Can AI Compose Music — And What Is the Musician Still For?

If AI can generate thousands of pieces of music, knowing which one is actually good becomes more important, not less.

Ask an AI system to create an uplifting instrumental, a melancholy piano piece or a dramatic soundtrack, and it may produce something remarkably convincing.

There may be a recognisable melody, a chord progression, percussion, a polished arrangement and even a singer. Services such as Suno offer both songs with vocals and instrumental music generated from written descriptions.

That raises an uncomfortable question for anyone who has spent years learning an instrument:

If a computer can produce all that, what is the musician still for?

My answer begins with another question:

What is this particular piece of music supposed to achieve?

A track can sound impressive and still be completely wrong for the job. It can overwhelm a spoken explanation, give a quiet scene the wrong emotional meaning or arrive at its dramatic climax several seconds after the important moment has passed.

AI can expand the supply of musical possibilities. Choosing, developing and using those possibilities requires judgement.

First, “AI Music” Covers Several Different Activities

There is a substantial difference between asking a chatbot for chord suggestions and asking a music generator to produce a finished recording.

There is also a difference between creating new material and using AI to separate an existing recording into component parts.

A useful distinction is:

• An ideas assistant: suggests themes, chords, lyrics, structures or approaches.

• An audio generator: produces music or sounds that you can hear.

• An analysis or production tool: helps inspect, separate or process recorded material.

These capabilities may overlap, but they are not interchangeable.

A written suggestion for a string arrangement is not a playable orchestral score. A stereo recording is not automatically a collection of editable instrument parts. A confident explanation of a chord progression is not proof that the analysis is correct.

The practical question is therefore: what kind of help do I need, and what will I actually be able to edit afterwards?

Ten Ways AI Can Help a Musician

1. Brainstorming themes: getting beyond the blank page

Sometimes the hardest part is deciding where to begin.

Suppose I am planning a short film about sailing on the Thames. “Write some sailing music” leaves a great deal unexplained.

A more useful starting point might be:

Suggest three musical approaches for a 45-second film showing an early morning river, a boat being rigged and the first moments under sail. Describe the mood, tempo, instrumentation and how the music could develop.

That could produce three directions:

• sparse piano and sustained textures for the quiet river;

• light percussion and a repeating figure for the preparation;

• a broader melody as the boat begins moving.

I might reject two and borrow one small idea from the third. That is still useful progress.

The value may lie in finding a direction, rather than accepting a finished answer.

2. Generating chord ideas: hearing different possibilities

Take a simple four-note phrase:

C – D – E – G

For a first experiment, give each note one beat.

An AI assistant could suggest several harmonisations. I could then play the phrase over C major, A minor or a more unsettled combination of chords and listen to the difference.

The important work happens at the keyboard.

Does the harmony support the phrase? Does it make the music too sentimental? Does a change of bass note create the movement I wanted?

On the Pergamon, this becomes an especially interesting experiment because I can change both the harmony and the registration.

However, I would change one thing at a time initially. Otherwise, I might mistake an attractive sound for a better musical idea.

3. Suggesting arrangements: deciding what happens next

A melody and some chords do not automatically provide an effective arrangement.

AI can suggest ways to introduce, develop and conclude material:

Outline a 30-second instrumental arrangement. Introduce a short motif, repeat it with a change of texture, build gently and finish with a clear ending. Leave space for narration.

That gives me something to test.

Perhaps the suggested percussion arrives too early. Perhaps the arrangement needs fewer layers. Perhaps the opening is already strong enough without an introduction.

An arrangement becomes useful when its development serves the purpose of the piece.

4. Analysing harmony: supporting learning

AI can also act as a discussion partner.

Provide a chord sequence and ask:

Explain the function of each chord in C major. Identify any ambiguity and suggest one alternative ending.

This can help a learner explore why a progression feels settled, unfinished, tense or surprising.

However, the explanation needs checking against the notes and the sound. A general chatbot may produce plausible musical terminology while overlooking an inversion or misunderstanding the key. An audio tool may also misidentify notes in a dense recording.

For teaching, I would ask a student to play the progression, explain it in their own words and compare alternatives.

That turns an answer into an investigation.

5. Creating lyrics: finding something worth singing

AI can suggest subjects, rhymes, verse structures and alternative lines.

That is helpful when the idea exists but the wording refuses to cooperate.

Yet lyrics must work in the mouth as well as on the page. A line may read beautifully but put the wrong syllable on the strongest beat. Another may rhyme neatly while saying very little.

A useful request would be:

Suggest four versions of this line, keeping the meaning and an eight-syllable pattern. Use everyday language and avoid clichés.

Then sing them.

Which words feel natural? Where does the singer breathe? Does the line sound like something this person would actually say?

Specific details often provide the individuality that a generic lyric lacks.

6. Producing reference tracks: making a conversation audible

Words such as “warm”, “cinematic” and “uplifting” can mean different things to different people.

A temporary reference track can help a client explain their preferences:

“The pace is right, but the brass is too grand.”

“I like the texture, but the melody draws attention away from the speaker.”

“That sounds like an adventure film. We are advertising a peaceful afternoon on the river.”

Those responses are useful.

A reference gives everyone something concrete to discuss. It should be clearly identified as a reference while the final musical direction is being agreed.

7. Separating stems: exploring an existing recording

A stem is an audio component or group, such as vocals, drums or bass.

AI separation tools can estimate these components from a mixed recording. Steinberg’s SpectraLayers, for example, provides an Unmix Song process for separating instrumental material into stems.

This can help with listening and study. Reducing the vocal may make an accompaniment easier to hear; isolating the bass may reveal how it supports the harmony.

However, separation does not recover the original studio tracks perfectly. Listen for watery sounds, softened attacks or fragments of other instruments.

It is also a production process, rather than composition in itself.

8. Generating sounds: finding an unfamiliar texture

Sometimes I need a sound rather than a tune.

Perhaps a soft metallic pulse, a slowly changing atmosphere or a brief transition between scenes.

Text-to-audio systems can generate music and sound effects; Stable Audio is one example.

There are two useful approaches here.

One is to generate audio and edit it. The other is to ask how to construct a sound using a synthesiser.

With VCV Rack, for example, a description of oscillators, filters, envelopes and modulation could become a practical patching exercise.

A generated audio file gives me a result to manipulate. Building the patch gives me control over how the sound behaves.

Both can be creatively useful.

9. Helping with orchestration: imagining who plays what

AI might suggest giving a melody to an oboe, placing sustained harmony in strings and adding a quiet horn response.

That can provide a starting point.

The musician then needs to consider range, balance, breathing, articulation and whether the parts make sense for the intended performers.

If I want real people to play the arrangement, I need to check it as a score. If I am producing it with virtual instruments, I still need phrasing and balance.

An orchestral sound library can make an awkward arrangement sound expensive. It cannot make every musical decision convincing.

10. Experimenting with styles: learning what changes

One short motif could become a waltz, a restrained electronic cue, a theatre-organ introduction or a brisk dance arrangement.

AI can suggest changes to metre, rhythm, harmony and texture.

For learning, I would describe musical features:

Keep the melody recognisable. Try triple metre, a gently moving bass and a light accompaniment.

That produces a more informative experiment than asking for a vague imitation.

The useful question is what changed in the music, and why that changed its character.

The Musician Still Decides What the Music Should Communicate

Consider three films containing the same boat:

1. A family enjoying an afternoon sail.

2. A racing crew approaching the start.

3. A documentary about the boat’s history.

The images share a subject, but the music has different work to do.

The family film may need ease and warmth. The racing sequence may need anticipation. The documentary may benefit from restraint, allowing voices and river sounds to carry much of the meaning.

Even “excitement” needs interpretation.

Do we want the audience to feel speed, tension, achievement or delight?

That decision shapes everything which follows.

For me, this is where musical work connects directly with filmmaking: understanding what the audience should notice and feel at each moment.

Rejecting a Good Track Can Be the Right Decision

A track may have an attractive melody and excellent production.

It may also be unsuitable.

Perhaps it competes with narration. Perhaps it sounds too solemn for a welcoming club video. Perhaps its large percussion makes a small business introduction feel absurdly dramatic.

This is why I would separate two questions:

Do I like this music?

Does it serve this project?

The answers do not always match.

If AI makes it easier to produce alternatives, we need a reliable way to choose between them. Otherwise, we can spend an afternoon generating variations without getting closer to a decision.

A clear brief helps us stop.

Phrasing, Timing and Instrumentation Give Music Its Character

A score records notes and durations, but performance involves decisions within them.

A slight hesitation before a phrase. A stronger first note. A softer answer. A bass line that moves forward while the melody seems to float.

AI can generate expressive performances too. The human contribution lies in deciding whether that expression communicates the intended meaning and changing it where necessary.

The same applies to instrumentation.

On the Pergamon, a melody can acquire a different character through registration, balance and articulation. A fuller combination of sounds may be exciting, yet a quieter registration may let the phrase speak more clearly.

That is one reason I remain interested in learning how sounds blend.

Having more sounds available makes listening carefully more valuable.

Matching Music to Pictures: A Practical Example

Imagine a proposed 30-second film of Champagne:

Time	Picture	Musical purpose
0–6 seconds	Quiet river and the boat at rest	Establish space and anticipation
6–14 seconds	Rigging and preparation	Introduce gentle movement
14–24 seconds	The boat sailing	Develop energy and the main idea
24–30 seconds	Closing image and title	Resolve clearly

A generated track might contain all the right ingredients but place them at the wrong times.

The strongest musical moment might occur during a close-up of a rope. The ending might arrive after the title has disappeared.

I could edit the audio, change the arrangement, record a new phrase or adjust the picture.

I might also decide that the opening needs only water, birds and the sound of preparation.

Music does not need to occupy every available second.

Try This: Three Versions, One Brief

A manageable experiment is more useful than generating twenty unrelated tracks.

Start with a brief:

Create music for a 20-second introduction to a welcoming sailing-club film. Convey curiosity and enjoyment. Leave room for one spoken sentence and finish clearly.

Then compare three approaches:

• a short phrase played yourself;

• the phrase developed using AI suggestions;

• an AI-generated audio version made from the brief.

Use the same pictures and keep playback levels reasonably similar. A louder version can seem more impressive simply because it is louder.

Ask:

• Does the emotion fit?

• Can I understand the speech?

• Does the music support the important picture changes?

• Is anything memorable?

• What would I change?

The aim is to discover which decisions improve the result.

Does Learning Music Still Matter?

Yes — because musical understanding gives us more precise ways to respond.

“This is wrong” becomes:

“The accompaniment is too busy.”

“The phrase needs room to breathe.”

“The bass movement weakens the ending.”

“The same motif should return with a different texture.”

You do not need advanced theory to begin. Listening, experimenting and describing what you hear are valuable starting points.

But greater knowledge expands your choices. It also helps you recognise when a tool has produced something plausible rather than something useful.

As a teacher, I would want AI to encourage that exploration. Ask for alternatives, play them, compare them and explain your decision.

AI Changes the Work — and Makes Judgement More Visible

It would be too comfortable to claim that every musical task will remain economically unchanged. Some clients will find a generated track adequate for their needs.

For musicians, that makes it useful to explain the value of work beyond producing an audio file: interpreting a brief, developing a recognisable identity, shaping a performance, responding to revisions and making music fit a particular production.

For client projects, I would also check the chosen tool’s current usage terms, permission for any uploaded material and how AI involvement should be described.

The creative standard remains straightforward:

Can I stand behind the result and explain why it belongs here?

Conclusion: More Music Makes Listening Matter More

AI can help us begin, explore, analyse and produce. It can offer an unexpected direction and make experimentation accessible to people who might otherwise never try composing.

That is exciting.

For my own interests — from the Pergamon and synthesisers to film, teaching and short musical identities — the attraction is the opportunity to explore more possibilities and understand them better.

But possibilities need purpose.

Someone must decide what the music should communicate, which ideas deserve development, how the phrases should move and when silence would do a better job.

The musician can be a performer, composer, editor, teacher and critical listener within that process.

If AI can generate thousands of pieces of music, musical judgement becomes one of the most valuable skills we can develop.

For a film, podcast or business introduction, the starting question is simple: what do you want your audience to feel?