access to evidence, and the ability to hold decision-makers accountable.
When discussing AI and democracy, distinguish between evidence and speculation. Ask who benefits from a claim, what data support it, whether the evidence is representative, and whether alternative explanations have been considered. The same critical method applies both to enthusiastic marketing and to alarmist predictions.
Environmental Impact
AI runs on physical infrastructure: chips, servers, data centers, networks, cooling systems, and electricity grids. Training and operating models require energy, while manufacturing hardware also requires materials and industrial processes. The environmental impact of an AI service depends on factors such as model size, hardware efficiency, data-center efficiency, utilization, cooling, location, and the electricity mix.
AI services depend on physical computing infrastructure, even when the user experiences them as cloud-based software.
It is difficult to assign a single universal energy figure to "AI" because systems differ greatly. Responsible analysis should therefore avoid treating one number as representative of every model or task. At the same time, energy demand from data centers is a real infrastructure issue. AI can also help optimize energy systems, scientific research, and industrial processes, so environmental assessment should compare both costs and potential benefits.
Ethics and Responsible AI
Core Ethical Questions
Ethics asks what ought to be done, not only what can be done. Important principles for AI include human dignity, autonomy, fairness, privacy, safety, transparency, accountability, inclusiveness, and environmental sustainability. These principles can conflict. For example, more transparency may expose sensitive data, or a highly accurate system may still be unacceptable if it violates rights.
A strong ethical analysis therefore identifies stakeholders, benefits, harms, uncertainty, power differences, and possible alternatives. It also asks who has a voice in the decision. People affected by an AI system should not be treated merely as sources of data.
A review of 200 AI ethics guidelines found recurring governance principles. The chart shows frequency of principles, not proof that the principles are implemented.
Human Oversight and Accountability
Human oversight is often proposed as a safeguard, but it must be meaningful. A person who lacks time, information, authority, or training may simply accept an automated recommendation. This is known as automation bias when people give too much weight to system output.
Accountability requires clearer responsibility. Organizations should know who is responsible for data quality, testing, deployment, monitoring, complaints, and corrective action. Documentation can support accountability by recording intended uses, limitations, evaluation results, and known risks.
People affected by important automated decisions should, where appropriate, receive understandable information and a route to challenge harmful or incorrect outcomes. Appeals and remedies are not technical features alone; they are institutional protections.
Governance and Regulation
AI governance includes laws, standards, organizational policies, audits, impact assessments, procurement rules, professional norms, and technical safeguards. Different frameworks operate at different levels.
UNESCO's Recommendation on the Ethics of Artificial Intelligence, adopted in 2021, emphasizes human rights, human dignity, fairness, transparency, human oversight, inclusion, and environmental sustainability. The OECD AI Principles promote innovative and trustworthy AI that respects human rights and democratic values. The U.S. National Institute of Standards and Technology provides the voluntary AI Risk Management Framework to help organizations manage risks. Its core functions are Govern, Map, Measure, and Manage.
In the European Union, the AI Act entered into force in 2024 and applies requirements in phases according to risk and system type. This risk-based approach illustrates an important governance idea: the safeguards expected for an AI system should depend partly on the seriousness of the potential consequences.
No framework removes the need for judgment. Rules can become outdated, organizations can comply formally without changing behavior, and new uses can create unexpected risks. Governance therefore needs monitoring, independent scrutiny, public participation, and the ability to revise decisions.
A Practical Framework for Evaluating an AI System
You can analyze a real AI system by asking the following questions:
- Purpose: What problem is the system supposed to solve, and is AI necessary for that purpose?
- Stakeholder analysis: Who benefits, who bears risks, and who has decision-making power?
- Training data: What data were used, who is represented, and what important groups or situations may be missing?
- Model evaluation: Which metrics are used, and do they reflect the harms that matter in the real setting?
- Privacy: What personal data are collected, inferred, stored, or shared?
- Transparency: What information can users and affected people reasonably understand?
- Human oversight: Can people meaningfully review, override, and challenge the system?
- Accountability: Who is responsible when the system fails or causes harm?
- Security: How could the system be manipulated, attacked, or misused?
- Environmental sustainability: What computing resources are required, and are lower-impact alternatives available?
- Alternatives: Could a non-AI process solve the problem with fewer risks?
- Monitoring: How will the organization detect new problems after deployment?
A good evaluation does not end with a single score. It explains trade-offs and uncertainty, compares alternatives, and recommends concrete safeguards.
Media Literacy for AI Claims
AI is surrounded by both hype and fear. To evaluate a claim responsibly, separate demonstrations from evidence. A polished demo may show what a system can do under selected conditions, but it does not reveal average performance, failure rates, cost, robustness, or social consequences.
Look for the denominator behind percentages, the population used in a study, the date of the data, and whether the comparison is fair. Ask whether the source has commercial or political incentives. Distinguish peer-reviewed research, official statistics, company announcements, journalism, and personal anecdotes.
For generative AI, verify factual claims with reliable sources. For predictions about the future, look for assumptions and ranges rather than treating a single number as a fact. Responsible media literacy means being willing to update your view when better evidence appears.
Future Directions
AI capabilities will continue to change, but social outcomes are not determined by technology alone. Institutions decide where systems are used, companies choose business models, regulators set rules, communities establish norms, and users shape adoption.
The most useful future-oriented question is therefore not "Will AI be good or bad?" but "Which AI system, used by whom, for what purpose, under which safeguards, with what evidence, and with what alternatives?" This question keeps attention on choices that people can actually debate and change.
Sources and Further Reading
- UNESCO Recommendation on the Ethics of Artificial Intelligence: A global human-rights-centered framework for ethical AI.
- NIST AI Risk Management Framework: A voluntary framework for identifying and managing AI risks.
- OECD AI Principles: Intergovernmental principles for innovative and trustworthy AI.
- European Commission AI Act overview: Official information about the European Union's risk-based AI regulation.
- IEA Energy and AI report: Evidence and analysis about data centers, electricity, and the energy-AI relationship.
Interactive Tasks
Quiz: Test Your Knowledge
Why is an AI system best understood as socio-technical? (Because technical components interact with people institutions and social choices) (!Because every AI system uses a humanoid robot) (!Because AI systems operate without human decisions) (!Because social effects can be predicted from code alone)
What happens during inference in a machine-learning system? (A trained model processes new inputs to produce outputs) (!The legal status of the system is decided) (!All training data are deleted automatically) (!Humans manually write every model parameter)
Why can a fluent generative AI answer still be unreliable? (The system can generate plausible content without verifying every claim) (!Fluent language proves that every source was checked) (!Generative AI can only repeat exact sentences from training data) (!Language models are unable to produce incorrect statements)
Which situation is an example of meaningful human oversight? (A qualified person can review evidence and override the automated recommendation) (!A person clicks approve without enough time to inspect the case) (!The system makes the final decision and hides its reasoning) (!The user is told that computers are always more objective)
Why can historical training data create unfair outcomes? (The data may contain patterns produced by earlier unequal treatment) (!Historical data always contain too few numbers) (!Old data cannot be processed by modern computers) (!Fairness is guaranteed whenever a dataset is large)
What is data minimization? (Collecting only the personal data needed for a legitimate purpose) (!Collecting every available data point for possible future use) (!Publishing personal data to make a system transparent) (!Replacing all human decisions with automated decisions)
What is a strong first response to a surprising AI-generated-looking video? (Check the source context and independent evidence before sharing it) (!Assume it is false because synthetic media exist) (!Assume it is true because the video looks realistic) (!Share it immediately so others can decide later)
Why should employment claims about AI be treated cautiously? (AI can automate augment and reshape tasks differently across sectors) (!Every occupation is affected in exactly the same way) (!Job outcomes depend only on model size) (!Employment effects can be known perfectly years in advance)
Which set names the four core functions of the NIST AI Risk Management Framework? (Govern Map Measure and Manage) (!Design Sell Predict and Replace) (!Collect Rank Monetize and Expand) (!Code Train Publish and Forget)
What is the best way to evaluate an AI system with serious consequences? (Combine technical evidence with analysis of context harms rights and accountability) (!Use only one accuracy score) (!Judge the system by how modern its interface looks) (!Assume automation is fairer than human decision-making)
Memory Game
| Training data | Examples used to fit patterns during model development |
| Model | Learned mathematical structure used to produce outputs |
| Inference | Use of a trained system on new inputs |
| Algorithmic bias | Systematic skew that can contribute to unfair outcomes |
| Human oversight | Meaningful review with power to question or override a system |
| Deepfake | Synthetic media that convincingly imitates a person's appearance or voice |
Drag and Drop
| Match the correct terms. | Topic |
|---|---|
| Human oversight | A qualified person can review and override a high-stakes automated recommendation |
| Data minimization | Personal information is limited to what is needed for a legitimate purpose |
| Content provenance | Evidence records where media came from and how it was changed |
| Accessibility | A system is designed so people with different abilities can use it |
| Red teaming | A system is deliberately tested for failures misuse and unexpected behavior |
...
Crossword Puzzle
| Bias | What word describes systematic skew that can contribute to unfair outcomes? |
| Privacy | What principle concerns appropriate control and use of personal information? |
| Dataset | What collection of examples may be used to train or evaluate a model? |
| Deepfake | What word describes convincing synthetic media that imitates a person? |
| Transparency | What principle concerns making important information about a system understandable? |
| Accountability | What principle asks who is responsible for decisions and harms? |
LearningApps
Cloze Text
Open-Ended Tasks
Easy
- AI Diary: Keep a one-day diary of five situations in which you encounter an AI-driven or algorithmic system, then describe what the system appears to optimize and what evidence supports your classification.
- Verification Challenge: Choose one surprising online image or claim about AI, trace it to reliable sources, and write a short explanation of the verification steps you used.
- Concept Poster: Create an image or poster that explains the difference between training, a model, and inference to learners who have never studied machine learning.
- Mini Interview: Interview a classmate, teacher, or family member about one AI tool they use and summarize one benefit, one risk, and one unanswered question.
Standard
- Bias Case Study: Investigate a documented case of algorithmic bias, identify the data or design choices involved, and produce a two-page evidence-based analysis with proposed safeguards.
- AI in the Workplace: Interview someone in a profession affected by automation or AI assistance and compare their experience with claims from at least two reliable published sources.
- Synthetic Media Video: Produce a short educational video that teaches viewers how to verify suspicious synthetic media without claiming that visual inspection alone can prove authenticity.
- School AI Policy: Draft a balanced school policy for generative AI that distinguishes learning support, prohibited assistance, disclosure, privacy, accessibility, and assessment integrity.
Advanced
- Algorithmic Impact Assessment: Choose a hypothetical high-stakes AI system and create an impact assessment covering stakeholders, benefits, failure modes, bias, privacy, security, appeals, and alternatives.
- Model Evaluation Experiment: Use a safe publicly available classifier or language model to test the same task across systematically varied inputs, record results, analyze limitations, and explain why your sample cannot support universal conclusions.
- Public Institution Visit: Visit or contact a library, local government office, university lab, consumer organization, or data-protection body and investigate how it approaches automated decision-making, digital rights, or AI literacy.
- Citizen Hearing Simulation: Organize and record a mock public hearing in which students represent developers, affected citizens, workers, regulators, and civil-society groups debating whether a proposed AI system should be deployed.
Learning Assessment
- System Analysis: Analyze a real AI application by connecting its technical purpose with stakeholders, incentives, data practices, possible harms, safeguards, and non-AI alternatives.
- Fairness Reasoning: Given a case in which two fairness metrics point to different conclusions, explain the value conflict and defend which criterion should matter most in that context.
- Evidence Evaluation: Compare a company claim, a news report, and an independent research source about the same AI technology, then judge which conclusions are supported and which remain uncertain.
- Governance Design: Design a governance plan for a high-stakes AI system that includes documentation, human oversight, monitoring, appeals, and clear responsibility for corrective action.
- Media Literacy Transfer: Apply lateral reading and source verification to a piece of synthetic or potentially synthetic media, then explain how your method could transfer to another subject area.
- Sustainability Trade-Off: Evaluate whether an AI application with measurable energy costs is justified by its expected social benefit and compare it with at least one lower-compute alternative.
Evidence of Learning
- Knowledge
- You can explain training, inference, models, generative AI, algorithmic bias, privacy, deepfakes, human oversight, accountability, and risk-based governance in your own words.
- Skills
- You can verify claims, read charts critically, compare sources, identify stakeholders, evaluate trade-offs, recognize uncertainty, and distinguish technical performance from social acceptability.
- Products
- Strong evidence may include an AI diary, verification report, poster, interview summary, bias case study, policy draft, impact assessment, experiment report, educational video, or public-hearing simulation.
- Reasoning
- You can justify a recommendation with evidence rather than relying on hype, fear, or a single performance metric.
- Transfer
- You can apply the same evaluation framework to unfamiliar AI systems in education, work, healthcare, public administration, culture, media, or environmental management.
OERs on the Topic
Linked Learning Areas
