Why I'm Optimistic About AI-Assisted Music Composition

Boo!

At UCF’s recent commencement, the Arts and Humanities graduates made national news when they booed Gloria Caulfield’s statement that artificial intelligence [AI] is “the next industrial revolution.” I had just finished a year teaching composition, orchestration, and private composition lessons there, so their reaction landed close to home.

Rick Beato’s creation of “Sadie Winters” using suno.com was my wake-up call about how good music generative AI had become. I invited lyrics from my pop songwriters and we had fun generating dozens songs in different voices, tempi, instrumentation and subgenres to see if it was really that good, and it was. If you can write strong lyrics, and type in some musical direction or hum a musical idea, the results can be astonishing. They can also be hilariously abused, like screamo on delicate poetry - the AI does what you ask of it - all with great vocal and band performances, arrangements, and mixes. For all of us who started with manuscript paper, tape reels and razor blades, generative A.I. is a miracle. That which required careful years of experience and budget now taking a few seconds is mind blowing - and fun.

After working with several AI generators and tools for composers, I have become interested in and encouraged by their promise. In March, I devoted a class to my UCF composers to bring them up to speed about Suno’s capabilities and limitations. Being informed would ease their fears of career erosion. We could explore its questionable ethics. They did not share my optimism. Far from it. They were largely silent, which, coming from this brilliant, opinionated group, said volumes: we don’t want AI!.

It is = it is

I do not believe the right response from the creative arts is ever denial. Music AI is not alive, sentient, or imaginative. It only sounds like it. When human collaborators tag enough examples of emotional response, the system will learn cause-and-effect relationships to reproduce musical gestures associated with them: tenderness, triumph, grief, intimacy, sincerity, awe—It does not feel those things. It imitates them convincingly. Where a composer judges, music AI correlates. Where a composer evaluates, music AI predicts. Where a human composes, music AI distributes sounds based on probability.

And, Artificial Intelligence has arrived: the result of a decades-long evolving set of techniques using math, coding and probabilty. $1.8 trillion last year alone was invested into its development. Scientists, engineers of all disciplines, business execs, hospitals, and even 84% of high school students are embracing its abilities. It is everywhere and growing. In 2017, “Transformer” architecture was introduced that allowed pattern completion at extraordinary scale. The machine ‘learns’ by calculating relationship values between every identifiable element of a text or media file to build a dense map of what tends to go with what in context of when. That’s just like what a human writer or composer does. AI’s are able to generate language, videos, sounds, and now, music equal to all but expert humans.

Putting Suno Studio to the test

My most useful composition and AI production experiment so far is a short, decide-as-you-go Suno-Studio ($30/mo) collaborative I decided to call “The Seeds of Joy and Peace.” The first minute and forty-four seconds were uploaded at 16bit/48kHz, taken from an archived 2004 film score.

I prompted Suno Studio to generate a sung performance over it. For the text, I used lines from the Metta Sutta—“May all be happy and secure; May all beings be happy at heart”—and chose a Musical Theater style so the words would remain clear over my busy orchestral foundation. The first result was astoundingly good, with an attractive melody using Disney-quality soprano and baritone virtual performances, with zero musical instructions! The production sound was intimately detailed, maybe a little over-hyped, but, emotionally aligned with the lyrics and the changing dynamics of the score. Amazing.

Suno, unfortunately, added things into the track that I did not ask for and would have to remove: a few new piano chords, bass pizzicati - all useable, but, unnecessary. Plus, their rhythms were a bit too ad lib and floaty and could benefit from being closer to the rhythmic grid. I exported the audio, brought it into Logic Pro, used stem separation to isolate the voice, automated its volume and tightened its rhythmic synchronization. A glitch in the word “strong” remained to be the only telltale it was virtual. I could have done the stem separation in Suno, by the way.

For the second half, I chose a Thich Nhat Hanh quotation and generated several outputs of each voice to see what the range of variation that the same prompt might create. Every generation was musical, but had little in common with the first half. The virtual soprano vocalist was different. Some of these generations were so different that their key centers wandered. Also, notice how the word “string” has invaded the lyrics (a typo from the site). Not the fault of suno, but if I was working with real singers, they would have caught it before recording a take - but my virtual singers only followed instructions!

I finally decided to prompt both female and male melodies to match the 1^st Violins so there was some musical connection with the first half. As usual, unasked-for sounds were added: strings, more bass pizzicati, a male vocal “ahh,” even a funny “BUH!” at the entrance.

After cleaning the tracks in LogicPro like before, I used them both, sometimes singing separately and sometimes together. Lastly, I created a final mix and added a final bass pizz. Here it is the final version of my first collaboration with an AI.

Why Suno fails as a composition/ production tool

This experience taught me something important. It is impressive, fast, and often musically persuasive, but the user of Suno Studio becomes less a composer than a curator. You cannot simply change one note or get a precise revision. It is not controllable enough to serve a musician who knows what they want. You often have to roll the dice again. That is a serious limitation because professional work is built on revisions.

That assessment is not a balanced criticism of Suno as a whole because it was not designed to be a composer’s tool. It was created to open a consumer market in AI-assisted song generation entertainment. In that it succeeds.

Ethics and Legality

The ethical problem of AI is enormous. The music creators and rights holders whose work makes these results possible are usually neither credited nor paid: large-scale-robbery? The black-box nature of its training makes attribution impossible. AI Companies have taken advantage of a legal gray area by refusing to reveal exactly what their systems ingested on trade secret grounds. They also claim that they did not infringe on copyrights because they did not illegally copy or distribute their IP in the learning process, just gleaned their statistical relationships. U.S. lawmaking has been too slow to keep up, and current politics, which favor big tech, might be less likely to act in favor of musicians. That will not last. Lawsuits, regulation, and market pressure are already forcing change. For example:

• Lawsuits are forcing disclosure case by case. According to Reuters, recent litigation suggests that training on lawfully obtained material may receive stronger fair-use protection. Pirated use without proper permission opens that company to risk.

• Europe created the EU AI Act, requires that general use AI must publish a summary of the content used for training.

• In June, the Danish government proposed an amendment to its copyright law that gives its people rights to their own image and voice so individuals could shut down deepfakes and digital soundalikes.

• Consumer remix apps like Starstruck will let fans use AI to create their own variations of songs using data from participating artists. The royalties get paid to all rights holders. Cover lets the user choose a new song for an artist to perform that they’ve never recorded. Reimagine keeps the lyrics but rewrites the composition. Remix puts the song in a new genre with the same artist. Create lets the users write their own lyrics and pair them with an artists voice and performance style. The biggest challenge to Starstruck’s success will be that its users will not be able to make copies or share their creations with their friends unless both are within Starstruck’s “walled garden” of membership.

• Spotify, is hinting at a similar AI-remixing model for a new “super-premium” tier. With 290 million listeners who depend on the service for all of their listening - it has a better chance of success than Starstruck because users and their friends may already live in the popular Spotify garden.

I worry that AI generated music could become a real threat to less-established popular artists working inside familiar genres. In my experiments, Suno generations were equal to, and occasionally more interesting musically and sonically than the productions being released by real artists. Virtual artists and AI-generated bands are already entering the marketplace with songs, images, and videos that listeners will not easily distinguish from human acts. For real artists to thrive, they will need to deepen what machines cannot provide: live presence, personal story, community, vulnerability, and emotional exchange with an audience.

For now, AI is strongest where the style boundaries are most statistically reinforced: popular genres. The threat will be less for modern classical concert music, modern jazz, experimental electronic music, and other less commercial fields. These traditions value idiosyncrasy, process, unusual tuning systems, extended forms, irregular rhythms, noise, improvisation, philosophical framing, and a willingness to sound unlike the marketplace. When I asked Suno for a three-minute string quartet in the style of Bartók’s Fourth, it produced a string quartet, but an extremely dull one with essentially no Bartók-like character.

In Closing

Composers have throughout history used tools that extend their imagination: notation, improvisation, recording studios, samplers, synthesizers, algorithmic systems, sequencers, digital audio workstations, and now machine-learning models. Every tool changes how we think. Every tool has temptations and blind spots. Every tool can be used lazily or beautifully.

Though Suno Studio is not aimed at a professional musician or composer, the qualities I expected it to fail at—emotion, sincerity, breath, warmth—were the very qualities it produced most convincingly. That does not make the machine an artist. It means that the traditional producer questions can still be asked of it: What do I want the music feel like? Is it sincere? Does it breathe? Does it carry the message?

I am optimistic about AI-assisted composition because I do not see it as a replacement for musical judgment. I see it as a sketching partner, a mirror, a provocation, a fast generator of alternatives, and sometimes a way of hearing around corners. Its output may require editing, recomposition, orchestration, mixing, coaching, human performance, or complete rejection. But those processes are not outside composition. They are composition.

The question is no longer, “Can AI make music?”

The better question is: Can AI help us become more conscious of what we value in music?