1. What is Creative AI?

Why this matters¶

In the autumn of 2022, a sentence typed into a web form could produce a photographic-looking image of a place that does not exist. A few months later, a paragraph could be drafted by a model that had read most of the open web. By 2024 we had short videos from text, songs from text, and code from text. By 2026 these tools are everywhere — in editors, in browsers, in phones — and the question is no longer whether they can produce something but what we should do with them.

This week we ask: what is Creative AI, and why is it now suddenly the subject of an undergraduate course at UiO?

Three slippery words¶

Each of the three words in Creative AI hides a long argument.

Creativity¶

Creativity is one of the oldest contested concepts in the humanities. A common working definition is that something is creative when it is novel and valuable in some context. The cognitive scientist Margaret Boden distinguishes between two senses Boden, 2004:

• P-creativity (“psychological”): novel to the person who produced it.
• H-creativity (“historical”): novel to humanity.

A child’s first drawing of a face is P-creative; the invention of perspective in early Renaissance painting is H-creative.

Boden also distinguishes three kinds of creativity:

• Combinational — putting familiar ideas together in unfamiliar ways (collage, mash-up).
• Exploratory — moving around inside an existing conceptual space and discovering its corners.
• Transformational — changing the conceptual space itself, so that ideas previously impossible become thinkable.

We will return to this triple repeatedly. Generative AI is excellent at the first two and ambiguous about the third.

Artificial intelligence¶

“AI” is even older as a label (the Dartmouth workshop coined the phrase in 1956) and even harder to pin down. A pragmatic working definition is:

Artificial intelligence is the study and construction of systems that perform tasks we would otherwise consider to require human cognition.

That definition is deliberately slippery: as soon as a task becomes routine (chess, OCR, spam filtering, speech-to-text), it tends to lose its “AI” status. Russell and Norvig describe four canonical framings: systems that think like humans, act like humans, think rationally, or act rationally Russell & Norvig, 2021.

For our purposes, the AI in Creative AI almost always refers to machine-learned, often generative, systems based on neural networks trained on large datasets. We unpack each of those words in chapters 2 and 3.

Creative AI¶

Putting the two together, we can define the field of study as follows:

Notice three things about this definition:

1. It is medium-agnostic. We are not building yet another course on “AI for music” or “AI for writing”. The whole point of this generation of models is that the same architectures move between media.
2. It treats AI as a material, not only as a topic. We will use these tools, not just read about them.
3. It includes both the technical and the cultural side. Without the technical side, we cannot evaluate claims about what AI can or cannot do. Without the cultural side, we cannot evaluate what it should or should not be doing.

A short, opinionated history — from Dadaism to diffusion¶

The history of Creative AI is older than it looks, and it has two intertwined strands. One strand is the technology — rule-based programs, neural networks, transformers, diffusion. The other is the art-historical lineage of practices that welcomed chance, machines, systems, and procedures into the studio long before any of this technology existed. Without that second strand, the first looks like it appeared from nowhere in 2022.

The art-historical strand¶

• 1910s — Dadaism. Tristan Tzara writes a poem by drawing words out of a hat. The point is precisely that chance is allowed into the work, and that the artist’s role becomes setting the conditions rather than choosing every word.
• 1950s–60s — Concrete music, serialism, and Fluxus. Composers like Iannis Xenakis use stochastic processes to compose pieces; the Fluxus group treats instructions as the artwork (“Drip Music”, “Composition 1960 #7”).
• 1960s–70s — Generative art. Vera Molnar, Manfred Mohr, and Frieder Nake produce drawings with algorithms and plotters; Sol LeWitt writes “Sentences on Conceptual Art” — “the idea becomes a machine that makes the art”.
• 1990s–2010s — Algorithmic art and creative coding. Casey Reas and Ben Fry release Processing; generative art becomes a stable category and the lineage that today flows directly into AI-augmented creative coding (chapter 8).

The shift to generative AI is therefore not a break with art history — it is the latest entry in a long tradition of artists delegating parts of the work to systems, procedures, and machines. What changes is how much gets delegated, how powerful the systems are, and who owns them.

The technical strand¶

• 1957 — Illiac Suite. Lejaren Hiller and Leonard Isaacson compose what is often cited as the first piece of music generated by a computer, using rules and pseudo-random choices.
• 1973–present — AARON. The artist Harold Cohen develops AARON, a rule-based system that draws and later paints autonomously McCorduck, 1991.
• 1980s — Markov models and expert systems are used in music composition (David Cope), in story generation (TALE-SPIN, MINSTREL), and in design.
• 2014 — Generative Adversarial Networks (GANs). Goodfellow and colleagues introduce a new way to train generators by pitting them against discriminators Goodfellow et al., 2014. This kicks off the first wave of “neural” creative AI.
• 2015 — DeepDream. Google researchers turn an image classifier inside out and produce hallucinated, dog-eyed pictures that go viral. For the first time, a wide public sees what a neural network “thinks”.
• 2017 — Transformer architecture. Vaswani et al. publish “Attention Is All You Need” Vaswani et al., 2017; within a few years it becomes the dominant architecture for language, image, audio, and code models.
• 2020 — Diffusion models mature Ho et al., 2020, and large language models pass the threshold where they become useful for general writing Brown et al., 2020.
• 2022 — The generative turn. Stable Diffusion (open-weight), DALL·E 2, Midjourney, and ChatGPT all land within a few months. Generative AI moves from research labs into the hands of millions of users.
• 2023 onwards — Multimodality and scale. Models now handle text, image, audio, and video together, run on phones, and are integrated into operating systems, browsers, and creative software.

Two patterns are worth pulling out of this list. First, the medium-specific waves are converging: by 2024 the same model family powers writing, drawing, coding, and speaking. Second, public visibility lags research by years: every “sudden” public moment (2015 DeepDream, 2022 ChatGPT) sits on top of a decade of slower academic and industrial work.

Three concepts that thread through the course¶

Out of these two strands, three concepts run through every remaining chapter:

• Intentionality. Why are you making this? A model can produce a thousand variations cheaply. The interesting question is which of them you meant.
• Aesthetic control. How precisely can you steer the system toward the artefact you actually want? Most of the technical content of this course — prompts, conditioning, sampling, editing, reference images, LoRAs, agents — is in service of this single question.
• Ethical authorship. Who is the author when a model trained on millions of other people’s work assists you? What do you owe them, your audience, and yourself in how you describe the work?

We will return to these three words repeatedly. They will also appear, almost verbatim, in your process memos.

What is not Creative AI¶

To sharpen the working definition above, here are a few things this course is not about:

• AI in general. We will not survey self-driving cars, medical diagnosis, or fraud detection.
• The mathematics of deep learning. A few equations help and we will show them, but you do not need to derive backpropagation to pass the course.
• AGI/ASI predictions. Speculation about super-intelligence is a fine pub conversation but not a useful basis for a 12-week practical course in 2026.

We will, however, take seriously:

• Critical perspectives from the humanities and social sciences on what these systems do to labour, copyright, attention, education, and the environment Crawford, 2021Bender et al., 2021O'Neil, 2016.
• Hands-on use of the systems, so that the criticism is grounded in experience.

This week’s lab: Reflect, Explore, Create¶

This first lab gets you over the activation energy of opening accounts, running generations, and writing the first entry in the practice log you will keep all semester.

Reflect (≈ 30 min, in lab + your weekly log)¶

Pick one of the following and write 150–300 words in your weekly log. We will use these answers to seed next week’s discussion.

1. Compare your one-sentence definition of Creative AI (from the start of this chapter) with the working definition we ended up with. What did you leave out? What did you include that we did not?
2. Pick one item from the short history above and look it up. What was the historical context (technical, cultural, economic)? Is there a similar context for the 2022 generative turn?
3. In your discipline, what counted as “creative work” five years ago? Which parts of it are most affected by generative AI today?

Explore (≈ 60 min, in lab)¶

1. Pick one text tool and one image tool from the lists in the introduction. Make a free account if needed.
2. Same brief, two media. Write a single short brief (1–2 sentences) for a small creative task. For example: “A flyer for a student concert at Chateau Neuf featuring a jazz trio.” Use the text tool to draft the flyer text; use the image tool to draft a visual.
3. Vary one thing. Re-run each generation with one parameter changed (a different style, a different tone, a longer or shorter prompt). Save both versions.

Create (≈ 30 min, in lab + carry-over to your portfolio)¶

1. From your variations, assemble one small artefact — a single flyer combining your best image with your best caption, or a two-image diptych with one shared title. Aim for something you would be willing to put on your portfolio page in week 12.
2. Write the first entry of your practice log with these fields:
   • Brief used.
   • Tools used (with version, if shown).
   • Prompts, exact wording.
   • One sentence on what you got, one sentence on what you expected, one sentence on the gap.
   • The artefact (or a link to it).
3. Tell us a story. In the last 20 minutes of class, pair up and present each other’s artefacts to a third student. We will then collect surprises on the board.

Going further¶

• Aaron Hertzmann, “Can Computers Create Art?” Hertzmann, 2018 — short, open-access, the right next thing to read after this chapter.
• Margaret Boden, The Creative Mind: Myths and Mechanisms Boden, 2004 — definitive philosophical treatment.
• Walter Benjamin, The Work of Art in the Age of Mechanical Reproduction Benjamin, 1968 — the historical precedent the current debates keep rhyming with.
• Sol LeWitt, “Paragraphs on Conceptual Art” LeWitt, 1967 — five pages; the conceptual lineage in one sitting.
• Lev Manovich, AI Aesthetics Manovich, 2018 — short and accessible.
• Kate Crawford, Atlas of AI Crawford, 2021 — for the political side.
• Stanford HAI, The AI Index Report Maslej et al., 2024 — annual snapshot of the field.

Open tools you can install this week:

• Ollama Ollama, 2024 — open-weight chat models on your laptop.
• Hugging Face Spaces Hugging Face, 2024 — many free image- and text-model demos in the browser.