07 DTC 101

Artificial Intelligence

Artificial intelligence is a broad term for computer systems that perform tasks associated with human intelligence, such as recognizing patterns, classifying information, making predictions, generating language, producing images, or helping people make decisions. AI systems do not work like human minds, but they can imitate some of the outward signs of intelligence because they are built from data, computation, and statistical patterns.

This chapter introduces artificial intelligence as both a technical system and a cultural force. It follows the earlier chapters on digital technology, media, networks, social platforms, and data because AI is made from all of these. AI depends on hardware and software, large collections of data, networked platforms, and the cultural habits of people who produce and share images, text, sounds, and code.

The goal of this chapter is not to turn you into a computer scientist. Instead, it gives you a practical and critical vocabulary for understanding how AI works, what it can and cannot do, and how it is changing writing, art, research, work, entertainment, and everyday communication.

7.1 A Brief History of AI

The dream of artificial intelligence is older than computers. Myths, automata, mechanical calculators, puppets, and fictional robots all express a desire to create artificial beings or intelligent tools. During the twentieth century, this dream became connected to computing. Once information could be represented as code, researchers began asking whether reasoning, language, perception, and creativity could also be represented as computational processes.

Early AI research often focused on symbolic AI, sometimes called Good Old-Fashioned AI or GOFAI. These systems attempted to model intelligence through explicit rules, logic, and symbolic representations. If a computer could be given enough rules, perhaps it could solve problems in ways that resembled human reasoning.

In 1950, Alan Turing proposed what became known as the Turing Test, asking whether a machine could imitate human conversation well enough that a person could not reliably distinguish it from another human. Early AI researchers believed that intelligence could be recreated through carefully designed rules and logical reasoning. These symbolic systems performed well in limited domains but struggled with the ambiguity and complexity of everyday life.

ELIZA

ELIZA was an early program created in the 1960s by Joseph Weizenbaum. It mimicked a conversation with a therapist using simple pattern matching. While it could simulate conversation, it did not understand the meaning behind the words. ELIZA showed both the promise and the limits of early symbolic AI.

ELIZA: natural language processing program by Joseph Weizenbaum

By the 1990s and 2000s, growing computing power, larger datasets, and improved algorithms shifted AI research toward machine learning. In machine learning, computers are not programmed with every rule in advance. Instead, they are trained on examples. Rather than telling the computer exactly how to identify a cat, translate a sentence, or recommend a video, machine learning systems look for statistical patterns in large collections of data.

Some important milestones in this shift include:

1950 – Alan Turing proposes the Turing Test.
1966 – ELIZA demonstrates conversational pattern matching.
1997 – IBM's Deep Blue defeats world chess champion Garry Kasparov.
2012 – AlexNet dramatically improves image recognition using deep neural networks.
2016 – DeepMind's AlphaGo defeats world champion Lee Sedol using deep learning and reinforcement learning.
2017 – Google researchers introduce the Transformer architecture, which becomes the foundation for modern large language models.
2022 – ChatGPT brings conversational AI to hundreds of millions of users.

AlphaGo

AlphaGo showed that AI could master a game of deep strategy and intuition. It learned not only from rules, but from examples, self-play, and reinforcement learning. Its victory over Lee Sedol in 2016 became a turning point in public understanding of AI because the system made moves that surprised even expert human players.

AlphaGo: The Movie, Google DeepMind

From Rules to Patterns

The history of AI can be understood as a movement from rules to patterns. Rule-based systems work well when the world can be clearly described in advance. Pattern-based systems excel when the world is complex, ambiguous, and too variable to describe with explicit rules. Contemporary AI depends on this pattern-based approach.

“What we want is a machine that can learn from experience.”
— Alan Turing, Intelligent Machinery (1948)

For students of digital technology and culture, the important question is not only whether AI is intelligent, but what kinds of human activities are being translated into data, prediction, automation, and generation. The next section looks at how the pattern-based approach actually works.

7.2 Machine Learning and Neural Networks

Machine learning allows computers to improve their performance by learning from examples rather than following a fixed set of rules. Instead of programming every possible situation, developers train a system using many examples so it can recognize patterns and make predictions about new data.

Image recognition provides a familiar example. Rather than telling a computer exactly how to identify a face, tree, or stop sign, a machine learning system is trained on thousands or millions of labeled images. Over time, it learns statistical patterns that allow it to recognize similar images it has never seen before.

Training, Models, and Neural Networks

In AI, a model is the result of the training process. During training, the model repeatedly compares its predictions with known answers and gradually improves its performance. Rather than storing a simple collection of facts, the model learns relationships among patterns in the training data.

Most contemporary AI systems use neural networks. Although inspired by the organization of neurons in the brain, neural networks are mathematical systems made up of layers of connected nodes. Each connection has a numerical weight that represents its importance. As the network trains, these weights are continually adjusted so that the model makes increasingly accurate predictions. This process of gradually reducing error is what allows the network to learn.

Neural Networks: a short introduction

Training data provides examples.
The model makes a prediction.
The prediction is compared with the correct answer.
The network adjusts its internal weights to reduce error.
After many repetitions, the model becomes increasingly accurate.

The learning process can be visualized as the wearing of paths through a landscape. The demonstration below represents a small neural network as a field of possible connections between input words and output words. Each connection is drawn as a line, and its thickness represents the strength—or weight—of that connection. At first, every path is nearly equal. As the model trains, some paths become stronger while others fade away, gradually revealing the patterns the network has learned. Although real language models are vastly more complex and use mechanisms such as attention (introduced in Section 7.3), the underlying idea remains the same: learning consists of strengthening useful pathways while weakening less useful ones.

Interactive Demonstration: The Worn Path

A small neural network shown as a field of connections. Run a forward pass to send a signal through the network, then "backpropagate to adjust the weights. Repeat the cycle and the deepest paths from input to output will emerge. The network's task is to predict the word mat from the input the cat sat.

loss– prediction– targetmat steps0

Untrained. The field is unmarked.

signal moving forward error moving backward channel depth (the weights)

In an actual model, the patterns are discovered by the training process rather than set in advance. Once the channels are carved, the weights freeze, and every later use of the model is a forward pass through a fixed landscape.

A trained model is not a memory store but a landscape of channels worn by repeated training. When the model is asked to produce a result, the signal follows the deepest channels available. For a familiar or generic request, those channels lead to the most expected output, which is one reason why bland prompts tend to produce bland responses.

Machine Vision

Machine vision is the branch of AI that enables computers to interpret photographs and video. It was one of the first areas where deep neural networks dramatically outperformed earlier AI techniques. In 2012, a system called AlexNet showed that a deep neural network could classify photographs far more accurately than previous methods, reshaping AI research almost overnight. Within a few years, similar techniques powered face recognition, photo search, medical image analysis, manufacturing inspection, and the perception systems used by autonomous vehicles.

The process resembles the wearing of paths through a landscape. A neural network trained on millions of labeled images gradually strengthens connections that detect increasingly complex visual patterns. Early layers respond to simple edges and colors, later layers recognize textures and shapes, and deeper layers respond to complete objects such as faces, bicycles, or stop signs. Rather than being explicitly programmed with these concepts, the network learns them through repeated exposure to examples.

Machine vision also reminds us that AI learns from human-created datasets. If the training images reflect social biases, unequal representation, or labeling errors, the model can reproduce those patterns. The same system that accurately recognizes one face may perform poorly on another, demonstrating that the quality and diversity of training data are as important as the algorithm itself.

7.3 Large Language Models

The worn-path picture from Section 7.2 applies not only to images, but also to sounds, code, and language. When the same pattern-learning machinery is trained on enormous collections of digital text, the result is a large language model (LLM). LLMs generate language by predicting likely sequences of word-parts, called tokens. A token may be a whole word, part of a word, a punctuation mark, or another small unit of text.

When you type a prompt into a chatbot, the model does not look up a fixed answer in a database. Instead, it predicts which token is most likely to come next, given the prompt, the conversation so far, and the statistical patterns learned during training. It generates one token, adds it to the response, and then repeats the same process again. A complete response is simply this prediction repeated, one token after another, until the model decides to stop.

How Large Language Models Work

Next-Token Prediction

The phrase next-token prediction sounds simple, but when it is performed by a very large model trained on enormous amounts of language, it can produce surprisingly capable results. LLMs can summarize, translate, explain, imitate genres, write code, brainstorm ideas, and carry on dialogue because human language already contains patterns of reasoning, instruction, style, and knowledge. The model is not reasoning in the same way a person does. It is continuing patterns of writing in which other people reasoned.

One surprising discovery is that sufficiently large language models develop abilities that were never explicitly programmed. Researchers call these emergent behaviors. As models become larger and are trained on more diverse data, they may acquire new capabilities such as translation, summarization, coding, or simple forms of reasoning. These abilities emerge from learning statistical patterns across language rather than from separate software modules designed for each task.

“A stochastic parrot is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data.”
— Emily M. Bender, Timnit Gebru, Angelina McMillan-Major & Margaret Mitchell, On the Dangers of Stochastic Parrots (2021)

This metaphor reminds us that fluent language is not the same as understanding. LLMs can help people write, learn, code, organize information, and explore ideas. They can also produce convincing falsehoods, flatten personal style, imitate authority, and obscure the difference between knowledge and plausible-sounding language. The same statistical machinery produces both fluent truths and fluent errors.

Transformers and Attention

Most contemporary LLMs are built on a machine learning architecture called the transformer, introduced in the influential 2017 paper Attention Is All You Need. A transformer uses a process called attention to weigh the relationships among tokens in a sequence. Rather than processing one word at a time, it considers how all of the tokens in the input relate to one another, allowing the model to track meaning across long passages and respond to context.

The same architecture can be adapted to many kinds of data. With appropriate training, transformers power today's leading systems for language (GPT), images (DALL·E), audio (Whisper), video, and code. This shared architecture is one reason contemporary AI is increasingly described as multimodal.

For this course, the key idea is that LLMs are cultural technologies built from language. They do not simply process information. They process the accumulated traces of human writing, conversation, documentation, argument, imagination, and error. The next section turns to a different application of the same pattern-based approach, this time applied to images rather than words.

7.4 Image Generation and Diffusion Models

Where large language models generate text by continuing patterns in language, diffusion models generate images by learning patterns in visual data. During training, the model learns how to gradually transform real images into random noise. It then learns the reverse process: how to remove that noise step by step until a recognizable image appears. When you enter a prompt, the model begins with random noise and repeatedly removes it, gradually producing an image that matches the prompt.

The diffusion process. During training, the model learns to add noise to images. During generation, it starts with random noise and gradually removes it to create a new image.

Unlike a camera, which records light from the world, or a paint program, which places pixels directly onto a canvas, a diffusion model synthesizes an image from learned statistical relationships among objects, colors, textures, styles, and text descriptions. The model does not retrieve an existing image from memory. It generates a new one by navigating patterns learned from millions of training examples.

AssemblyAI: Diffusion Models Explained in Four Levels of Difficulty

Latent Space

Many generative AI systems operate within a latent space, a mathematical space built from vectors. A vector is simply a list of numbers that describes the features of an object. For example, an animal might be represented by values such as size, speed, habitat, diet, friendliness, or whether it is a predator. Each feature becomes one dimension of a much larger mathematical space. Real AI models may represent an image or a word using hundreds or even thousands of numerical dimensions rather than just a few.

During training, the model learns where images, words, styles, and concepts belong within this multidimensional space. Ideas that seem very different to people may end up close together because they share many underlying features. A text prompt guides the model toward a region of this latent space, where it gradually transforms noise into an image that best matches the requested concepts.

This is why image generators can combine seemingly unrelated ideas—a cathedral made of glass, a medieval manuscript depicting a spaceship, or a portrait painted in the style of an old photograph. The model is not imagining these scenes as a person would. It is navigating learned visual relationships and synthesizing an image that best matches the prompt.

Synthetic Media

AI-generated images, video, voices, music, and animation are all forms of synthetic media. These technologies support creative experimentation, rapid prototyping, accessibility, scientific visualization, and new forms of storytelling. They can also be used to create convincing misinformation, impersonate individuals, generate non-consensual imagery, or flood digital platforms with low-quality content.

Understanding diffusion models helps us see that AI images are not magic. They are the result of training data, mathematical models, prompts, and countless learned relationships among visual patterns. The next section turns from how these systems work to how people can work creatively and critically with them.

7.5 Human-AI Collaboration

AI tools are often described as replacements for human work, but they are increasingly used as collaborators that extend human thinking, creativity, and productivity. In creative, educational, and professional settings, people use AI to brainstorm ideas, revise writing, generate variations, translate concepts into different media, write code, analyze data, and create images, sound, and video.

Prompting is the practice of giving instructions or context to an AI system. A prompt can be a question, a command, a role, an example, a constraint, or a creative direction. Good prompting is less like pushing a button and more like setting up a conversation, experiment, or design process.

The worn-path image from Section 7.2 helps explain why prompting matters. A generic prompt asks the model to follow only the deepest channels of its training, often producing the most predictable and derivative results—sometimes called slop. A specific, constraint-rich, or unexpected prompt encourages the model to move beyond those well-worn paths. The resulting output is less predictable, but often more original, more focused, and more clearly shaped by human intention.

Critical Use

Human–AI collaboration requires judgment. AI systems can produce useful drafts, but they can also generate errors, stereotypes, clichés, unsupported claims, fabricated citations, and confident nonsense. The user remains responsible for evaluating, revising, contextualizing, and deciding what belongs in the final work.

Use AI to generate possibilities, not final truth.
Consider how the prompt itself shapes the response.
Ask for alternatives and compare results.
Check facts, sources, code, and citations.
Revise outputs in your own language and style.
Disclose AI assistance when required or ethically appropriate.

The most valuable AI literacy may not be knowing how to produce the fastest output. It may be knowing how to slow down, question the output, and use AI to support more thoughtful human work.

AI, Power, and Labor

AI systems are not neutral technologies. They are built by corporations, governments, universities, and researchers with particular goals, trained on data created through human labor, and deployed within economic and political systems that distribute benefits and harms unevenly. Understanding AI means understanding not only how models work, but also who builds them, who benefits from them, and who bears their costs.

Who decides: Decisions about what AI systems are built, what data they are trained on, and how they are deployed are made by a relatively small number of corporations, governments, and research institutions. Those decisions increasingly affect everyone.

Data labor: Workers who label training data—identifying objects in images, flagging harmful content, and transcribing audio—are often low-paid contractors performing repetitive and sometimes psychologically difficult work. Their labor makes AI systems appear seamless.

Environmental cost: Training and operating large AI models requires enormous amounts of electricity and water. The environmental costs are often borne by communities located near data centers and energy infrastructure.

Minerals and supply chains: AI hardware depends on lithium, cobalt, rare earth elements, and other minerals extracted through global supply chains that carry significant environmental and labor impacts.

Labor markets: AI is reshaping work. Some jobs are automated, while others increasingly involve supervising, prompting, checking, or correcting AI systems. As with earlier technological revolutions, new opportunities emerge even as existing forms of work change.

Kate Crawford and Vladan Joler: The Hidden Costs of AI (Museum of Contemporary Art Australia, 2025)

"AI is neither artificial nor intelligent."
— Kate Crawford, Atlas of AI (2021)

AI changes what people can create, but it does not eliminate the need for human intention, judgment, responsibility, or imagination. These technologies expand our capabilities, but the purposes we choose—and the values we bring to their use—remain human decisions.

7.6 Unit Exercise: Human–AI Collaboration

Generative AI can produce text, images, code, sound, and video in seconds. But speed is not the same as creativity, understanding, or judgment. In this exercise you will investigate both the strengths and the limitations of AI by comparing your own ideas with those produced through collaboration with an AI system.

Part 1. Create Without AI

Choose one of the following:

Write a short paragraph or story.
Create a sketch or simple illustration.
Design a poster or logo.
Write a poem.
Develop a concept for a short video.

Spend 15–20 minutes working without AI. Save this first version.

Part 2. Collaborate with AI

Use one or more AI tools to explore the same idea.
Begin with a simple prompt and save the first output.
Revise your prompt at least three times by adding constraints, references, examples, or stylistic directions.
Experiment with changing tone, audience, medium, or perspective.
Choose the strongest AI result.
Edit, rewrite, remix, or redesign the output yourself rather than accepting it unchanged.

Part 3. Compare

Compare your original work with your AI-assisted version.

What ideas came from you?
What ideas came from the AI?
What became stronger through collaboration?
What became weaker or more generic?
Where did you disagree with the AI?
What required your own judgment?

Reflection

Write a one-page reflection that considers both the benefits and limitations of generative AI.

Address several of the following questions:

Did AI surprise you?
Did it reinforce familiar ideas or generate genuinely new ones?
When did the AI become repetitive, cliché, or confidently incorrect?
How much of the final work feels like your own?
Did the AI help you think differently, or simply work faster?
Could you explain and defend every decision in your final piece without referring to the AI?
What skills remain uniquely human?

Your submission should include:

Your original (non-AI) work
Your prompt history
The AI outputs
Your revised final version
Your written reflection

Discussion Questions

What kinds of creativity can AI imitate, and what kinds remain difficult for it?
When does AI act as a useful collaborator, and when does it become a substitute for thinking?
How do prompts shape—not merely request—the output?
How should AI-generated work be credited or disclosed in school, journalism, research, and the arts?
Who benefits economically from generative AI, and who bears its environmental and labor costs?
If AI becomes better at generating media, what human abilities become more valuable rather than less?

7.7 Glossary

artificial intelligence: computer systems designed to perform tasks associated with intelligence, such as recognizing patterns, generating media, making predictions, or supporting decisions.
symbolic ai: an approach to artificial intelligence based on explicit rules, logic, and symbolic representations.
machine learning: a method of creating computer systems that learn patterns from data rather than following only fixed instructions.
model: a trained AI system that can produce outputs based on patterns learned from data.
training data: the examples used to teach a machine learning system how to recognize patterns or generate outputs.
neural network: a machine learning system made of connected layers that transform input data into output.
weight: a number stored inside a neural network that represents the strength of a connection between two parts of the model. Training adjusts millions or billions of weights to encode the patterns the model will use.
backpropagation: the process of working backward through a neural network to adjust its weights after each training example. Repeated many times, backpropagation is how a model carves out the patterns it will later use to generate output.
large language model: an AI model trained on large collections of text to generate and respond to language.
token: a small unit of text, such as a word, part of a word, or punctuation mark, used by language models.
next-token prediction: the process of generating language by predicting the next likely token in a sequence.
transformer: a machine learning architecture used by many contemporary language models to process relationships across sequences of text.
attention: a process that helps an AI model weigh relationships among parts of an input, such as words in a sentence.
diffusion model: a generative AI system that creates images by learning how to transform noise into visual patterns.
latent space: a compressed mathematical space where relationships among data, images, words, or concepts can be represented.
synthetic media: media generated or significantly altered by computational systems, including AI-generated images, video, voices, music, or text.
prompting: the practice of giving instructions, context, examples, or constraints to an AI system.
hallucination: an AI output that appears plausible but is inaccurate, unsupported, or invented.
bias: a pattern of unfairness or distortion that may appear in an AI system because of training data, design choices, or social inequalities.
generative ai: AI systems that create new text, images, audio, video, code, or other media from prompts or input data.

7.8 Bibliography

Boden, Margaret A. AI: Its Nature and Future. Oxford University Press, 2016.

Broussard, Meredith. Artificial Unintelligence: How Computers Misunderstand the World. MIT Press, 2018.

Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021.

Mitchell, Melanie. Artificial Intelligence: A Guide for Thinking Humans. Farrar, Straus and Giroux, 2019.

Noble, Safiya Umoja. Algorithms of Oppression: How Search Engines Reinforce Racism. NYU Press, 2018.

Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019.

Christian, Brian. The Alignment Problem: Machine Learning and Human Values. W. W. Norton, 2020.

Li, Fei-Fei. The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI. Flatiron Books, 2023.

Mollick, Ethan. Co-Intelligence: Living and Working with AI. Portfolio, 2024.

Pasquinelli, Matteo. The Eye of the Master: A Social History of Artificial Intelligence. Verso, 2023.