does ChatGPT understand music theory?
partly. ChatGPT and other general chatbots can explain most music theory concepts well, and they usually get textbook facts right, such as what a dominant seventh is or what a plagal cadence sounds like. they go wrong more often when the answer depends on the actual notes of your score, because they can't see it, and on careful counting and spelling, which they do loosely.
written . we did not run a scored test of our own. the figures below come from a published study, and chatbots change often, so use this page as a list of things to check before you trust an answer.
- safe ground: definitions, terms and textbook rules.
- check hard: rhythm, note spelling, voice leading and transposed parts.
- best fix: give it the actual notes, then play its answer back.
can ChatGPT read sheet music?
only partly. a chatbot reads sheet music as text you type, as a file such as MusicXML, or as a photo it has to decode first, and each step can drop or change notes without warning. it is far more reliable explaining a concept than reporting exactly which notes are in bar 12.
what a published test found
in a paper posted in March 2025, Liam Pond and Ichiro Fujinaga of McGill University gave three chatbots the questions from the August 2024 Royal Conservatory of Music (RCM) Level 6 theory exam. the models were OpenAI's GPT-4o, Anthropic's Claude 3.5 Sonnet and Google's Gemini 1.5 Pro. the music in each question was given to them as a file in one of four formats (ABC, Humdrum, MEI or MusicXML), and each model answered once with no help and once with written guides and worked examples added to the prompt (Pond and Fujinaga, arXiv 2503.22853).
- with no help, the best total was 52%, from GPT-4o (Pond and Fujinaga). the authors write that "none of the models would have met the passing threshold of 60% without the help of the contextual prompts". the national average for Level 6 exams written in 2024 was 88.0%.
- with the guides added, Claude 3.5 Sonnet did best, at 75% (Pond and Fujinaga).
- every model scored 100.0% on the terminology and history question, in every format, with or without help (Pond and Fujinaga).
- the rests and rhythms question never scored above 10.0% for any model or format (Pond and Fujinaga). in the authors' words, "we were unable to teach the models rests and rhythmic groupings".
- questions on chords "showed limited improvement" when the guides were added.
those models have since been replaced by newer ones, and the authors note that "these results could quickly become out of date". the pattern matches the rest of this page: definitions are safe ground, and counting rhythm is where to check hardest.
where general chatbots tend to slip
they only see what you type
ask "why does bar 12 sound wrong?" and a chatbot has to guess from your summary. some chatbots accept a photo or PDF of a page, but reading notation from an image is its own step, and a misread note or rhythm there flows straight into the answer without any sign that it happened.
note spelling
on a piano, D♯ and E♭ are the same key, but in written music they are different notes, and the key decides which one is correct. the third of a B major chord is spelled D♯. chatbots know the rule, yet they sometimes spell chords in sharp-heavy or flat-heavy keys with the wrong letter name, especially in longer answers.
counting beats
checking whether a bar adds up is careful arithmetic: dotted notes, ties, triplets and compound meters such as 6/8 (six eighth notes, felt as two beats) all have to be counted exactly. a language model writes its answer word by word without doing the sum, so a confident "yes, that bar is correct" deserves a second look.
several voices at once
spotting parallel fifths or octaves means following two or more lines through every chord change. from a text description of a four-part passage, a chatbot can miss a real problem or report one that isn't there.
transposing instruments
a B♭ clarinet or trumpet reads a major second higher than concert pitch, and an E♭ alto sax reads a major sixth higher. the rule is easy to state and easy to apply in the wrong direction, so a B♭ part ends up written a step below concert pitch, when it belongs a step above.
how to check what a chatbot tells you
| kind of question | what the chatbot works from | how to check the answer |
|---|---|---|
| definitions and concepts | what it learned from text | usually reliable; compare with a textbook if it matters |
| chord and scale spelling | rules applied in the answer | one letter name per scale step; no letter repeated or skipped |
| does this bar add up? | your description of the rhythm | count it yourself against the time signature |
| voice-leading errors | the notes you typed, if any | play each pair of voices on its own and listen |
| parts for transposing instruments | the interval rule | check the direction: B♭ and E♭ parts are written higher than concert pitch |
| what's wrong with my piece? | your summary of it | give it the actual notes, then play its suggestion back |
an illustrative example
the exchange below was written for this page to show the spelling problem. it is not a log of a real chat with any chatbot.
- question: what are the notes of a B major chord?
- a mistaken answer of the kind to watch for: B, E♭ and F♯.
- what's wrong: the sound is right, but a B major triad is spelled B, D♯, F♯. the chord is built on every other letter (B, D, F), so the third has to be some kind of D.
getting better answers from any chatbot
- give the key and time signature first.
- type the notes with their spelling, octave and length, bar by bar.
- ask it to show its working one bar at a time, so you can see where an error starts.
- play anything it suggests before you keep it.
a tool that reads your score
the alternative is an assistant that works from the score itself. notei's music-theory assistant reads the score you have open in the editor and answers questions about it in plain language, so it doesn't depend on your description of the notes. it can still be wrong, and the same habit applies: hear its answer with notei's playback before you trust it. on the paid tiers it can also harmonise, arrange, extend or vary a line, and the results land as editable notes. more on the features page.
if a chatbot's transposed part looks suspicious, our free transposer is a quick second opinion.
common questions about chatbots and music theory
can ChatGPT write sheet music?
it can write short passages as text formats such as ABC or MusicXML, which a notation program can then show as a score. the result needs checking, especially the rhythm. our page on whether AI can write sheet music covers what to check.
can a chatbot check my music theory homework?
it can explain the concepts behind a question, which is useful. it is less reliable at marking answers that depend on counting beats or spelling chords, so check those parts yourself against the table above, and ask your teacher whether AI help is allowed.
why do chatbots get rhythm wrong?
a language model writes its answer word by word without adding up the beats in a bar, so it can state that a bar is correct when it is a beat short or long. counting is the step to do yourself.
is notei's assistant always right?
no. it reads the score you have open, so it avoids the guessing a general chatbot does, but it can still be wrong. play its answer back with notei's playback and keep only what sounds right.
sources
- Pond and Fujinaga, Teaching LLMs Music Theory with In-Context Learning and Chain-of-Thought Prompting, arXiv 2503.22853 (2025)
- Open Music Theory: triads (chord spelling)
- Open Music Theory: notating rhythm
related guides: how to write a chord progression, can AI write sheet music?, how to harmonize a melody and the chord identifier.
this guide is part of the how to write sheet music series. notei publishes it, and notei is our product.
written and checked by the notei team. how these guides are written and corrected.