What this page covers
People worry about three things with AI-generated music: what happened to the rights of the songs used for training, who is responsible if the output resembles an existing song, and whether an AI-made song is your own work at all.
Each is summarized below from public official documents. This is not legal advice. If you need a ruling on your case, please consult a professional.
Concern 1: Training data
Generative AI is built by learning from large amounts of existing data. For music, the dispute is whether that data includes recordings and compositions used without the rights holders' permission.
In Japan, Article 30-4 of the Copyright Act generally allows use for AI development and training as long as the purpose is not to enjoy the expression itself. There is an exception where rights holders' interests would be unreasonably harmed, and what counts is decided case by case.
Elsewhere, the music industry has pushed back. Major labels sued the AI music services Suno and Udio in 2024, and settlements with licensing terms with some of the labels have since been reported. The situation keeps changing, so please check recent reporting.
Concern 2: The output resembles an existing song
Japan's Agency for Cultural Affairs, in its document on AI and copyright (March 2024), says infringement by AI output is judged like human-made works: by similarity and reliance on the existing work.
Similarity means the creative expression is shared. Sharing only ideas or common expression is generally not enough. Reliance means the work was made based on the existing one.
This is where training data matters. The same document indicates that even if the user did not know the existing song, reliance can be found if the AI used it for training. So with an AI whose training data you cannot see, it is hard for a user to confirm safety.
Concern 3: Is an AI-made song your own work?
The US Copyright Office, in its January 2025 report (Copyright and Artificial Intelligence, Part 2: Copyrightability), says output without sufficient human creative contribution is not protected by copyright. Entering prompts alone is generally not enough in its view.
On the other hand, parts where a person used AI as an aid and then selected, edited, performed or arranged can be protected. Treatment differs by country, and in Japan too it is judged case by case.
In short, the more a person adds to the output, the easier it is to claim the result as your own work.
Check your distributor's rules too
Some music distributors set conditions on AI-generated songs, such as requiring human creative input and that training data does not infringe third-party rights. Each service decides approvals and its rules change, so check the latest guidelines before release.
An option if you want no training data
If you want to avoid the training data question, you can choose a tool that does not use a trained model in the first place.
ChordMelo is one. It estimates the key from your chords, combines music theory rules with rhythm patterns to build candidate melodies, scores them, and shows the top three. It has no training data and does not read existing songs to imitate them.
The output is MIDI, a sequence of notes. Choosing sounds, editing, and adding performance or vocals is your work, so the human contribution in Concern 3 comes naturally. Generating, playing and saving MIDI all run in your browser, and your chords are not sent to a server.
Risks that remain
Having no training data does not bring the risk to zero. Because the tool builds melodies from theory rules, common note patterns can end up resembling an existing song by chance.
ChordMelo assumes you will check for similarity to existing songs yourself before releasing or selling, and its terms say the same. We do not claim that using it is always safe.
The concerns, and how ChordMelo compares
| Concern | With trained generative AI | With ChordMelo |
|---|---|---|
| Training data | Hard to know whether unlicensed songs are included | None (rules and scoring) |
| Resembling existing songs | A resemblance to a trained song is hard for users to see | Reads no existing songs. Chance resemblance remains, so check before release |
| Is it your own work? | Output from prompts alone is hard to protect | Output is MIDI. Editing, performing and arranging are yours |
| Reproducibility | The same prompt rarely gives the same result | The same settings and seed always give the same result |
Try melody generation with no training data
Pick a chord progression and press Generate. No sign-up, and you can save the ones you like as MIDI.
Q&A
- Does making a song with generative AI infringe copyright?
- It depends. In the Japanese agency's framework it is judged by similarity and reliance. With an AI whose training data is unknown, a user may find it hard to rule out reliance. If you are worried, consult a professional.
- Is ChordMelo generative AI?
- No. It uses no trained model. It is a program that builds candidates from music theory rules, scores them and picks the best.
- Is a melody made with ChordMelo safe from a copyright standpoint?
- There is no training data issue, but we cannot guarantee safety. A chance resemblance to an existing song remains possible. Please check it yourself before releasing or selling.
- Is this page legal advice?
- No. It is a general summary based on public documents from Japan's Agency for Cultural Affairs and the US Copyright Office, as of November 2026. Laws and lawsuits change. For your own case, consult a professional.