In 1888, George Eastman put a camera called the Kodak on sale for twenty-five dollars and changed the meaning of the word "photographer." Before that moment, making a photograph required a studio, chemicals, glass plates, and years of technical apprenticeship. After it, anyone who could press a button could capture an image. Professional photographers warned that the craft would be devalued, that the masses would produce nothing but mediocrity. They were partly right and mostly wrong: the overall volume of creative output exploded, new genres emerged that the professionals had never imagined, and a century later the photographic tradition is considered richer, not poorer, for the democratization.
Something structurally similar is happening right now with language, code, music, and image. In November 2022, OpenAI released ChatGPT to the public; within two months it had a hundred million users — the fastest consumer-product adoption ever recorded. By mid-2023, Adobe had embedded generative AI into Photoshop, Spotify was hosting AI-voiced podcast translations, and publishers were debating what "authorship" even meant. The tools are not finished, the legal frameworks are not settled, and the creative questions are genuinely open.
This course is a working introduction to AI as a creative collaborator. You will learn how to write prompts that actually produce what you intend, how to iterate when the first result is wrong, how to use AI for visual and musical ideas, and how to think clearly about what remains yours when a machine helped. It will not make every creative problem easy. It will give you a real starting vocabulary and a set of practiced instincts — so that whatever these tools become next, you are not starting from zero.
In January 2023, the novelist Kazuo Ishiguro was asked in a BBC interview whether AI could write a novel. He said he doubted it — not because the sentences would be bad, but because a novel is made of things the author genuinely does not know yet, things that surprise the writer during the writing. The machine, he argued, does not not-know anything. It produces the statistically expected next word. That is a different activity than discovery.
Ishiguro is probably right about the deepest layer. But there is a large territory between "producing statistically expected words" and "writing a novel as an act of self-discovery" — and that territory is where most actual creative work happens. Drafting. Generating options. Breaking blocks. Finding a phrase you then improve. This is where AI tools are genuinely useful right now, and this is where we will spend most of our time.
The critical insight is that the quality of what comes out is tightly coupled to the quality of what goes in. A vague prompt produces vague output. A specific, contextually rich prompt produces something you can actually use — or at least usefully react to. Learning to write that prompt is the skill this course is built around.
Large language models — GPT-4, Claude, Gemini, Llama — were trained on enormous text datasets by a process that, simplified, involves predicting which token comes next in a sequence. After enough of that training, the model develops internal representations that allow it to generalise: it can complete sentences it has never seen, adopt registers it was not explicitly taught, and reason about topics by analogy to patterns it absorbed. This is not the same as understanding in the human sense, but it is also not simple autocomplete.
Image generators like Midjourney, DALL·E 3, and Stable Diffusion work differently: they learn to map between text descriptions and image features through a process called diffusion — starting from noise and iteratively removing it according to what the prompt implies. Music generators like Suno (launched publicly in 2023) and Udio use related architectures applied to audio tokens.
You do not need to understand the mathematics to use these tools well. But you do need to understand one practical implication: the model has no intention. It does not want to help you write a good story. It is producing the most probable continuation of the context you gave it. Your job, as the human, is to construct a context — a prompt — that makes the probable continuation also the useful one.
Researchers at Anthropic, OpenAI, and independent prompt-engineering teams have converged on a rough taxonomy of what makes prompts work. The following components are not rules — they are levers. You choose which to pull for a given task.
In 2023, Stanford researchers studying GitHub Copilot (the AI coding assistant used by over a million developers) found that programmers who wrote more specific, contextual prompts accepted AI suggestions at twice the rate of those who wrote brief prompts — and the accepted suggestions required significantly less editing. The pattern holds in writing and design tasks too. Specificity is the primary quality lever.
Professional prompt engineers do not write one prompt and expect a final result. They treat the first output as a draft — material to react to. The discipline is: run, evaluate, diagnose, refine. If the output is too formal, add "conversational and direct" to the format instruction. If it missed a key point, add that point to the context. If the structure is wrong, show an example of the right structure.
This iterative loop is the closest analogy to how editors work with writers, or how directors work with actors. The AI is not the author. You are. The AI is the person in the room who will try anything you ask without complaint, at high speed. Your judgment about what works — your taste, your knowledge of your audience, your sense of your own voice — is the irreplaceable ingredient.
The practical implication: do not judge AI creative tools by their first output. Judge them by what you can coax from them across three or four iterations. Most beginners quit too early. Most experienced users have a specific diagnostic vocabulary — they know which lever to pull when the output is flat, when it is generic, when it is structurally wrong — and that vocabulary is learnable.
You are not learning to operate a machine. You are learning to direct a very fast, very capable collaborator who has no taste, no intention, and no investment in the outcome. The directing skill — specific, contextual, iterative — is what this module is about. Everything else is application.
Ishiguro's point deserves honesty. AI tools do not have genuine aesthetic preferences. They do not know what it feels like to be bored at 2 a.m. trying to find the right word, which is part of where the right word comes from. They do not have the lived experience that gives creative work its specificity and authority.
More practically: they hallucinate. They produce confident-sounding text that is factually wrong. They average across their training data, which means they trend toward the expected. They do not know what happened yesterday. Their outputs on contested cultural or personal topics can be generic to the point of uselessness.
Knowing these limits is not pessimism — it is calibration. A surgeon who knows what a scalpel cannot do is a better surgeon. The limits define where your judgment is not optional: verifying claims, bringing specificity from your own knowledge, and making the final call about whether something is actually good. These remain entirely human responsibilities.
You are going to write a prompt for a real creative task — then refine it at least twice based on feedback. The AI assistant here knows the six-component prompt framework from Lesson 1 and will help you diagnose why a prompt works or doesn't, suggest which lever to pull, and respond to your drafts as if it were the creative AI you are prompting.
Start by describing a creative thing you actually want to make: a piece of writing, a concept for an image, a title, a product description, anything. Then build your prompt using the framework components. The assistant will both evaluate your prompt structure and show you what the output would look like.
In March 2023, a team at Columbia Journalism Review analyzed a sample of AI-generated press releases submitted to newsrooms. They found a distinctive set of phrases appearing at statistically anomalous rates: "delve," "leverage," "cutting-edge," "in today's rapidly evolving landscape," "it is important to note." These phrases had not been common in human business writing before 2022. They were emerging as AI fingerprints — the model's statistical centre of gravity made visible.
The same pattern shows up in student essays, marketing copy, and social media posts. When a model averages across millions of texts, it tends toward the median — the phrasing that appears in the widest variety of contexts. That median is not bad writing exactly. It is competent, inoffensive, and utterly unmemorable. Your job is to push the output away from the median and toward something that sounds like you knew something specific and cared enough to say it precisely.
The reason AI output tends toward blandness is that voice, in writing, is largely made of specific choices: a preference for short sentences or long ones, a habit of starting paragraphs with questions, a tendency to anchor abstractions in particular kinds of concrete detail, a characteristic relationship to humour or irony. These choices are not arbitrary — they are the accumulated trace of a person's reading, thinking, and life. The model doesn't have that trace. It has the average of everyone else's traces.
The practical solution is to treat voice as information you supply in the prompt. This means: giving the model examples of your own writing, describing your preferences explicitly ("I write in short declarative sentences and avoid adverbs"), and using the model's output as a raw material that you then rewrite rather than publish directly.
The rewrite step is not optional if you care about voice. The most effective workflow used by professional writers who use AI — including those at publications like The Atlantic, which discussed AI-assisted workflows openly in 2023 — is to use the model to generate structure and options, then rewrite the passages that will be read closely in their own voice.
These are not theoretical — they are practiced techniques with documented effectiveness:
The novelist Gary Shteyngart wrote about using ChatGPT for his 2023 novel "Our Country Friends" promotional work in The New Yorker. He found it useful for generating options he then heavily edited, but noted that the model consistently made his voice "warmer and less strange" — flattening exactly the qualities that make his work distinctive. His solution was to use AI for structural drafts and then aggressively rewrite the surface text. This is now a common professional pattern.
Not every writing task requires a distinctive voice. Terms and conditions do not need your personality. A first draft of a meeting agenda does not need to sound like you at your literary best. A summary of a document for internal use can be competently generic without anyone losing anything.
Part of the skill is knowing which category you're in. Voice-critical writing — things your audience will actually read and remember, things attached to your reputation, things meant to build a relationship with a reader — requires the extra step of voice preservation. Functional writing — structural scaffolds, first drafts for editing, administrative documents — can use AI output more directly.
Making this distinction deliberately, rather than applying the same workflow to everything, is itself a sign of creative maturity with the tools.
Your voice is made of specific knowledge and specific choices. The model has neither. Voice preservation is not about fighting the tool — it is about supplying the information the tool lacks, and then doing the rewrite work that no tool can do for you. The result is faster than writing from scratch and more distinctly yours than accepting the first output.
In this lab you will practice identifying and correcting AI voice problems. The assistant will generate a deliberately generic AI-style paragraph on a topic you choose, then help you apply one or more voice-preservation techniques to transform it into something that sounds specific and human.
You can also bring your own AI-generated text that you think sounds too generic, and the assistant will help you diagnose which techniques would help most.
In August 2022, an image generated by Jason Allen using Midjourney won first place in the digital art category at the Colorado State Fair. The image, "Théâtre D'Opéra Spatial," showed figures in elaborate period costume before a circular portal of light. It beat human artists. Allen had entered it under the name "Jason Allen via Midjourney." The judges did not know it was AI-generated. When they found out, the debate that followed was not primarily about Allen's ethics — it was about what "making" an image means when the making is linguistic: you describe, the model renders.
The uncomfortable fact that the controversy surfaced was this: Allen had spent hours refining his prompts. He was not typing one sentence and accepting the result. He was iterating across dozens of generations, selecting, recombining, adjusting aspect ratios and style references and lighting descriptors. The creative labour was real — it was just invisible, and it was expressed in language rather than in brush strokes or code.
Models like Midjourney, DALL·E 3, and Stable Diffusion are trained on enormous datasets of image-text pairs: a photograph of a red barn at dusk paired with the caption "red barn at dusk, golden hour, pastoral landscape." Through billions of such pairings, the model learns which visual features correspond to which descriptive concepts.
At generation time, the model starts with random noise and iteratively denoises it — removing randomness in the direction that makes the image more consistent with the prompt. The number of denoising steps, the degree of prompt adherence (called "guidance scale" or "cfg"), and the random starting seed all affect the output.
Practically this means: your prompt is a set of instructions to a denoising process. Concepts that appear early in the prompt and are weighted heavily (in Midjourney, by repetition or the :: weighting syntax) have more influence on the output. Vague abstract concepts ("beautiful," "interesting") have less effect than concrete visual descriptions ("diffuse afternoon light coming from camera-left, long shadows").
Adobe's internal research shared at their MAX conference in October 2023 showed that Firefly users who included lighting descriptors and style references in their prompts rated their outputs significantly more satisfactory than those who used subject-only prompts — and required fewer regenerations. Lighting and style reference are the highest-leverage additions to an image prompt.
Tools like Suno (released in public beta December 2023) and Udio (launched April 2024) generate music from text prompts. The prompting grammar is different from image generation but the underlying principle is identical: specific descriptors outperform vague ones, and you are guiding a probabilistic system toward a region of its training space.
Effective music prompts typically include: genre ("neo-soul," "Appalachian folk," "Berlin-style techno"), instrumentation ("fingerpicked acoustic guitar, upright bass, no drums"), tempo and feel ("slow, rubato, contemplative"), mood or reference ("melancholic, reminiscent of late-period Talk Talk"), and if using lyric generation, subject matter and emotional arc.
The key limit to understand with current music AI: it is very good at genre and texture, less reliable at structure (verse-chorus-verse following a specific emotional arc), and essentially cannot yet be given a melodic line to harmonise in a controlled way. Understanding this limit tells you where to use it — texture and mood exploration — and where not to — precise compositional structure.
Whether you are prompting for text, images, or audio, the underlying discipline is the same: translate your intention into specific, concrete, technically-informed language that narrows the model's probability space toward the region you actually want. The vocabulary differs by medium. The logic is identical.
In this lab you will build an image prompt for a visual idea you have, using the six visual components from Lesson 3: Subject, Environment, Lighting, Style Reference, Composition/Camera, and Negative Prompts. The assistant will evaluate your prompt, identify which components are missing or weak, and help you build a stronger version — and then show you what the refined prompt would likely produce.
You can also bring an image prompt that didn't work and we'll diagnose it together.
In February 2023, the US Copyright Office issued a registration for Kris Kashtanova's graphic novel "Zarya of the Dawn" — and then immediately walked it back. The Office reexamined the work after public attention revealed that its images had been generated by Midjourney. The final ruling: the text and arrangement of the work could be copyrighted (Kashtanova wrote it and made structural choices), but the individual AI-generated images could not, because copyright requires human authorship of the specific expressive element. In March 2023, the Copyright Office issued further guidance: AI-generated material produced "without creative control" is not copyrightable. Material where a human made "sufficient creative choices" in selection, arrangement, and modification may be.
The legal landscape is still forming — multiple court cases are active as of 2024 — but this ruling established the operational distinction that matters for creators: the process of prompting and selecting is where your authorship lives, and the question is whether that process is sufficient to constitute the "creative control" the law requires.
The legal and ethical questions around AI authorship do not have clean universal answers yet. But there are three practical questions that help clarify your position in any specific case:
Different fields have moved at different speeds. The Associated Press issued AI guidelines in 2023 requiring disclosure whenever AI generated any part of a published work. The National Eating Disorders Association made news for replacing human staff with an AI chatbot in 2023, then reversed course after public criticism — a case study in how disclosure failures compound reputational damage. The Academy of Motion Picture Arts and Sciences ruled in late 2023 that AI-assisted films are eligible for Oscar consideration, but that human creativity must be "predominant."
The working principle that emerges across fields: disclose more rather than less. The cost of over-disclosure is low. The cost of discovered concealment — particularly as detection tools improve — is high. This is not primarily a legal calculation; it is a trust calculation. Your creative reputation rests on what people believe about how you work.
An ongoing and unresolved issue: the models were trained on copyrighted material without license in most cases. Getty Images sued Stability AI in January 2023; the New York Times sued OpenAI in December 2023; multiple class-action suits from authors are active. Using these tools does not make you party to those lawsuits, but it does mean the tools exist in contested legal territory. This is worth knowing as a factual matter, separate from whether you choose to use them.
The most useful frame for thinking about authorship when AI is involved may be the one already used for other collaborative situations. When a director makes a film, they did not write the score, operate the camera, or design the costumes — but no one questions whether the film is "theirs." The director's contribution is conceptual, selective, and coordinative. AI-assisted creative work can sit in a similar relationship: the human provides the concept, the taste, the selective judgment, the editorial eye, and the final decision about what is good enough to share.
What this means practically: the more you treat AI as a tool in service of a vision that is clearly yours — where your specific choices, your knowledge of your audience, your aesthetic judgment are the determining factors in the output — the stronger your claim to the creative work, legally, ethically, and in terms of what you can honestly say about how you made it.
The blank page was always a collaboration — with language, with tradition, with influence, with the tools of the moment. AI is a new kind of collaborator, faster and stranger than any previous one. The questions it raises about authorship are not new in kind, only in urgency. You already know how to make creative work that is yours. The discipline is applying that same judgment to a faster, weirder process.
The blank page just got a partner. That partner is fast, tireless, and has no taste. What it produces becomes yours when your specific choices, your knowledge, and your judgment are the deciding factors in what the world finally sees. That responsibility has not changed. Only the speed of the process around it has.
In this lab you will work through an authorship audit of a real or hypothetical AI-assisted creative project. The assistant will ask you structured questions about your creative choices — what you decided, what you selected, what you changed — and help you build a clear account of where your authorship sits and how you would disclose AI's role accurately in your specific context.
This is also where you can bring harder questions: what to say in an artist's statement, how to handle a client who asks about AI, what disclosure looks like in your field.