The History and Future of Creative AI

Yesterday, I wrote about the history and future of AI in general. Today, I am continuing my explorations by examining the role of AI in the arts and the impact of the arts on AI. The exploration is based on this notebook, with 60 sources collected by NotebookLM.

AI in the Arts

AI has a multi-generational history in the arts, transitioning from centuries-old mechanical automatons to symbolic rule-making and eventually to deep learning-based approaches. In the following, I will go through some of the works picked out by NotebookLM. Some of them I know well, others I hadn’t heard about before. A complete list of detected works is at the end of the post.

Fine and Visual Arts

Early AI art was pioneered by Harold Cohen, who developed AARON beginning in the late 1960s to codify the act of drawing through symbolic rules. I still remember a weeklong workshop with Harold Cohen while I was studying at Chalmers in Göteborg in the early 2000s. We had good discussions about the possibilities and limitations of rule-based systems and how he had gradually developed his system over decades. Here are examples of how AARON paints:

In the modern era, Generative Adversarial Networks (GANs) have become central, with artists like Refik Anadol creating immersive “data paintings” from massive datasets and Anna Ridler using GANs to explore financial speculation through AI-generated imagery. I hadn’t heard about Ben Snell before, but I love his sculpture Dio, which was first developed by a computer model and then cast using the ground-up remains of the computer that designed it.

Dance

I have been highly inspired by choreographer Merce Cunningham in my dance studies and even tried his LifeForms software back in the days. I hadn’t made the connection to AI before, though, but recall how he used a machine to “sketch” movement ideas that his aging body could no longer create, leading to the discovery of “impossible” movements for his live dancers.

Similarly, Wayne McGregor collaborated with Google and other labs to create generative choreographic tools and a living archive of movement.

Theatre and Acting

I don’t have a good overview of theatre and acting in general, and even less so when it comes to theatre and acting. NotebookLM brought up examples of how AI has been used to generate scripts, such as the surreal short film Sunspring (2016) written by an LSTM model named “Benjamin.”

There are fewer historical references to the use of AI in theatre, but, of course, today, AI supports all parts of the process, from casting through voice analysis, automates complex lighting and sound cues, and even provides “live-scripts” to actors via earbuds in real time.

Music

There are many examples of AI in music, so I am surprised that NotebookLM didn’t find much beyond the most obvious ones. It starts with what is arguably the first computer-composed piece, the Illiac Suite, written in 1957, which used probabilistic rules. Another “classic” is David Cope’s EMI project (Experiments in Musical Intelligence), which successfully mimicked the styles of classical masters like Mozart and Bach.

When it comes to real-time performance, composer–performer George E. Lewis developed the interactive improvisation system Voyager, modeling real-time musical dialogue between human performers and software. That system is primarily rule-based, as far as I know. When it comes to learning-based human–computer interaction, I think Francois Pachet’s Flow Machines are great examples:

More recently, Suno enables the creation of complete music tracks with simple prompts.

Film, TV, and animation

I was surprised that NotebookLM didn’t find many references to old film and animation. Instead, it points to relatively new systems and productions, including Massive, which uses AI agents and “fuzzy logic” to simulate realistic, large-scale battle scenes involving hundreds of thousands of autonomous individuals, most famously in The Lord of the Rings. AI is also used for de-aging actors (e.g., The Irishman) and creating hyper-realistic digital humans and digital actors via motion capture. Tools like Tilt Brush (Google) is a VR painting and sculpting platform used to create immersive 3D artworks.

Again, these are just a few examples highlighted by NotebookLM. Check the complete list below for more.

The impact of art on AI development

While AI has indeed been used to create art over the years, I think it is also interesting to reflect upon how artistic use—and misuse—has fed back into AI development, pushing technical boundaries and advancing theoretical frameworks.

By attempting to encode the artistic process, like in AARON for painting and Voyager for music, artist–researchers gained insights into how human artistic decision-making can be translated into code. Those systems were developed by the artists themselves, though. Other examples, including the development of LifeForms, were influenced by Merce Cunningham’s discussions with the developers.

I need to dig deeper into whether it is true, but NotebookLM claims that the development of “fuzzy logic” was directly inspired by the challenges of creating realistic crowds in films within Massive. This software allowed AI agents to move from binary true/false reactions to a spectrum of “shades of gray,” enabling digital entities to interact in unpredictable, lifelike ways that traditional programming could not.

Story generation remains an exploratory research stage in AI. Creating narratives is easy for humans but difficult for machines. Early attempts, such as James Meehan’s TALE-SPIN, helped researchers identify the “brittleness of symbolic AI.” Bizarre outcomes in these generated tales highlighted that AI required vast, “mundane” real-world knowledge to simulate human-like reasoning.

One of my big arguments for working on creative AI is that artistic projects are excellent “laboratories” for exploring human–computer interaction. You can quickly get a sense of whether something “works” and test it in real-life settings, such as installations or performances.

Key challenges in art and AI

Artistic work on and with AI both creates and exposes numerous challenges, many of which we will address in MishMash. Some key challenges include:

• Authorship: Who is actually responsible for an artwork created (largely) by AI? What are the relationships between the underlying data and algorithms and the human and machine creators? How does an audience perceive these uncertainties?
• Ethical and legal issues: Many of today’s large AI models are trained on copyrighted data without the consent or compensation of the original artists. There are also numerous challenges related to privacy, and questions about ownership of one’s own voice, appearance, and expression.
• Cultural bias and homogenization: Given the skewness in training material, current large models are skewed towards Western, popular cultural expressions. Thus, new artistic outputs will carry this bias. There is also the risk of “aesthetic convergence,” in which AI models reinforce dominant norms in their training data, potentially erasing minority or radical voices.

These challenges are not easy to solve, but it helps to begin by identifying and criticizing them. Focusing on human–machine co-creation also helps. Such a “human-in-the-loop” approach ensures the human remains in control while leveraging AI-based tools to handle repetitive tasks or generate initial “sketches” for further refinement.

In sum, while I think we should remain critical, I also see significant potential in using AI in artistic projects. Ultimately, artistic (ab)use can also help AI in general move forward.

Catalogue of AI art

Here is a list of all the AI-based artworks NotebookLM found for me. This is undoubtedly not complete, and probably biased, but I keep it here for reference.