Two creators can use the same AI tool and get opposite outcomes. One receives a registration. The other gets a refusal, or a certificate that can be attacked later. The difference is not the presence of AI. The difference is whether the final track contains enough human expression to be recognized as human-authored.

For copyright AI-generated music, that boundary matters more than genre, platform, or prompt quality.

A prompt directs a machine; authorship fixes an expression.

The D.C. Circuit’s decision in Thaler and the Copyright Office’s recent guidance point to the same rule: AI cannot be the author. A human must make the expressive choices that shape the work. That does not mean every AI-assisted song is unprotectable. It means protection follows the human layer, not the machine layer.

Why the Prompt Is Not the Song

A prompt is creative direction. It can be detailed, iterative, and artistically informed, but it is still a request for output. Telling a model to make a melancholy bedroom-pop track with layered harmonies and a 92 BPM groove is closer to commissioning than composing.

That distinction explains why so many applications fail. The applicant feels creative because the output matches a vision. The law asks a different question: did the human choose the actual expression, or only the idea behind it?

Ideas are not copyrightable. Expression is.

If the machine decides the melody, the chord movement, the phrasing, and the arrangement, the human has supplied a concept, not authorship. If the human wrote the words, sang the hook, structured the verses, and chose among multiple generated stems, the human has contributed protectable expression. The line is not philosophical. It is what the registration examiner is looking for.

Where Human Authorship Actually Appears in Music

In music, human authorship usually shows up in concrete, audible decisions:

  • Original lyrics written by the creator
  • A melody composed by the creator before or alongside AI assistance
  • Vocal performance, including phrasing, timing, tone, and delivery
  • Arrangement choices such as verse order, chorus placement, build, drop, and bridge structure
  • Selection among multiple AI outputs when the creator curates a specific sequence or combination
  • Substantial editing of generated material, such as rewriting harmony, revoicing parts, or replacing sections
  • Creative production decisions that shape the musical expression, not just the loudness or polish

Those are not cosmetic changes. They are the places where the human mind leaves a trace on the finished work.

By contrast, these actions usually do not create a strong copyright claim on their own:

  • Typing prompts
  • Clicking regenerate until one result feels right
  • Changing EQ, compression, or loudness only
  • Making minor tempo or key adjustments
  • Exporting the raw AI output with no expressive revision

That last list matters because many creators confuse technical control with authorship. A polished result is not automatically an authored result. Sounding professional is not the same as being protectable.

The Strongest Claims Are Layered, Not Total

The cleanest AI music registrations are rarely based on the entire song as a single undifferentiated creation. They are based on separable human contributions.

A common example is a songwriter who writes the lyrics and vocal melody, then uses AI for the backing track. The lyrics and melody can be claimed. The instrumental generated by the model should be excluded.

Another example is a producer who generates dozens of instrumental fragments, selects the best sections, rearranges them into a new song structure, and records original vocals over the top. The producer may have a credible claim in the human-authored lyrics, vocal performance, arrangement, and perhaps the original structure. The AI-generated stems remain outside the claim.

A third example is a vocalist who performs over AI accompaniment. The vocal performance is human expression fixed in a sound recording. The backing track may not be. The song can still receive protection, but the claim has to match the human contribution instead of pretending the machine part was authored by a person.

This is the practical rule that keeps registrations alive: claim what you actually created, not what the finished track makes it tempting to claim.

Why Overclaiming Causes Rejection

The U.S. Copyright Office is not looking for your artistic ambition. It is looking for a truthful allocation of authorship.

Overclaiming creates two problems at once. First, it can trigger an examiner to ask for clarification or reject the application because the deposit and the claim do not match. Second, it can create a later vulnerability if a competitor, distributor, or licensee challenges the registration.

The easiest way to damage an otherwise valid filing is to write something broad like entire musical work when parts of the recording were machine generated. That kind of wording collapses the distinction between human and AI output. Once the claim becomes fuzzy, the Office has to ask whether the human actually controlled the expressive elements or merely approved a generated result.

A precise claim does the opposite. It tells the examiner exactly where the human contribution lives. It may be narrower, but it is stronger.

A Simple Test for Whether a Track Can Be Claimed

A useful way to evaluate an AI-assisted track is to ask a few blunt questions:

  1. Could the human contribution be identified without mentioning the AI tool?
  2. Does the human contribution involve expression, not just preference?
  3. Can specific lyrics, notes, takes, edits, or arrangement choices be pointed to?
  4. Would the work still have an original human layer if the machine-generated parts were removed?
  5. Is the final claim limited to what the creator actually fixed in a tangible form?

If the answer to most of those is yes, the work is probably sitting in protectable territory, at least for the human-authored parts. If the answer is no and the only real contribution was prompting, the registration case is weak.

That test is more useful than asking whether AI was used at all. AI usage is not the issue. The issue is whether the human contribution is separable, concrete, and expressive.

Disclosure Is Not an Admission of Weakness

Many creators try to hide AI involvement because they assume disclosure makes the work look less original. In practice, the opposite is usually true.

A clean disclosure helps the Copyright Office evaluate the claim. It signals that the applicant understands the boundary between human authorship and machine generation. It also makes the registration harder to attack later because the record already distinguishes what was claimed from what was excluded.

The strongest application reads like an honest credit sheet, not a marketing pitch. It says, in effect: the human wrote these lyrics, composed this melody, performed this vocal, and arranged these sections. The model generated this instrumental backing. The claim covers the first group and excludes the second.

That level of precision is what turns a fragile filing into a defensible one.

The Real Takeaway

AI does not make copyright impossible. Vagueness does.

The law is still built around authorship that can be traced to a human being. If the human shaped the lyrics, melody, performance, arrangement, or other expressive elements, those parts can often be protected. If the machine supplied the expression and the human only supplied a prompt, the claim is thin.

The safest way to protect AI-assisted music is to make the human layer unmistakable, then describe that layer with precision. Copyright law is not asking creators to avoid AI. It is asking them to identify where the human creativity actually lives.

A track that survives review is one where the author can still be heard, named, and proven.