In 1888, when George Eastman introduced the Kodak No. 1 camera at $25 and the slogan "You press the button, we do the rest," professional portrait photographers were furious. They had spent years mastering wet-plate collodion chemistry, exposure calculations, and darkroom technique. Here was a device that handed the act of image-making to anyone willing to wind a key. The Photographic Society of London debated whether photography done by machine could even be called art. Within two decades the argument had dissolved — not because the camera won, but because photographers discovered that the machine freed them from the mechanics of capture so they could concentrate entirely on light, composition, and meaning. Alfred Stieglitz was making photographs that hung in the Metropolitan Museum by 1910.
That pattern — a tool arrives, a profession panics, the tool gets absorbed and the work deepens — has repeated with the electric guitar in the 1930s, magnetic tape editing in the 1950s, the synthesizer in the 1970s, desktop publishing in the 1980s, and digital audio workstations in the 1990s. Each time, the objection was that the machine was doing what the artist should be doing. Each time, artists who engaged with the tool early set the terms for everyone who followed. Brian Eno built an entire compositional philosophy around the limitations of the early synthesizer. Trent Reznor used Pro Tools to make Nine Inch Nails records that could not have existed on tape.
This course is about AI in creative work — writing, visual art, music, design, and the creative decisions embedded in almost every professional role. We will not pretend AI is neutral, limitless, or simple. We will examine what it actually does technically, where it genuinely helps, where it reliably fails, and how working practitioners are integrating it today. By the end you will have a clear framework for deciding when to reach for an AI tool and when to put it down — and you will have practiced both.
If you finish every module, here's who you become:
In January 2023, The Atlantic published a piece by staff writer Stephen Marche documenting his experiment using GPT-4 to co-write a short story. Marche, a novelist, spent three weeks feeding the model fragments, revising its output, and steering it away from clichés. His conclusion was precise: the model was "an extraordinarily sophisticated autocomplete." It never surprised him the way a collaborator surprises you — by bringing something it had genuinely experienced. But it also never got tired, never needed encouragement, and never forgot a constraint he'd specified. What he had was not a co-author. It was something harder to name: a system that had processed more narrative structure than any human could read in a lifetime, and that surfaced patterns from that structure on demand.
That description — patterns surfaced on demand — is the most accurate plain-English account of what a large language model does. The technical name is next-token prediction. The creative implications take a full course to unpack. This lesson starts with the mechanism.
A large language model (LLM) is trained on a massive corpus of text — in GPT-4's case, an estimated several hundred billion words drawn from books, websites, academic papers, code repositories, and digitized human writing going back decades. During training, the model learns one thing obsessively: given a sequence of words (or sub-word tokens), what token is most likely to come next?
This is not storage and retrieval. The model does not keep a copy of its training data. Instead it compresses statistical relationships between tokens into billions of numerical weights distributed across a neural network architecture called a transformer. When you prompt the model, it uses those weights to compute a probability distribution over its entire vocabulary — roughly 50,000 tokens for most current models — and samples from that distribution. It then appends the chosen token and repeats the process. A 500-word response involves approximately 650 individual prediction steps.
What this means for creative work: the model is not imagining, intending, or evaluating. It is sampling from learned distributions. High-probability outputs tend to be familiar and fluent. Low-probability samples (controlled by a parameter called "temperature") can be surprising but are also more likely to be incoherent. The creative practitioner's job is to understand this mechanism well enough to work with its tendencies rather than against them.
The "temperature" parameter — available in most AI APIs and some consumer interfaces — controls how sharply the probability distribution is peaked. Temperature 0 always picks the highest-probability token (deterministic, repetitive). Temperature 1 samples proportionally (default behavior). Temperature above 1 flattens the distribution, increasing randomness. Most creative writing tasks benefit from temperatures between 0.7 and 1.1.
Text models predict the next token. Image generation models — Midjourney, DALL-E 3, Stable Diffusion — use a related but distinct mechanism called diffusion. During training, the model learns to reverse a process of adding random noise to images. At inference time, it starts from pure noise and iteratively denoises, guided by a text prompt, until a coherent image emerges. The guidance comes from a text encoder (often CLIP, developed by OpenAI and published in January 2021) that has learned relationships between text descriptions and image features across hundreds of millions of image-caption pairs.
Audio models like Suno (launched publicly in 2023) and MusicLM (Google, published February 2023) apply similar principles to spectrograms — visual representations of sound. The model learns what combinations of frequency, rhythm, and timbre patterns co-occur with text descriptions of musical genres, moods, and instruments.
In every case, the underlying operation is the same: learned statistical associations between human-generated outputs and the conditions that accompanied them. There is no listening, no seeing, no reading in any phenomenological sense. There is pattern matching at extraordinary scale.
Because these systems are pattern-matchers, they excel at tasks that have clear patterns in their training data — genre fiction conventions, standard logo layouts, common chord progressions — and struggle with tasks that require genuine novelty, factual accuracy outside their training window, or understanding of context that exists only in the real world. Knowing this lets you assign tasks intelligently.
In this lab you'll have a conversation with an AI assistant about the mechanics of AI text generation. The goal is not to receive a lecture — it is to test the boundaries of the explanation. Ask it to describe what it does when it generates a creative sentence. Then push back: ask whether it "understands" what it wrote, whether it could be wrong, and how temperature would change the output. Aim for at least 3 exchanges.
In 2022, Cosmopolitan magazine published what it described as the first AI-generated magazine cover, produced using DALL-E 2 in collaboration with digital artist Karen X. Cheng. The cover — depicting an astronaut on the moon — took Cheng about two days of iterative prompting and manual refinement in Photoshop. The image itself was striking; the process was labor-intensive in ways the press coverage largely glossed over. Cheng's own account, published on LinkedIn, described spending hours crafting precise prompts, discarding dozens of unsatisfactory generations, and spending significant time in post-production correcting anatomical errors in the astronaut's glove. The story of that cover is not "AI made a magazine cover." It is: a skilled artist used AI to access a visual style she couldn't have produced as quickly with traditional tools, while providing all the judgment the tool cannot.
A September 2023 study by researchers at MIT and Stanford (published in Science, lead author Shakked Noy) measured the effect of access to ChatGPT on professional writing tasks including cover letters, short reports, and analytical memos. Workers with AI access completed tasks 37% faster and received quality ratings 18% higher from blind evaluators. The gains were concentrated among workers who started at lower performance levels — the AI raised the floor of the distribution more than it raised the ceiling. High performers improved modestly; lower performers improved substantially.
A separate 2023 study from Boston Consulting Group (published in Harvard Business Review, September 2023) tested GPT-4 against consultants on creative and analytical tasks. On tasks within the model's "frontier" — brainstorming, writing, synthesizing — AI-assisted consultants outperformed unassisted consultants by 40% on quality. On tasks outside that frontier — causal reasoning, real-world judgment calls — AI assistance made performance worse, not better. Consultants who trusted the AI uncritically on out-of-frontier tasks performed below the control group.
In the BCG study, "quality" was evaluated by experienced consultants who did not know which documents were AI-assisted. They rated documents on clarity, persuasiveness, and originality of ideas. AI-assisted work scored higher on the first two dimensions. On originality, the gap narrowed. The evaluators noted that AI-assisted work tended toward "competent but predictable" framings on novel problems.
The enthusiasm in the productivity literature collides with a different body of evidence from creative practitioners. In a 2023 survey of 1,000 professional writers conducted by the Authors Guild, 90% said they were concerned about AI's impact on their livelihood; fewer than 10% reported that AI tools had improved the quality of their published work (as distinct from their speed of drafting). The discrepancy between "faster drafting" and "better published work" is significant: speed in early-stage drafting may not translate to quality in the finished artifact that readers encounter.
In music, Suno and Udio (both launched publicly in early 2024) can generate full songs from text prompts in seconds. What they cannot yet do is generate a song that grows from a specific lived experience — the structural quality that distinguishes a Townes Van Zandt lyric from a competent genre exercise. The pattern-matching mechanism produces music that sounds like the genre it was prompted toward. It cannot produce music that sounds like a specific human life.
The honest summary: AI reliably helps with volume, variation, and the mechanics of production. It is unreliable for tasks requiring originality in the deepest sense — work that could only come from a specific person's specific experience.
Use AI for: first drafts, variation generation, style exploration, ideation at scale, mechanical production tasks (transcription, reformatting, summarization). Reserve human judgment for: final selection, voice consistency, factual accuracy, and any work where provenance — who made it, out of what experience — is part of the value.
Screenwriting: In 2023, the WGA strike foregrounded AI's specific threat to TV writers' rooms — particularly the practice of using AI to generate first-draft "outlines" that human writers would then be paid less to polish. The WGA contract reached in September 2023 prohibited studios from requiring writers to use AI and from using AI-generated material as a basis for setting compensation. This is a documented instance of AI affecting not just the work but the economic structure around the work.
Visual design: Adobe introduced Generative Fill in Photoshop in May 2023, allowing designers to extend images, remove objects, or generate background fills using Firefly, Adobe's in-house diffusion model trained on licensed images. Within six months, Adobe reported that Firefly had generated over 3 billion images. Professional designers reported using Generative Fill primarily for comp mockups and background extension — tasks previously requiring either stock photography licensing or manual retouching.
Marketing copy: Jasper AI, founded in 2021, raised $125 million in October 2022 at a $1.5 billion valuation based primarily on enterprise adoption for marketing copy generation. CMOs at companies including Levi's and HubSpot publicly described using AI to generate hundreds of copy variations for A/B testing — a use case where volume and variation matter more than the singular authorial voice that distinguishes literary writing.
This lab asks you to test the BCG study's finding about the "frontier" — the boundary between tasks where AI helps and tasks where it doesn't. Have the AI assist you with a task clearly within its frontier (generating marketing headline variations for a product you describe), then give it a task outside the frontier (explaining why a specific real creative decision was made by a real artist). Compare the quality and confidence of responses.
When Refik Anadol created "Unsupervised" for the Museum of Modern Art in New York — installed in January 2023 — he and his studio spent months engineering the prompts and parameters that would guide the piece's real-time AI generation of visual forms. The work runs on a custom machine learning system trained on MoMA's collection of 200 years of art. Anadol described his process in a February 2023 interview with The New York Times as closer to "conducting an orchestra" than to making images: he was specifying mood, rhythm, and transformation rules, not individual outputs. The installation required the same kind of sustained intentional work that any large-scale artwork requires. The fact that a machine was executing the individual frames did not reduce the directorial labor — it transferred it to the specification stage.
That transfer — from execution to specification — is what makes prompting a genuine creative skill, not a technical shortcut.
Research on prompt engineering converged on a set of components that consistently improve output quality. A 2022 paper from Google DeepMind (Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models") demonstrated that simply asking a model to "think step by step" before answering significantly improved performance on reasoning tasks. This was not a magic phrase — it was an instruction that changes what the model treats as the relevant prior context for each subsequent token.
For creative tasks, the components of an effective prompt are:
Role: What perspective or expertise should the model adopt? ("Write as a crime fiction editor reviewing a first chapter" produces different output than "Write as a crime fiction fan.")
Task: What is the specific output format and scope? ("Write a 200-word pitch paragraph for the following concept" is more useful than "Write about this concept.")
Constraints: What should be excluded or bounded? Constraints are often more valuable than positive specifications because they prevent the model from defaulting to high-probability generic patterns. ("Avoid clichés about isolation. Do not use the word 'journey'." is actionable in a way that "Be original" is not.)
Examples: Providing one or two examples of the target style — called "few-shot prompting" — is the single most reliable technique for aligning output with a specific voice or format. OpenAI's own documentation consistently ranks few-shot prompting as the highest-leverage prompting technique for style replication.
Evaluation criteria: Telling the model how the output will be judged ("The reader is a skeptical editor who has seen 500 pitches this year") shapes the probability distribution toward outputs that would satisfy that evaluator.
In a 2023 experiment documented by writer and educator Ethan Mollick (Wharton School, University of Pennsylvania) in his Substack "One Useful Thing," Mollick found that prompts that specified what NOT to do before specifying the task produced output rated more original by blind reviewers than prompts that only specified what to do. The mechanism: negative constraints push the model away from its highest-probability (most generic) outputs.
One of the most common errors in AI-assisted creative work is treating the first output as the output. In reality, the first generation is a starting position — a rough draft of a draft. The workflow that produces usable creative work almost always involves multiple rounds of prompting, each one informed by what the previous round revealed about the model's tendencies and limitations.
In September 2023, designer and researcher Linus Ekenstam published a detailed account of building an illustrated children's book using Midjourney v5.2 and GPT-4 for text. The project took approximately 80 hours across three weeks. Ekenstam documented 340 distinct image generations, of which 23 were used. The ratio — about 7% — is consistent with reports from other practitioners using diffusion models for professional work. The creative work was almost entirely in the curation and iteration, not in the initial prompts.
This is not a criticism of the tools. It is an accurate description of how creative work with any tool actually functions. A photographer on a professional shoot might take 800 frames and use 12. The ratio is not evidence of inefficiency; it is evidence of the selective judgment that distinguishes professional from amateur work.
The skill shift AI creates is from generating to directing. A skilled prompt engineer is doing what a film director does with a cinematographer — specifying intent, evaluating output, redirecting when the output misses the intent, and maintaining a vision across many iterations. The model is the cinematographer: highly capable at execution within a brief, unable to supply the vision itself.
You'll practice the components of an effective prompt by building one iteratively. Start with a vague prompt for a short piece of creative writing. Then rebuild it with: a specific role, a concrete task, at least one negative constraint, and one example sentence. Compare the outputs. Aim for at least 3 exchanges.
In February 2023, the US Copyright Office issued a partial registration for the graphic novel "Zarya of the Dawn" by Kristina Kashtanova — and simultaneously revoked copyright protection for the AI-generated images within it, retaining protection only for Kashtanova's text and the selection and arrangement of the images. The Office's reasoning was precise: copyright requires human authorship; the images were generated by Midjourney based on Kashtanova's prompts, but the Office found that prompting did not constitute "sufficient creative control" over the specific expressive elements in each image to qualify as authorship. This was not a judgment about whether the work was good. It was a determination about the legal structure of who made what.
That determination established a precedent that has shaped every subsequent copyright discussion in the US about AI-generated creative work: prompting alone does not establish authorship. The human contribution must be specific, intentional, and traceable to particular expressive elements in the output.
As of 2024, the US Copyright Office's position is that works generated entirely by AI are not eligible for copyright protection. Works in which a human makes "sufficient creative contributions" may receive copyright protection for those human-contributed elements. The Office has not defined a clear threshold for what constitutes "sufficient" — it is evaluating cases individually.
The European Union's AI Act (passed March 2024) includes disclosure requirements for AI-generated content but does not directly address copyright ownership. The UK Intellectual Property Office concluded a consultation in 2023 and has not changed its existing position that computer-generated works (without a human author) receive a separate, shorter protection period of 50 years — a position that predates generative AI by decades, from the Copyright, Designs and Patents Act of 1988.
In the training data question — whether AI companies infringed copyright by training on human-created work without license — multiple lawsuits are pending as of this writing. The New York Times filed suit against OpenAI and Microsoft in December 2023, alleging that GPT models were trained on Times articles. Getty Images filed suit against Stability AI in January 2023. These cases have not reached final verdicts and the legal landscape is actively evolving.
If you intend to register copyright for AI-assisted work, document your human creative decisions at each stage. The Copyright Office's guidance suggests that the more you can demonstrate specific creative choices — selecting, arranging, editing, and modifying AI output — the stronger your claim. Work that amounts to "I typed a prompt and used the first result" is unlikely to receive protection.
The legal question of copyright protection is distinct from the ethical question of attribution. Several professional communities have established disclosure norms that go beyond what the law requires. The journal Science announced in January 2023 that AI-generated text cannot be listed as an author and must be disclosed in the methods section. Nature published a similar policy. The Associated Press style guide was updated in 2023 to require disclosure of AI use in news content.
In literary publishing, the situation is less settled. After a flood of AI-generated submissions in early 2023, the science fiction magazine Clarkesworld temporarily closed submissions in February 2023 — editor Neil Clarke documented receiving hundreds of submissions he identified as AI-generated within a short period, overwhelming the slush pile. Many literary magazines updated their submission guidelines to require disclosure or to prohibit AI-generated work. The science fiction professional organization SFWA (Science Fiction and Fantasy Writers of America) updated its membership requirements in 2023 to specify that AI-generated work cannot qualify toward membership eligibility.
The ethical principle underlying most of these policies: audiences, editors, and collaborators are making decisions based on an implicit understanding that the work represents human creative effort. Concealing AI generation violates that understanding regardless of whether it violates a law.
Beyond legal and ethical compliance, there is a pragmatic creative argument for transparency: audiences who know a work is AI-assisted evaluate it on different terms — and that can be advantageous. Refik Anadol's MoMA installation was celebrated partly because its AI process was central to its meaning. Karen X. Cheng's Cosmopolitan cover generated significant coverage because it was openly described as an AI experiment. Concealment forecloses the possibility of the process being part of the work's meaning.
The question "who is the author?" becomes most important when you answer it honestly for yourself, before any legal or ethical obligation compels you. Three questions worth asking before publishing AI-assisted creative work:
1. What did I contribute that the AI could not have contributed without me? Specific lived experience, a particular editorial perspective, a selection judgment rooted in knowledge the model doesn't have — these are human contributions. Rephrasing a prompt until you liked the output is a thin contribution.
2. Would the people encountering this work make different decisions if they knew AI was involved? If yes — if an editor would reconsider, if an audience would feel deceived, if a collaborator would object — that is a signal that disclosure is appropriate regardless of the legal floor.
3. Does the process reflect the values I want my creative work to represent? This is a question about craft identity, not legal compliance. Some practitioners are entirely comfortable with AI as a tool, disclosed or not. Others find that extensive AI generation undermines something they value about making work by hand. Both positions are coherent. What is not coherent is avoiding the question.
In this lab you'll conduct an authorship audit — a structured examination of who contributed what to a creative work. Describe to the AI a hypothetical project: a short piece of writing or an image created with AI assistance. Then work through three questions: What did the human contribute that was specific and traceable? Would the audience feel deceived if they learned AI was involved? What disclosure, if any, would be appropriate? Aim for at least 3 exchanges.