What Makes an AI-Generated Song Valuable?

Value after abundance

In Is Music Trading at a Premium?, we argued that AI changes the economics of music by expanding supply. If technically competent tracks become easier to make, some of the scarcity that supports catalogue value weakens. The next question is more practical. When a new song can be generated in minutes, what makes one AI-generated song worth choosing, licensing, saving or sharing?

The answer is not that AI music has no value. It is that value moves. As competent production becomes more common, the market has to look harder at qualities that are harder to reproduce. Novelty matters. So do cultural specificity, provenance, recognizable identity, rights clarity and genuine demand. A track can sound polished and still be highly substitutable. Another can be imperfect in conventional production terms but valuable because it carries a distinct musical identity that is difficult to generate faithfully at scale.

This distinction matters for AI artists, human musicians, platforms and rights holders because it affects what should be made, promoted and protected. The old shortcut was often sound quality. The new test is harder. It asks whether a song has a reason to exist beyond being another plausible entry in a familiar style.

Why familiar music becomes easier to reproduce

Calling an AI song an “average” of its training data is useful shorthand, but it is not technically precise. Music generation systems do not simply blend existing tracks into a mean song. They learn recurring statistical relationships from the material used to train them, then generate new outputs that are probable under those learned patterns and the user’s prompt or conditioning.

That distinction is important because the scale and composition of training data shape what a model can do well. In a 2024 court filing, Suno said its model had been trained on “tens of millions of recordings” and included essentially all reasonable-quality music files accessible on the open internet, according to the filing reported by Music Business Worldwide. If a system has absorbed huge quantities of pop, house, mainstream electronic music, bossa nova and other heavily represented styles, it should not surprise anyone when it produces convincing versions of those styles.

That is not a moral judgment on the output. It is a valuation problem. When many tracks share similar rhythms, chord movement, song structures, instrumentation and production conventions, the next competent version becomes easier to substitute. A generic house beat may still be useful for a video, a game lobby or a background playlist. But if thousands of similar tracks can be generated cheaply, the marginal value of any one of them is likely to fall unless it carries something more distinctive.

This is where similarity becomes an economic signal. A song that sits close to existing music and close to other AI outputs competes in a crowded pool. The more interchangeable it is, the more its value depends on price, convenience and platform placement rather than artistic identity.

Underrepresented music faces a different equation

The reverse may apply to music that is poorly represented in training data. A 2025 NAACL study cited in the research brief found that only 5.7 percent of the hours in surveyed music-generation datasets came from non-Western genres. It also found that models often imposed Western tonal and rhythmic structures on underrepresented traditions. Fine-tuning improved generation for Hindustani classical and Turkish makam music, but adaptation remained difficult and depended on the model and training process.

That finding should make the industry cautious about claims of universal generation. A model may imitate the surface of a Congolese, Iranian or Central Asian tradition without capturing its tuning systems, rhythmic logic, instrumental practice, performance context or cultural function. To an untrained listener, the result may sound plausible. To a musician or community familiar with the tradition, it may feel flattened or incorrect.

This creates a different kind of scarcity. If a style is hard for current systems to reproduce faithfully, then authentic knowledge, careful curation and culturally informed human involvement become more valuable. The scarcity is not just in the audio file. It is in the musical understanding behind the file.

For Lunar Boom, this is central to the long-term challenge of creating across a wide range of genres. A song in every genre cannot mean a superficial prompt in every genre. It requires attention to what makes a tradition musically and culturally specific, and it requires humility where the available data or model behavior is not enough.

Five tests for value

A useful framework for judging an AI-generated song begins with similarity. If the song closely resembles existing recordings, genre clichés or other AI outputs, it may be easier to replace. Similarity is not automatically a flaw, since listeners often want familiarity, but it limits scarcity.

The second test is genre representation. If the style is heavily present in training data, the model may produce competent results with little intervention. If the style is underrepresented, good results may require specialist prompts, fine-tuning, human arrangement, post-production or collaboration with knowledgeable musicians.

The third test is faithful reproduction. Some genres are not hard because they are obscure. They are hard because they rely on tuning, microtiming, improvisational rules, instrument techniques or social functions that are poorly captured by broad datasets. A song that solves those problems has more defensible value than one that merely adds an ethnic instrument preset to a Western pop frame.

The fourth test is human or cultural specificity. Provenance matters because listeners and licensees increasingly want to know where music came from, who shaped it and what rights are attached to it. This is also a legal and commercial issue. Sony Music filed a second lawsuit against Udio claiming the company copied more than 30,000 sound recordings without permission to train its models, according to Music Business Worldwide. The claims are allegations, but they show why provenance and rights clarity are becoming part of value, not an administrative afterthought.

Suno faces related scrutiny. The Verge reported lawsuits alleging that Suno used copyrighted material from platforms including YouTube, Genius and Deezer, while Suno has argued its use falls under fair use. Until courts and markets settle these questions, some users will place a premium on music with clearer permissions and cleaner origin stories.

The fifth test is demand. Scarcity does not create value by itself. A rare AI-generated track in an underrepresented style still needs listeners, a licensing use case, a community, a strong artist identity or another reason for people to choose it. The market does not reward difficulty alone. It rewards difficulty connected to desire.

What the next market will reward

AI will not make every kind of music equally cheap. It will make some music cheaper first, especially music that models can reproduce within dense and familiar training distributions. In those areas, value shifts toward brand, curation, speed, licensing convenience and audience relationship.

In less represented or more technically specific traditions, the calculation is different. Faithful generation may remain difficult, and that difficulty can preserve scarcity. But the strongest position belongs to music that combines scarcity with demand, cultural care, recognizable identity and credible rights.

For creators and AI artists, the lesson is not to avoid familiar genres. It is to understand when familiarity helps and when it makes a song disposable. For listeners and licensees, the useful question is not whether a track was made with AI. It is whether the track offers something that cannot be replaced by the next plausible generation.