Die Druckversion wird nicht mehr unterstützt und kann Darstellungsfehler aufweisen. Bitte aktualisiere deine Browser-Lesezeichen und verwende stattdessen die Standard-Druckfunktion des Browsers.

Artificial Intelligence

Introduction

Artificial Intelligence (AI) is the scientific and engineering field concerned with building computational systems that perform tasks associated with intelligent behavior, such as perception, reasoning, learning, language use, planning, and decision-making. AI is not a single algorithm or product. It is an interdisciplinary field drawing on computer science, mathematics, statistics, cognitive science, linguistics, philosophy, neuroscience, and engineering.

For university-level study, it is useful to distinguish a system's observable capability from claims about human-like understanding or consciousness. A model can generate fluent text, classify medical images, recommend products, or control a robot without possessing the full range of human cognition. Most deployed AI today is task-specific: it is designed or trained for a defined set of objectives. Artificial general intelligence remains a research idea rather than an established class of deployed systems.

The diagram below shows a common relationship among AI, machine learning, and deep learning. Machine learning is an important subfield of AI, while deep learning is a family of machine-learning methods based on multilayer neural networks. Not every AI method uses machine learning.

This introductory video gives a broad overview of what AI can and cannot do and is useful for checking your prior assumptions about the field.

Learning Outcomes

By the end of this aiMOOC, you should be able to explain major AI paradigms, distinguish core learning settings, interpret common model architectures and evaluation measures, design a basic AI workflow, critique claims about AI capabilities, identify ethical and governance risks, and propose evidence-based ways to evaluate an AI system in a real-world context.

You should also be able to connect technical choices to social consequences. For example, the choice of training data affects not only predictive performance but also whose situations are represented, while the choice of evaluation metric affects which kinds of errors are prioritized.

Foundations and History

From Computation to Artificial Intelligence

The history of AI combines formal logic, computation, psychology, statistics, engineering, and changing expectations about what machines can do. In 1950, Alan Turing proposed what became known as the imitation game as a way to discuss machine intelligence operationally. The 1955 proposal for the Dartmouth Summer Research Project on Artificial Intelligence, organized for 1956, helped establish the term artificial intelligence as the name of a research field.

Early AI research emphasized symbolic approaches: programs represented objects, facts, goals, and rules explicitly and manipulated these representations through search and logical inference. Later decades saw periods of rapid optimism and periods of reduced funding and interest, often called AI winters, when technical progress or practical impact fell short of expectations.

From the 1980s onward, statistical learning became increasingly influential. The growth of digital data, improved algorithms, and increasingly powerful processors supported major advances in machine learning. During the 2010s, deep neural networks produced strong results in areas such as speech recognition, computer vision, and natural language processing. The transformer architecture introduced in 2017 became especially important for large-scale language and multimodal models.

A key lesson from this history is that AI progresses through interactions among ideas, data, hardware, software, institutions, benchmarks, and economic incentives. Technical breakthroughs alone do not determine how AI is adopted.

Core AI Paradigms

Several paradigms coexist within AI. Symbolic AI represents knowledge explicitly using rules, symbols, logic, search spaces, or constraints. Probabilistic AI models uncertainty through probability distributions and statistical inference. Machine learning fits models from data rather than encoding all decision rules manually. Evolutionary computation searches for solutions through population-based optimization inspired by biological evolution. Robotics integrates perception, planning, learning, and control in systems that act in physical environments.

Modern systems often combine paradigms. A robot may use a neural network for perception, a probabilistic model for state estimation, a search algorithm for planning, and rule-based constraints for safety. Therefore, asking whether a system is symbolic or learning-based can be less useful than asking what representations, algorithms, objectives, and interfaces it uses.

Machine Learning

Learning from Data

Machine learning studies methods that improve performance on a task by learning patterns from data or interaction. A typical learning problem defines inputs, outputs or rewards, a model family, an objective function, an optimization method, and an evaluation procedure.

Supervised learning uses labeled examples. Classification predicts categories, while regression predicts continuous values. Unsupervised learning looks for structure in unlabeled data, for example through clustering or dimensionality reduction. Self-supervised learning creates learning signals from the data itself, such as predicting masked or next elements in a sequence. Reinforcement learning trains an agent to select actions in an environment so as to maximize expected cumulative reward.

The following regression plot illustrates the idea of fitting a simple predictive relationship to data points.

Decision trees are another important model family. They partition the input space through a sequence of decisions and can be used for classification or regression.

This Computerphile video contrasts supervised and unsupervised machine-learning methods.

Reinforcement Learning

In reinforcement learning, an agent observes a state or observation, takes an action, receives a reward, and encounters a new state. The learning challenge is not simply to imitate correct labels but to discover a policy that performs well over time. This introduces questions about exploration versus exploitation, delayed rewards, uncertainty, and credit assignment.

Reinforcement learning is used in domains such as games, simulation, robotics, resource allocation, and control. However, a reward function is only a proxy for the designer's intended objective. Poorly specified rewards can lead to behavior that optimizes the metric while violating the real goal, illustrating why objective design and monitoring matter.

The Machine-Learning Workflow

A rigorous workflow separates the creation of a model from the evidence used to judge it. Data are collected or selected, cleaned, documented, transformed, divided into appropriate subsets, and used to train candidate models. Hyperparameters are tuned using validation procedures. The final test data should remain separate from training decisions so that the reported performance better estimates generalization to unseen examples.

Important threats to validity include data leakage, unrepresentative samples, duplicated observations, label errors, class imbalance, distribution shift, and choosing metrics that do not reflect the real decision context. Reproducible experiments require versioned data, code, model configurations, and documented random seeds or sources of stochasticity where relevant.

Evaluation and Generalization

A model that performs well on training data may still fail on new data. Overfitting occurs when a model captures idiosyncrasies of the training set that do not generalize. Underfitting occurs when the model is too limited, poorly optimized, or insufficiently trained to capture important structure.

For classification, common metrics include accuracy, precision, recall, F1 score, and area-under-curve measures. No metric is universally best. In a high-stakes screening task, false negatives and false positives can have very different consequences, so metric selection should follow the application and error costs. For regression, mean absolute error and mean squared error are common but emphasize errors differently.

The bias-variance perspective helps explain why model complexity, dataset size, regularization, and noise affect generalization.

Good evaluation goes beyond a single average score. You should examine subgroup performance, uncertainty, calibration, robustness to plausible perturbations, behavior under distribution shift, and failure modes. When a system supports consequential decisions, human review and domain-specific validation may be essential parts of the evaluation design.

Neural Networks and Deep Learning

Artificial Neural Networks

An artificial neural network is a parameterized function built from connected computational units arranged in layers. Each unit typically combines inputs using learned weights and a bias, then applies a nonlinear activation function. During training, an objective or loss function measures error, and an optimization algorithm updates parameters to reduce that loss.

Backpropagation efficiently computes gradients of the loss with respect to network parameters by applying the chain rule through the computation graph. Gradient-based optimizers then use those gradients to update the parameters. Deep learning uses neural networks with multiple layers, enabling the system to learn increasingly complex representations.

The following visual explanation develops the intuition behind neurons, layers, weights, biases, and learned representations.

Different architectures encode different assumptions. Convolutional neural networks are effective for grid-like data such as images because they share local filters. Recurrent neural networks process sequences through recurrent state. Transformers use attention mechanisms to relate elements across a sequence and now play a major role in language, vision, audio, and multimodal AI.

Transformers and Attention

The transformer architecture uses attention and feed-forward layers rather than recurrence as its central sequence-processing mechanism. In self-attention, representations of tokens are transformed into queries, keys, and values. Attention weights determine how strongly each token draws information from other tokens in the context.

Transformers are flexible rather than magical. Their performance depends on training objectives, data, model scale, optimization, context, and deployment design. A language model predicts or generates tokens according to learned statistical structure. Fluent output is therefore not a guarantee of factual correctness, valid reasoning, or trustworthy citation.

For a more technical university-level introduction to modern deep learning, the following MIT lecture provides mathematical and architectural context.

Generative AI and Foundation Models

Generative AI produces new content such as text, images, audio, video, code, or structured data. Generative models learn patterns in training data and sample outputs from a learned distribution or conditional generation process. Examples include autoregressive language models, diffusion models, variational autoencoders, and generative adversarial networks.

The diagram below contrasts a discriminative model, which maps an input to a prediction, with a generative model, which can produce new samples.

Large models trained on broad datasets are often called foundation models when they can be adapted to many downstream tasks. Adaptation methods include prompting, retrieval-augmented generation, fine-tuning, tool use, and domain-specific system design. Each method changes the system's capabilities and risks in different ways.

A generative model may produce hallucinations: plausible-looking but unsupported or incorrect content. Retrieval and external tools can reduce some errors, but they do not eliminate the need for verification. In scholarly work, you should verify factual claims against primary or authoritative sources and follow your institution's rules for declaring AI assistance.

Human Feedback and Alignment

Developers can shape model behavior with preference data, supervised instruction tuning, reinforcement learning from human feedback, rule-based methods, and other alignment techniques. These methods try to make outputs more useful, safe, or consistent with specified goals, but they do not solve every problem of reliability or social values.

Human feedback itself can be inconsistent, culturally situated, or incomplete. An alignment process therefore involves choices about whose preferences are collected, how disagreements are handled, how evaluators are trained, and which failure modes are measured.

Reasoning, Search, and Knowledge Representation

Search and Planning

Many AI problems can be represented as a state space containing possible situations and actions. Search algorithms explore this space to find a path from an initial state to a goal. Breadth-first search and depth-first search use different exploration strategies, while informed methods such as A* use a heuristic to estimate progress toward a goal.

Planning extends search by reasoning about actions, preconditions, effects, and goals. In complex domains, the challenge is often not merely to find any plan but to find one that is efficient, robust, safe, and adaptable when the environment changes.

Knowledge and Uncertainty

Knowledge representation concerns how a system encodes facts, concepts, relations, rules, or ontologies so that they can be used for reasoning. Classical logic provides precise inference under explicit assumptions, while probabilistic methods represent uncertainty. Bayesian networks, for example, model dependencies among variables and support probabilistic inference.

The choice of representation constrains what a system can express and which computations are feasible. A symbolic representation may support transparent rule tracing, while a learned vector representation can capture statistical similarity without yielding an obvious human-readable explanation. Hybrid systems can use both.

AI Applications

Language, Vision, Robotics, Science, and Decision Support

Natural language processing includes translation, information extraction, question answering, summarization, dialogue, and text generation. Computer vision includes image classification, object detection, segmentation, tracking, and visual generation. Robotics combines perception and decision-making with physical action. AI also supports scientific discovery through tasks such as pattern detection, surrogate modeling, simulation, and candidate generation.

In professional settings, AI often functions as decision support rather than autonomous decision-making. A system may rank cases, draft text, detect anomalies, or recommend actions while a human remains responsible for interpretation and approval. This division of labor should be designed deliberately rather than assumed.

Applications in healthcare, finance, employment, education, public services, and infrastructure can have high stakes. In these settings, performance on a benchmark is insufficient evidence for safe deployment. You must consider data provenance, population fit, user competence, error severity, monitoring, contestability, and accountability.

AI as a Sociotechnical System

An AI model operates inside a larger system of people, policies, interfaces, incentives, datasets, organizations, and institutions. A technically accurate model can still produce harmful outcomes if it is used for the wrong purpose, presented with misleading confidence, integrated into an unfair process, or deployed without recourse for affected people.

Thinking sociotechnically means asking not only Can the model predict? but also Who defines the objective?, Who benefits?, Who bears the errors?, Can decisions be challenged?, and What happens when conditions change?

Responsible AI and Governance

Bias, Fairness, and Accountability

AI systems can reproduce or amplify patterns found in training data and institutional processes. Bias can enter through data collection, measurement, labeling, sampling, model design, evaluation, or deployment. Fairness is not a single mathematical property: different fairness criteria can conflict, especially when groups have different base rates or when the underlying labels reflect unequal social conditions.

Joy Buolamwini's talk below illustrates how performance disparities in facial-analysis systems can reveal broader questions about representation, measurement, and accountability.

Responsible practice includes documenting data provenance, testing across relevant subgroups, involving domain experts and affected stakeholders, recording known limitations, enabling meaningful human oversight, and establishing processes for appeal or correction when automated outputs affect people.

Privacy, Security, Transparency, and Safety

Privacy concerns how personal information is collected, processed, retained, inferred, and shared. AI can create new privacy risks by extracting patterns or sensitive attributes from large datasets. Privacy-preserving techniques, access controls, minimization of collected data, and clear governance can reduce risk, but their suitability depends on the application.

Security concerns attacks on models, data, and surrounding infrastructure. Examples include adversarial inputs, data poisoning, model theft, prompt injection in tool-using systems, and unauthorized access to sensitive outputs. Security testing should therefore cover the whole AI system, not just the model.

Transparency can include documentation of purpose, training or evaluation data, limitations, metrics, uncertainty, and human oversight. Explainability aims to make a model's behavior or decision basis more interpretable. Different audiences need different explanations: a developer debugging a model, a clinician reviewing a recommendation, and a student receiving an automated assessment do not need the same information.

Safety means reducing the probability and severity of harmful outcomes. This can involve pre-deployment testing, red teaming, constrained permissions, staged rollout, monitoring, incident response, and the ability to stop or roll back a system.

Risk-Based Governance

Governance turns principles into operational responsibilities. A practical risk-management process identifies the system's intended use, affected stakeholders, foreseeable misuse, severity of possible harms, uncertainty, legal and organizational constraints, evaluation criteria, and monitoring plans.

Risk should be assessed in context. A recommendation model for entertainment and an AI system that influences access to medical care require different evidence, oversight, and tolerance for error. Governance should also cover the full lifecycle: design, data collection, training, testing, deployment, monitoring, updating, and retirement.

From Prototype to Deployment

MLOps, Monitoring, and Change

A model that works in a notebook is not yet a dependable production system. Deployment introduces versioning, reproducibility, interfaces, data pipelines, latency constraints, access control, testing, monitoring, incident response, and maintenance. MLOps applies software-engineering and operations practices to machine-learning systems.

After deployment, input data and user behavior may change. Data drift refers to changes in the distribution of inputs, while concept drift refers to changes in the relationship between inputs and target outcomes. Monitoring should therefore track not only uptime but also data quality, model performance, subgroup behavior, safety events, and unexpected usage patterns.

A mature AI lifecycle includes clear ownership, thresholds for intervention, procedures for retraining or rollback, and records that allow decisions to be audited. Continuous improvement should not mean uncontrolled change.

Interactive Tasks

Quiz: Test Your Knowledge

Which description best captures artificial intelligence as a field? (Computational systems designed to perform tasks associated with intelligent behavior) (!Only humanoid robots that imitate people) (!Any program that stores a large amount of data) (!A single algorithm used for all intelligent tasks)

What is the defining feature of supervised learning? (Training uses examples paired with target labels or values) (!Training always requires interaction with a physical robot) (!Training has no objective function) (!Training uses only randomly generated data)

What does overfitting mean? (A model fits training-specific patterns that do not generalize well) (!A model is always too small to learn the training data) (!A model uses no parameters) (!A model can only perform regression)

What mechanism is central to transformer architectures? (Attention) (!Binary search) (!K means clustering) (!Rule chaining only)

Which metric asks what fraction of predicted positives are actually positive? (Precision) (!Recall) (!Mean squared error) (!Training loss only)

What distinguishes reinforcement learning from ordinary supervised learning? (An agent learns from rewards while interacting with an environment) (!Every example comes with a correct class label) (!The model never receives feedback) (!The model can only process images)

What is a generative model designed to do? (Model and produce new samples consistent with learned data patterns) (!Store every training example without transformation) (!Guarantee that every generated statement is factually true) (!Replace all evaluation with human intuition)

Why can biased training data create problems? (It can cause a model to reproduce or amplify unrepresentative patterns) (!It guarantees equal performance for every subgroup) (!It removes the need for evaluation) (!It makes optimization mathematically impossible)

Why should final test data remain separate from model tuning? (To obtain a less biased estimate of performance on unseen data) (!To make the training set larger) (!To prevent the model from having parameters) (!To eliminate all uncertainty from deployment)

What is an important reason for human oversight in high stakes AI use? (Consequences and context may require judgment beyond a model score) (!Human oversight guarantees that all decisions are correct) (!AI systems cannot process numerical data) (!Oversight makes data governance unnecessary)

Memory Game

Supervision	Learning from examples with target labels or values
Clustering	Grouping data by similarity without predefined target classes
Attention	Mechanism that weights relationships among elements in a context
Policy	Strategy that maps observations or states to actions
Calibration	Agreement between predicted confidence and observed frequency
Regularization	Technique used to discourage overly complex fitted solutions
Inference	Use of a trained model to produce an output for new input

Drag and Drop

Match the correct terms.	Topic
Classification	Predict a discrete category
Regression	Predict a continuous quantity
Clustering	Discover groups in unlabeled data
Generation	Produce new content from a learned distribution
Reinforcement learning	Learn actions from reward and interaction

...

Crossword Puzzle

Algorithm	What general term describes a defined computational procedure for solving a problem?
Inference	What is the use of a trained model to produce outputs for new inputs called?
Dataset	What is a structured collection of examples used for analysis or learning called?
Transformer	Which neural architecture is built around attention mechanisms?
Fairness	What concept concerns equitable treatment and performance across relevant groups?
Robotics	Which field integrates computation with sensing and physical action?

LearningApps

Cloze Text

Artificial intelligence is a broad field that includes both learning-based and

methods. Supervised learning relies on examples paired with

. A model should be evaluated on data that were not used to make training decisions so that researchers can estimate

. Reinforcement learning trains an agent using signals called

. Deep neural networks learn parameters through gradient-based optimization supported by

. Transformers rely heavily on

to relate elements in a context. Generative models can produce new content but may also produce unsupported claims called

. Responsible evaluation should examine performance, uncertainty, and relevant

. Privacy and security must be considered across the entire AI

. Human oversight is especially important when model outputs influence

decisions.

Open-Ended Tasks

Easy

AI concept map: Create a one-page concept map that connects artificial intelligence, machine learning, deep learning, symbolic AI, robotics, and generative AI; add one example and one limitation for each branch.
AI claim check: Choose three public claims about an AI tool and write a short evidence-based note for each claim explaining what evidence would be needed to verify it.
Model card sketch: Design a simple visual model card for a hypothetical university chatbot that states its intended use, users, data needs, limitations, and three risks.
AI explainer video: Produce a two-minute video that explains the difference between supervised learning, unsupervised learning, and reinforcement learning to first-year students.

Standard

Dataset audit: Inspect a small public dataset, document its variables and missing values, identify at least three possible sources of bias, and propose changes that would improve its suitability for an AI project.
Classification experiment: Train or simulate two simple classifiers on the same dataset, compare at least three evaluation metrics, and explain why the metrics lead to the same or different conclusions.
AI practitioner interview: Interview a researcher, engineer, librarian, lecturer, or other professional who works with AI and summarize how they evaluate reliability, handle data, and respond to mistakes.
Responsible AI field visit: Visit a university lab, library, makerspace, company, public institution, or virtual open day where AI is used; document the workflow and identify where human oversight enters the process.

Advanced

Distribution shift study: Design an experiment that changes an input distribution after training, measure how model performance changes, and propose a monitoring threshold that could trigger review or retraining.
Fairness trade-off analysis: Select a realistic decision-support scenario, compare two fairness criteria, show where they may conflict, and defend a governance approach that includes affected stakeholders.
Generative AI evaluation protocol: Create and pilot an evaluation protocol for a generative system that measures factuality, usefulness, harmful failure modes, uncertainty, and reproducibility across at least two prompt conditions.
AI governance project: Develop a lifecycle governance plan for a high-impact AI application, including ownership, documentation, risk assessment, validation, monitoring, incident response, appeal mechanisms, and retirement criteria.

Learning Assessment

Model evaluation case: Given a classification problem with unequal costs for false positives and false negatives, select suitable metrics, justify them, and explain what additional evidence would be needed before deployment.
Architecture comparison: Compare a symbolic rule system, a decision tree, and a neural network for the same application and reason about interpretability, data requirements, robustness, maintenance, and likely failure modes.
Data leakage diagnosis: Analyze a hypothetical experiment in which future information accidentally enters the training features, explain why the reported score is misleading, and redesign the evaluation procedure.
Sociotechnical impact analysis: Map stakeholders for an AI-supported university admissions process, identify technical and institutional risks, and propose safeguards that address both.
Generative model critique: Examine a set of generated outputs, categorize failure modes, test whether retrieval or prompting reduces them, and explain which errors still require human verification.
Deployment decision memo: Write a recommendation on whether to deploy an AI system, using evidence about performance, uncertainty, subgroup behavior, privacy, security, human oversight, and monitoring.

Evidence of Learning

Evidence area	What strong evidence looks like
Knowledge	You accurately explain core AI paradigms, learning settings, neural-network concepts, transformer architecture, evaluation principles, and major governance concerns.
Skills	You can frame an AI problem, inspect data, select suitable metrics, compare models, interpret errors, test assumptions, document limitations, and communicate uncertainty.
Products	You produce reproducible experiments, model or system documentation, visual explanations, evaluation protocols, risk analyses, and reasoned recommendations.
Transfer	You apply AI concepts to unfamiliar domains, recognize when benchmark performance is insufficient, and adapt evaluation or governance to the consequences of a new context.
Reflection	You can distinguish technical performance from social value, identify assumptions in an AI workflow, and explain how evidence could change your conclusion.

OERs on the Topic

The English Wikipedia article provides an openly accessible overview and links to major subfields, historical developments, applications, and debates. Use it as a starting point and follow its citations to primary or specialized sources for academic work.