La teoría musical y la IA
¿Vale la pena aprender teoría musical ahora que la IA compone canciones?
Music theory is still worth learning in 2026, but for a different reason than it was in 2019: generative models are now good at theory, and the skill that pays is being able to ask them for the right thing and hear when they got it wrong.
Principiante · 7 secciones
The Short Answer
La respuesta corta
Yes — and the honest reason is not the reassuring one. Music theory is worth learning because AI got good at it, not in spite of that. A model can spell chords that defeat most undergraduates; what it cannot do is decide what a piece should be, notice when its form collapses, hear that the chord chart it handed you is wrong, or make you the legal author of what it produced. Every one of those is a theory skill, and every one of them is now worth more than it was.
The question behind the question is usually about time. Learning harmony takes months, and a text box takes thirty seconds, so the arithmetic looks obvious. It is not, because the two are not alternatives: one produces a finished-sounding track and the other produces the ability to judge it. The moment you want a particular piece of music rather than a plausible one, those stop being interchangeable.
What follows is the case, with the evidence attached. It includes the parts that cut the other way — what has genuinely changed, what is genuinely at risk, and where the old objection to theory is right.
What People Are Actually Afraid Of
The fear is rarely stated as a question about theory. It is stated as a question about whether the effort will turn out to have been pointless — and that deserves a straight answer rather than a pep talk.
AI music is disposable generic soulless trash, even if it is technically correct, in accordance with the rules and conventions of music theory.
— a commenter on Hacker News, arguing against AI music
That is worth quoting because of the concession buried in it. The complaint is not that the machine breaks the rules; it is that following them is not the same as making something worth hearing. Both halves of that are true, and together they are the whole argument of this lesson: theory is what makes output correct, and correctness is a floor rather than a ceiling. You need it to get off the floor, and you need something else to get anywhere — which is precisely why the person who can tell the two apart is not the one being automated.
| The fear | What is actually the case |
|---|---|
| A model can already write a whole song, so learning harmony is wasted time | It can write a song. Getting a particular one requires naming what you want, and the names are theory |
| AI does not really understand music theory, so this will blow over | It largely does, and this is the assumption to drop. See §3 — the interesting gap is between knowing and reasoning |
| The market will be flooded and nobody will hear human music | Uploads flooded; listening did not. The gap between the two is the most useful number on this page (§7) |
| Composing and performing are about to stop being jobs | Employment projections are flat, not collapsing — and they were flat before this too (§7) |
| Anything I make with these tools is mine anyway | Not by default. Prompting alone does not make you an author, and that has a price attached now (§6) |
The Older Objection, Which Was Always Partly Right
Long before any of this, musicians argued that theory is unnecessary, and sometimes harmful — that it is a description of what composers did rather than a set of instructions, that plenty of great records were made by people who could not name a single chord in them, and that learning the rules first can leave you writing to satisfy them. All of that is correct, and an article that pretended otherwise would not be worth trusting on anything else.
What theory actually buys is not permission. It is speed and diagnosis: a name for the sound you already like, so you can ask for it again on purpose; and a reason when something is not working, so that fixing it is not guesswork. That was a convenience when you were the one playing every note. It becomes the whole interface when the notes are coming from somewhere else.
Does AI Actually Know Music Theory?
Mostly yes, and unevenly — which is a more useful answer than either of the ones usually given. The most informative published measurement separates two things that get run together: knowing music theory, and reasoning with it.
MusicTheoryBench is a college-level benchmark of 372 multiple-choice questions, split into 269 on music knowledge and 98 on music reasoning. Tested zero-shot, GPT-4 answered 58.2% of the knowledge questions correctly and 25.6% of the reasoning ones — against a random-guess baseline of 25%. On the reasoning half it was, in other words, at chance. Given chain-of-thought prompting and five worked examples it improved to 69.9% and 39.5%, which its authors note still falls "short of fully saturating the proposed benchmark".
58.2% vs 25.6%
Newer models have moved the knowledge half a long way — chord spellings that reliably catch out second-year students are now routine — and the reasoning half has moved less. So the shape of the thing you are working with is a collaborator with an enormous vocabulary and unreliable judgement about structure: excellent on "what notes are in a ", far shakier on "does this modulation arrive where the lyric needs it to".
And it will not tell you which mode it is in. A model reporting a chord it looked up and a model guessing at a form it never tracked produce prose of identical confidence. There is no tone of voice to listen for. The only way to know which you received is to be able to check — which is the whole of the argument, and the reason the next two sections are about asking and hearing rather than about knowing.
You Cannot Ask For What You Cannot Name
The strongest evidence that theory is still the control surface comes from the companies selling the alternative.
Suno — the tool most often named when someone says theory is finished — publishes a music glossary in its own help centre, which opens "Stuck on a prompt and want to try some musical terms?" and then lists the terms worth trying. Among them: modulation, cadence, ostinato, suspension, anacrusis, syncopation, polyrhythm, counterpoint, dissonance and resolution, arpeggio, augmentation, diminution. It groups them under headings including Melody & Harmony, Song Structure and Advanced Concepts, and gives tempo markings with their BPM ranges — Adagio 66–76, Andante 76–108, Allegro 120–168, Presto 168–200.
That is a vocabulary list for a first-year theory course, published by the vendor as the way to get better results. The prompt box is not an alternative to theory. It is a theory exam with no marks and no feedback.
Why a prompt can never be precise enough on its own
The US Copyright Office, examining whether prompting makes someone an author, quoted a generative music vendor’s own guide on this: prompts cannot fully define an output because "the same text describes an infinite number of possible audio tracks". Adding words narrows the set. It never reduces it to one.
Two chords, twice. The second version changes exactly one note. Play both, then read the table — the word you did not have is the sound you could not have asked for.
One note separates the two rows. The adjective cannot reach it; the name of the scale lands on it directly.
This is the mechanism, and it generalises past prompting. "Dreamy" is a guess about a result. is an instruction about a note. Every theory term you learn converts one adjective you were hoping for into one thing you can specify — and that holds whether the thing carrying out the instruction is a model, a session player, or your own hands.
You Cannot Fix What You Cannot Hear
Asking is the first half. The second is knowing what came back — and this is where the absence of theory costs the most, because the failure is silent. A wrong chord chart does not announce itself. It looks exactly like a right one.
The assistive tools are honest about this when you read their own documentation rather than their marketing. Hooktheory — a music theory company — describes what its own generator hands you as "some neat ideas to work with that are a great jumping off point". Not a finished piece: a starting point, which is only useful to someone who can take it somewhere.
The same caution applies to the analysis tools, and there the errors are specific rather than vague. Chord recognisers routinely simplify what they are unsure of, reporting a plain triad where the recording has a seventh or a suspension. Automatic transcription is bounded in ways that are stated but easy to miss — MuseScore’s audio-to-score feature, in beta, recognises solo piano and acoustic guitar and not an arrangement. Chord labels without harmonic function tell you what was played and nothing about why it worked.
None of that makes the tools bad; all of them are genuinely useful. It makes them instruments that need a reader. A flattened seventh in a chart you accepted becomes a wrong voicing in a song you release.
The first chip is a chord a labeller might report as the second. Play them one after the other. If the difference is obvious to you, you can audit a chord chart; if it is not, that is the skill this section is about.
And Then There Is Everything Below the Chord Symbols
Even a chart that is right is not a piece of music. The same chords, voiced differently, are the difference between something that sounds written and something that sounds assembled — and "fix the voice leading" is the most common note a human gives generated output.
The same two chords twice. In the first, every voice jumps; in the second, one note is held and the others move as little as they can. Nothing about the harmony changed.
| Version | First chord | Second chord | What moved | |
|---|---|---|---|---|
| Every voice jumps | All three voices leap upward; the top voice rises a fifth | |||
| One note held | stays put; the other two move by a step each |
Identical harmony, different voicing. This is the layer a chord symbol does not record and a generated arrangement most often gets wrong.
What Genuinely Changed, and What Did Not
An article that conceded nothing here would not be worth reading, so: plenty has changed, and one widely repeated fear turns out to be measuring the wrong thing.
The flood is real. Deezer reported in July 2026 that fully AI-generated tracks had passed "more than 50% of total uploads of new music" on its platform. The same report says those tracks account for "between 1 and 3% of total streams", and that "up to 85% of the streams generated by fully AI-generated tracks were in fact fraudulent in 2025".
over 50% of uploads, 1–3% of streams
The employment picture is similarly unromantic in both directions. The US Bureau of Labor Statistics projects music directors and composers to grow 1% between 2025 and 2035 — "slower than the average for all occupations" — with about 4,200 openings a year and median pay of $73,710. Musicians and singers are projected at 0%, "little or no change", with about 17,400 openings a year. Roughly 21,600 openings annually, in a field that is not growing. That is a flat market, not a collapsing one, and it was flat before any of this.
US Bureau of Labor Statistics, Music Directors and Composers · US Bureau of Labor Statistics, Musicians and Singers ·
What is genuinely at risk
Work whose value was that it was cheap and unremarkable — generic mood beds, background cues bought by the minute — is being commoditised, and saying otherwise would be dishonest. What is harder to substitute is work where a specific person’s judgement is the product: arranging, music direction, anything needing revision to a brief, and anything that has to be performed in a room.
The tools themselves are also less stable than they look. Following its licensing agreement with Universal Music Group, Udio disabled downloads and stem export — a capability its users had, removed by contract rather than by choice. Building a practice on a single vendor’s feature set is its own risk, separate from AI.
So the honest summary is neither "nothing has changed" nor "it is over". What changed is which part of musicianship is scarce. Producing plausible audio is now abundant and close to free. Deciding what it should be, noticing when it is wrong, and being able to claim it are not — and those are learnable, in that order.
What to Learn First, If You Are Starting Now
In this order, because each one makes the next cheaper. Nothing here takes months before it becomes useful.
- Keys and scales. Which seven notes a piece is drawing on. This is the single highest-leverage thing to know, because it turns every other question into a smaller one.
- Chords and roman numerals. Enough to write a progression down and read one back. This is the notation an expressive input is made of.
- Modes and characteristic notes. The vocabulary that replaces adjectives with instructions — the of §4.
- Ear training, alongside all of the above rather than after it. Naming a seventh on paper and hearing one are different skills, and it is the second that lets you audit what a tool hands back.
- Voice leading and form. The two things generated output most reliably gets wrong, and therefore the two that most reliably mark human revision.
Para seguir
- Name a chord you can hear but cannot yet spell — the first thing a prompt cannot do for you — Identificador de acordes
- Why learn theory at all — the version of this argument that does not mention AI
- Keys: how a piece commits to a pool of notes, and how to tell which
- Roman numerals: writing a progression down so it can be moved, sent and asked for
- Character notes: the one altered degree that gives each mode its sound
- Voice leading: why the same chords can sound written or assembled
- Form: the level at which generated music most often stops making sense
- The shortest case for the first item on that list — Por qué entender las escalas es la clave de cada canción que tocas
Puntos clave
- Music theory is worth learning in 2026 because generative models became good at it, not despite that: the scarce skill moved from producing music to specifying and checking it.
- Models know theory far better than they reason with it, and they report both in the same confident voice — so the only way to tell which you received is to be able to check.
- A prompt cannot name what you have no word for, which is why the company most often said to have made theory obsolete publishes a theory glossary for prompting.
- Generated output fails silently: a chord labeller flattens a seventh it is unsure of, and a wrong chart looks exactly like a right one until someone can hear the difference.
- The US Copyright Office holds that prompting alone does not make you an author, and that what you own is your own expression perceptible in the output — a melody, a chart, an arrangement, an edit.
- AI passed half of all new uploads on at least one streaming service while accounting for 1–3% of actual listening, so the real problem is upload spam and streaming fraud rather than audiences defecting.
Actualizado el 23 de septiembre de 2026
Preguntas
- Do I need music theory to use Suno?
- Not to generate something. To generate something specific, yes — and Suno itself says so: its help centre publishes a glossary of musical terms to use in prompts, covering modulation, cadence, counterpoint and tempo markings. Without those words you get the average of a genre; with them you get what you asked for.
- Does AI understand music theory?
- It knows it much better than it reasons with it. On MusicTheoryBench, GPT-4 scored 58.2% on knowledge questions and 25.6% on reasoning ones, where guessing scores 25%. Newer models are far stronger on the knowledge half. Structural judgement — form, modulation, whether a piece goes anywhere — remains the weak part.
- Will AI replace musicians?
- The projections do not show it. US Bureau of Labor Statistics figures published in 2026 put music directors and composers at 1% growth to 2035 and musicians and singers at 0%, together around 21,600 openings a year. That is a flat market rather than a shrinking one — and it was flat beforehand.
- Can I copyright a song I made with AI?
- Only the human part. The US Copyright Office states that prompts alone do not provide sufficient control for authorship, and that copyright does not extend to purely AI-generated material. What you do own is expression you authored that is perceptible in the output, plus your creative selection, arrangement and modifications.
- Doesn’t learning theory limit creativity?
- Theory describes what music does rather than prescribing what it must do, so it forbids nothing. What it changes is that a sound you liked by accident becomes one you can reproduce on purpose, and a passage that is not working becomes diagnosable instead of mysterious. Rules you cannot name are the ones that constrain you.
- Is it too late to start learning music theory?
- No, and the useful part arrives early. Knowing which scale a song draws on is a first-week idea that immediately makes both its melody and its chords legible. You do not need a complete theoretical education to direct a tool or audit its output — you need enough vocabulary to be specific and enough ear to check.
Apréndelo bien
Capítulo 1: Fundamentos de la Teoría Musical