In 1839, when Louis Daguerre announced the daguerreotype to the French Academy of Sciences, the painter Paul Delaroche reportedly declared that painting was dead. It wasn't. But something did die: the assumption that capturing a likeness required a human hand. Within a decade, portrait studios had opened in every major city, and working painters faced a genuine economic threat — particularly those who made their living on commissions. What followed wasn't extinction but reorganization. Painters moved toward abstraction, symbolism, and the inner life of subjects, territory photography couldn't yet reach. The technology didn't end creativity; it relocated it.
The same pattern surfaced in 1997 when Garry Kasparov lost to IBM's Deep Blue. Chess commentators predicted the game would lose its audience once a machine could beat any human. Instead, chess viewership expanded dramatically — partly because AI analysis tools made the game's depth newly visible to casual fans, and partly because human players began training against engines to find moves no grandmaster had considered. The threat became the curriculum. By 2023, the best human players were still playing — and still drawing audiences — in part because of what they had learned from machines.
This course is about that same inflection point, arriving now in writing, music, visual art, design, and code. AI tools in 2024 — Midjourney, Suno, Claude, Runway, GitHub Copilot — are not finishing the conversation about creativity; they are changing who gets to start one. This course won't pretend the changes are small or that every creative tradition survives unchanged. It will show you specifically what these tools can and cannot do, where human judgment remains irreplaceable, and how to build a genuine working practice around them rather than just a set of prompts.
In January 2023, Nick Cave received a fan-generated lyric written by ChatGPT "in the style of Nick Cave." Cave's response, published on his Red Hand Files newsletter, became one of the most widely cited artistic statements of that year. He didn't call the lyric bad. He called it soulless — not because the words were wrong but because they had cost nothing. "Songs arise out of suffering," he wrote, "by which I mean they are predicated on the complex, internal human struggle of creation." Cave's objection wasn't to AI existing. It was to the confusion about what a song is and where it comes from.
That same month, Holly Herndon — the composer who had spent years building AI models trained on her own voice — argued the opposite position. Herndon had released proto in 2019, an album made collaboratively with an AI she named Spawn, and she was already licensing her voice model to other artists. For Herndon, the question was never whether a machine was involved. It was whether a specific human's choices, taste, and creative history shaped the outcome. Her work answered yes. Cave's fan's work answered less clearly.
The tension between these two positions defines the field right now. Co-creation with AI is not a single thing. It is a spectrum — from using AI as a spell-checker at one end, to having AI generate 90% of a finished work at the other. Where you sit on that spectrum, and why, determines what the work means and who is responsible for it.
Researchers at Stanford and MIT have been studying human-AI creative collaboration since at least 2019, and one consistent finding is that most people have no coherent model of what they are actually doing when they use a generative tool. They either over-attribute the output to themselves ("I made this") or under-attribute it ("the AI made this") when the reality is almost always somewhere in between and depends heavily on the specific decisions made during the process.
A useful frame comes from the work of Margaret Boden, the cognitive scientist who has studied computational creativity for four decades. Boden distinguishes three types of creative contribution: combinatorial (recombining existing ideas in new ways), exploratory (pushing the boundaries of an existing conceptual space), and transformational (changing the rules of the space itself). Current generative AI is genuinely powerful at combinatorial creativity. It is occasionally capable of something resembling exploratory creativity. It has not demonstrated transformational creativity — the capacity to fundamentally redefine what a domain is.
This matters practically. If you are using an AI to generate image variants from a concept you defined, the combinatorial work is shared. The exploratory work — deciding which variant is right, why it fits the larger project, what it says — is yours. The transformational work, if there is any, is entirely yours. Understanding which layer you are working at helps you make better decisions and explain your process honestly.
In April 2023, the Recording Academy updated its Grammy eligibility rules to state that AI-generated music is not eligible for awards — but music that uses AI as a tool, with a human as the primary creative force, may be. The distinction they are trying to draw, imperfectly, is the same one Boden's framework illuminates: who is doing the exploratory and transformational work?
The word prompt has become shorthand for the entire human contribution to AI-generated work, which is a significant distortion. A prompt is an input, but creative agency is not just the moment of input. It is the full sequence of decisions: what problem to solve, what constraints to impose, which outputs to reject, how to combine and edit results, what context the work enters, and what claim the creator makes about the relationship between their intentions and the final object.
When photographer Boris Eldagsen submitted an AI-generated image to the Sony World Photography Awards in April 2023 and won — then declined the prize and publicly revealed it was AI-generated — he was making an argument about precisely this. His point was not that AI images are bad. It was that a photography competition implies a specific kind of human agency (the photographer's eye, their physical presence in a moment, their technical command of a camera) that his winning image did not involve. The competition had no framework for the distinction. He forced one.
The lesson for working with AI is direct: know what kind of agency you are claiming, and make sure your process actually produces it. If you are calling yourself the author of something, be able to articulate what decisions you made, what you rejected, and why the final version reflects your judgment rather than a default.
Co-creation is not a binary. It is a set of choices you make at every stage of a project. The aim of this course is to give you enough fluency with specific tools — and enough conceptual clarity about what you are doing with them — that your choices are actually choices, not defaults.
You'll use the AI below to examine a real creative decision you have made — or are making — and map where your agency actually lives in that process. This isn't about defending your choices; it's about seeing them clearly.
Think of a recent creative project (or one you're planning): a piece of writing, a design, a photograph, a piece of music, a presentation, anything. You'll walk the AI through your process and it will help you identify where the combinatorial, exploratory, and transformational layers of decision-making occurred — and what that tells you about your authorship of the work.
In March 2023, a team at Runway ML ran an informal study comparing outputs from novice and experienced users of their Gen-2 video generation tool. The prompts were structurally similar in length and vocabulary. The outputs were dramatically different in quality. The experienced users weren't using secret syntax. They were specifying mood, camera angle, lighting, pacing, and relationship between elements in ways that reflected genuine visual fluency. The AI wasn't performing better — it was receiving better instructions. The skill gap had migrated, not disappeared.
This pattern appears consistently across generative domains. The 2023 paper "Large Language Models as Creative Writing Teachers" from researchers at the University of Edinburgh found that prompt quality correlated strongly with the author's pre-existing domain knowledge — writers who understood structure, voice, and scene construction wrote better prompts and got better outputs, even when they were new to AI tools specifically. The AI amplified existing skill more than it substituted for missing skill.
A prompt is a compressed creative brief. The same principles that make a good creative brief make a good prompt: clarity of objective, specificity of constraints, articulation of audience and tone, and explicit statement of what success looks like. The difference is that a prompt is also a communication to a system that has no shared context with you — it cannot ask clarifying questions the way a human collaborator can, and it will fill ambiguity with statistical likelihood rather than creative judgment.
This means that anything you leave unspecified gets decided by the model's training data — which reflects the average of what it has seen, not your specific intentions. Ambiguity in a prompt is not creative freedom; it is creative abdication. The model will make choices, but they will not be your choices.
The most productive mental model is to think of prompting as simultaneous translation: you are translating a mental image, feeling, or argument into language precise enough that a system with no access to your interior experience can approximate it. The translation skill is where your craft lives. The generation is what the AI does with the translation.
In October 2022, when Midjourney v4 launched, the artist community began documenting what became known as "prompt engineering" as a serious skill. Users like Andrei Kovalev began publishing annotated prompt breakdowns showing how specifying lighting conditions ("golden hour," "Rembrandt lighting"), aspect ratios, artist references, and mood terms transformed outputs from generic to intentional. These weren't tricks — they were craft vocabulary applied to a new instrument.
Constraints are not limitations on creativity; they are its structure. Sonnets are not worse than free verse because they have a fixed form — they are different, and the constraint creates the pressure that makes certain kinds of meaning possible. The same principle applies to prompting. Explicit constraints — style restrictions, format requirements, things to exclude — produce more controllable and more interesting outputs than unconstrained requests.
Negative prompting (explicitly telling a system what not to produce) is as important as positive specification. In image generation, the negative prompt space controls quality artifacts. In language model prompting, framing what you don't want ("do not use bullet points," "avoid clichéd metaphors," "do not summarize — analyze") shapes the response at least as much as the positive request.
The iterative dimension of prompting is also underused. A single prompt is rarely the best version of a prompt. Experienced users treat the first output as diagnostic — it reveals what the model understood, where it defaulted, what it emphasized. The second prompt is a correction, not a restart.
Prompt craft is the new surface where domain expertise becomes visible. You don't need to know code. You do need to know what you want, why you want it, and how to say it precisely enough that a system with no access to your interior life can approximate it.
You'll practice the anatomy of a prompt by working with the AI to build, diagnose, and refine a prompt for a real task you want to accomplish. The goal is to see the components of a good prompt clearly and understand what each one does.
Choose a creative task you actually want to do — generate an image concept, draft a piece of writing, design a structured document, compose a music brief, anything. You'll start with a rough version and systematically improve it.
In February 2023, CNET was revealed to have been quietly publishing AI-generated finance articles since November 2022. The articles were grammatically correct, well-structured, and covered real topics accurately at the surface level. The problem emerged when fact-checkers found that multiple articles contained subtle financial errors — not hallucinations of invented facts, but wrong applications of correct information. The AI had applied a tax rule accurately described in its training data to the wrong type of account. The error was not random; it was plausible. It looked right. CNET's editors, who were supposed to be reviewing the content, had missed it because plausible-looking text passes a casual reading.
The incident illuminated something precise: fluency creates the illusion of correctness. A model trained to produce coherent text produces coherent text even when the underlying reasoning is wrong. The reader's brain, which evolved to treat fluency as a signal of competence, is not naturally calibrated to detect this kind of error. This is not an argument against using AI. It is an argument about where the human's role becomes non-negotiable.
Based on documented cases of AI-assisted creative and professional work since 2020, four functions consistently require human judgment that current AI systems cannot reliably provide:
1. Consequential truth-checking. AI systems confabulate with confidence. They produce plausible-sounding claims that are wrong, and they do not know the difference. Any output that will be acted on — that carries consequences in the real world — requires a human to verify claims against primary sources, not against the model's self-assessment of its own accuracy.
2. Contextual fit. An AI model has no access to the specific relationships, history, stakes, and unspoken constraints of your actual situation. It cannot know that this particular client has had a bad experience with this specific word, that the audience for this piece has a specific cultural context the text must navigate, or that the joke that would land perfectly in one room would be disastrous in another. Contextual knowledge is held by humans who are embedded in situations.
3. Ethical and relational responsibility. When a decision affects people — when the content will hurt or help, include or exclude, elevate or diminish — a human must own the decision. Not review it, own it. The distinction matters because ownership implies that someone can be held accountable, can change their mind, and can be confronted by those affected. A model cannot be confronted.
4. Taste as a coherent system. Individual taste can be mimicked at the surface level, but a body of work has a coherent internal logic — a set of values that run through every choice, including the choices not to do things. That coherent system is a human construction accumulated over time, and it is what separates a distinctive voice from a capable imitation of one.
In a 2023 study published in the journal Cognition, researchers found that people rated AI-generated explanations as more credible when they were written in fluent prose than when the same information was presented in plain, simple language — even when the content was identical. The implication for creative and professional work is significant: fluency is a persuasion mechanism, not a quality signal. Evaluating AI output requires actively resisting the fluency effect.
One practical consequence of understanding where human judgment is irreplaceable is that editing AI output is not a secondary task. It is where the primary creative work happens. The writer Gordon Lish famously transformed Raymond Carver's manuscripts by cutting up to 70% of the text and restructuring what remained. Carver resisted and eventually published the original versions — but the Lish versions became more celebrated precisely because they embodied a sharper vision. The editing was the creative act.
When you edit AI-generated work, you are performing the same function: imposing a coherent vision on raw material. The material happens to come from a generative model rather than from a first draft. The editorial work — cutting what is merely plausible, keeping what is specifically right, restructuring around a genuine argument — is still entirely yours.
This frame matters practically because it recasts what AI-assisted work looks like as a process. The generation phase is fast and cheap. The editorial phase is slow and expensive and requires everything that makes you specifically you. Inverting the time allocation — spending most of your time in editorial rather than generation — produces better work and a more honest creative relationship with the tools.
Use AI for generation. Use yourself for judgment. The clearer you are about which phase you're in at any given moment, the less likely you are to confuse fluency with quality or speed with good work.
You'll practice evaluating AI-generated content for the fluency trap — learning to distinguish between text that sounds right and text that is right. This is an active resistance exercise against the cognitive bias the lesson described.
You can bring a piece of AI-generated content you've produced or received, or ask the lab AI to generate something in a domain you know well so you can audit it properly. The goal is to practice the editorial eye.
In 2023, the architecture firm Bjarke Ingels Group (BIG) began publicly discussing their use of Midjourney and other generative tools in early-stage design exploration. Principal Kai-Uwe Bergmann described the shift precisely: the tools had not replaced their designers, but they had changed which conversations happened earlier. Concepts that previously required weeks of sketching to communicate were now visualized in hours, which moved the critical design discussion — the conversation about what a building should mean and how it should relate to its context — to week one instead of week six. The quality of the final work depended on whether the team treated that earlier conversation as more important, not less, than it had been before. The risk, Bergmann noted, was the opposite tendency: to keep generating variations rather than committing to a direction, because generation had become cheap.
This dynamic — generation becoming cheap while judgment remains expensive — is the defining structural condition of AI-assisted creative work. The teams and individuals who use it well have developed explicit practices for managing that asymmetry: deliberate protocols for when to generate and when to stop, for how to evaluate options, and for when to override the most impressive-looking output in favor of the specifically right one.
When generation is cheap, the default mode becomes generating more. More options, more variations, more iterations — because they are available and because choosing feels harder than producing. This is a version of what psychologists call the paradox of choice, and it appears consistently in documented AI-assisted design, writing, and music production workflows.
The practical consequence is decision fatigue without a corresponding increase in quality. Research by Stanford professor Bob Sutton on innovation processes consistently shows that constraint increases creativity more reliably than expanded option sets. The professional tool is knowing when to close the generation loop and enter the judgment loop — not as a concession but as a deliberate move.
Experienced AI users report setting explicit generation limits: a maximum number of image variants before committing, a maximum number of draft paragraphs before editing, a fixed prompt-refinement cycle before accepting or abandoning a direction. These are not arbitrary restrictions. They are structural defenses against the generativity trap.
Musician and producer Grimes released her AI vocal model in April 2023, inviting other artists to use it on the condition that they split royalties 50/50. The decision was controversial, but the underlying structure was coherent: a defined relationship between the human contribution (the model trained on her voice, her identity, her aesthetic context) and the AI contribution (the generated vocals). It was a practice decision, not just a tool decision. It defined the terms of authorship in advance.
A working practice with AI has the same components as any professional creative practice: a defined scope of what the tools are for and not for, a set of quality standards that the tools must meet before their outputs advance, a feedback loop for improving your own prompting and editorial skill over time, and an honest account of the relationship between what you make and what the AI generates.
The scope question is worth taking seriously. Many creators use AI for everything — because it's available for everything. A more deliberate approach is to identify the specific phases of your work where AI assistance produces real leverage: where generation is genuinely faster than hand-drafting, where having more options earlier genuinely improves the final decision, where the AI's range exceeds your own. Then use it specifically there and not elsewhere.
The quality-standards question is equally important. AI outputs should meet the same bar as any other raw material you work with. A generated image that is technically impressive but generically composed should be treated the same way you would treat a stock photo that is technically competent but has no specific relationship to your project: as a starting point at best, probably as a source of components rather than a final asset.
The goal of this course is not to give you access to more AI tools. It is to give you a deliberate working relationship with the tools you already have or will encounter — one where you know what you are using them for, where your judgment is doing the work, and what the result says about you as a creator.
You'll use this session to draft a genuine working protocol for AI use in your own creative or professional context. Not a set of rules imposed from outside — a set of commitments you actually believe will produce better work for you specifically.
This is the most personal and high-stakes lab in the module. The output is a short document — three to five points — that defines your scope, your quality bar, your generation limits, and your authorship position. You don't need to share it with anyone. But you should mean it.