Part 1. What is Artificial Intelligence (AI)?
A Brief Origin of AI and Computer Music
The origin of the term AI can be traced back to the 1956 Dartmouth Summer Research Project on Artificial Intelligence. One of the first pieces of computer-generated music was written shortly after: in 1957, Lejaren Hiller and Leonard Isaacson wrote a string quartet entitled “Illiac Suite” at the University of Illinois Urbana-Champaign using an ILLIAC I computer.
Other notable composers that contributed to the history of computer-generated music include Iannis Xenakis (1922–2001), who used computers to create mathematical models that guided his compositional process, and David Cope (1941–2025), who began developing a computer program that could imitate the style of various composers, called EMI (Experiments in Musical Intelligence), in the 1980s.
Toward a Definition of AI
It is difficult to analyze the impact of AI on music (or anything else, for that matter) without working to understand what AI is and how it works. Reductive understanding risks overlooking the fact that, like all technology, AI is a non-neutral actor. Part of uncovering why that is requires a closer look at how AI works. Consider this definition of AI from IBM:
“Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.”
This definition points to some of the issues with the term “AI” as we use it today, namely that it applies to a vast array of technologies and applications, many of which we do not consider particularly intelligent (for example, auto-focus on iPhone cameras, spellcheckers, and chess software). This is partially because of a phenomenon termed the “AI effect:” from a socio-cultural point of view, when AI is sufficiently effective and effortless, we no longer consider it AI.
In other words, although AI has been part of our lives for decades, we are no longer impressed by the applications we are accustomed to. On the other hand, we can comfortably assign the AI label to applications that are not fully developed or “solved,” such as large language models (LLMs). Consequently, “AI” as an umbrella term can have limited utility.
Under the Hood
AI as we know it today is usually built using a type of machine learning called deep learning. During the training phase of deep learning, models learn to find patterns, or “features,” in their training data. Possible features for music include (but are not limited to) timbre, form, harmonic progression, melodic contour, spatialization, rhythm, and texture.
We are typically referring to Generative AI (GenAI) when we use the term AI today. Generative AI models create new content; for example, large language models output text and music models output audio. Deep learning is also used to build discriminative models, which excel at classification tasks.
Deep learning models are essentially very sophisticated prediction algorithms. When you use a music generation model, the features it learned from its training data form the basis of the predictions it makes about how the audio should be constructed. For instance, if you ask for a modern EDM track, the generated track may have a dramatic drop as part of the song structure. Conversely, if you ask for a modern pop song, it is unlikely to include an extended free jazz saxophone solo in the middle of the song.
To complicate matters further, there are a variety of components that influence the output of models, including:
Training data (including its breadth and makeup, see “Limitations” below)
Model architecture
Reinforcement learning from human feedback (e.g., when a user “thumbs up” or “thumbs down” LLM output)
Settings (temperature, system prompts)
Limitations: Bias and Homogeneity
I find the most useful and effective AI tools in music available now to be some of the least flashy and mostly assistive in application, such as stem separation, audio enhancement, and sample search. I find most text-to-music models to be frustratingly limited in the scope of their output.
Part of the reason is that deep learning, as a form of predictive modeling, prioritizes data that appears frequently. While this is crucial to formulate coherent output, it internalizes biases and omissions within its training data.
There is a lot of music in the world, and many music datasets are biased toward certain genres, geographical regions, tuning systems and tonalities, etc. For example, a 2024 paper by researchers at the Mohamed bin Zayed University of Artificial Intelligence (UAE) analyzed audio datasets and research papers and found that “approximately 86% of the total dataset hours and over 93% of researchers focus primarily on music from the Global North” (“Missing Melodies: AI Music Generation and its “Nearly” Complete Omission of the Global South”).
Additionally, there is concern that widespread use of generative AI could lead to a homogenization of culture (for example, see “Generative AI enhances individual creativity but reduces the collective diversity of novel content”). For in-depth research on the ways that generative AI limits musical possibilities, see “Reductive, Exclusionary, Normalising: The Limits of Generative AI Music.”
Part 1: Wrapping Up
“AI” refers to a diverse range of technologies and applications. When we use the term AI today, we are generally referring to deep learning-driven generative AI. Deep learning relies on statistical analysis of training data, so GenAI models inherit the biases and omissions inherent to their training data.