AI and music: who owns a song that no one wrote?

Sofia takes three months to compose a song. Marco takes 30 seconds with Suno to achieve the same result. But who does that song belong to? While major record la

Sofia is an independent composer. She spent fifteen years studying harmony, endless hours at the piano, sleepless nights to find the right melody. Her latest song took three months: writing, rewriting, arrangement, production. She published it on Spotify. Two hundred listens the first month.

Marco opened Suno. He wrote a prompt: "Melancholic pop ballad, female voice, theme of heartbreak, Billie Eilish style". Thirty seconds later: complete song, professional production, convincingly emotional voice. He uploaded it to Spotify. Five thousand listens the first month.

Marco's song sounds almost identical to Sofia's. Same mood, same structure, same emotional target. But Marco doesn't know how to play any instrument. He hasn't studied music theory. He doesn't know what a modulation to the fourth is. He only wrote a sentence in English.

Who is the author of Marco's song? Him, who wrote the prompt? Suno, which trained the algorithm? The artists whose songs trained the model without consent? The question is not academic. It is at the center of a billion-dollar legal battle that is redefining what it means to create, own, and make a living from music in the age of artificial intelligence.

The Explosion of Algorithmic Music

Musical AI is no longer a lab experiment. It is an active, rapidly expanding industry, with millions of users. Suno AI has generated over 10 million songs in a few months. Udio, a direct competitor, similar numbers. ChatGPT can now generate music. Stable Audio produces custom tracks in seconds.

The technology behind it is sophisticated but conceptually simple: generative models trained on millions of existing songs. The algorithm analyzes melodic, harmonic, timbral, and structural patterns. It learns "how a pop ballad sounds," a blues, a jazz piece. When the user describes what they want, the model generates new audio by synthesizing the learned patterns.

Crucial difference with traditional music synthesis: these systems do not use programmed synthesizers or samplers. They generate audio waveforms directly, imitating style, timbre, even human voice with unsettling realism. The result sounds "human-made" because it was trained on human-made music.

And this is where the problems begin. Those millions of songs used for training? They were not released for this purpose. The artists did not give permission. The record labels did not license the rights. The composers were not compensated.

It's the musical equivalent of photographing all the paintings in a museum, analyzing them with AI, then generating "new paintings in the style of the masters" without ever paying or asking the original authors. Technically possible. Legally controversial. Ethically devastating.

As discussed in the article on AI and Copyright, when the algorithm generates the work, the traditional categories of authorship collapse.

The RIAA vs Suno/Udio Case: The Mother of All Battles

June 2024: the Recording Industry Association of America (RIAA) sues Suno and Udio for massive copyright infringement. The majors (Universal, Sony, Warner) united against musical AI startups. This is not a routine legal skirmish. It is an existential battle for the future of the industry.

The accusations are devastating:

1. Training on millions of protected tracks without a license RIAA claims that Suno and Udio massively downloaded music protected by copyright – including tracks from YouTube via "stream-ripping" – to train the models. Millions of songs, from Taylor Swift to the Beatles, from Drake to Beyoncé, used without permission, without compensation, without even notifying the authors.

Suno and Udio do not deny using existing music for training. They argue it is "fair use" – permissible use for transformative educational purposes. A complex but shaky legal argument when the final product is a commercial product that directly competes with the originals.

2. Generation of output that violates copyright Not only is the training problematic. The outputs themselves violate rights. Tests conducted by RIAA show that with the right prompts, Suno generates tracks almost identical to famous protected songs – same melodies, same harmonic progressions, same recognizable hooks.

Not "in the style of" but "copy of." It's as if the algorithm memorized the most popular songs and regurgitated them with minimal variations. Direct copyright infringement, not an interpretive controversy.

3. Unfair competition with existing catalogs Suno offers a $10/month subscription for unlimited personalized music. Spotify costs $10/month to listen to existing music. But Spotify's music has compensated authors, producers, labels. Suno's music has not.

It is direct competition based on costs zeroed out by systematic rights violation. Economic dumping facilitated by industrial intellectual theft. Hard to compete when your competitor doesn't pay for the raw material you had to purchase.

Suno responded to the court asking to dismiss the "stream-ripping" accusations as a strategic "gambit" by the majors to protect their oligopoly. It argues that labels fear innovation, that copyright law does not apply to automatic training, that AI generation is sufficiently transformative to constitute a new work.

But US courts have begun rejecting the generalized fair use argument: the Thomson Reuters vs Ross case (November 2024) ruled that "using protected content for AI training without permission is NOT fair use." A significant precedent that weakens the legal defense of generative platforms.

German GEMA has also sued Suno and OpenAI for failing to remunerate authors and lyrics used to train models. The legal battle is rapidly globalizing.

As highlighted in the article on AI and Generative Art Ethics, when technology generates works derived from training on uncompensated others' work, the ethical question precedes the legal one.

Who is the Author When the Author is an Algorithm?

But even if we solved the training problem – imagine all platforms paid licenses for the data – the deep philosophical question would remain: who owns the generated song?

US legal analysis addresses the problem: The US Copyright Office requires "human authorship" – only humans can be authors. AI cannot own copyright. But this creates absurd situations:

Scenario 1: The prompt user is the author Marco wrote "melancholic pop ballad." Is that enough for authorship? Did he exercise creativity? Did he make artistic choices? Or did he just give generic instructions that any user could have given?

If a prompt constitutes authorship, then anyone who writes "paint a red sunset" to DALL-E becomes the author of a painting. The creativity threshold dramatically lowers. Copyright becomes automatic, trivial, inflated.

Scenario 2: The company that developed the model is the author Suno programmed the algorithm, collected the data, trained the model. It is significant intellectual work. Perhaps Suno should own the copyright on everything the system generates?

But then Suno owns millions of songs generated by unaware users. Monstrous concentration of intellectual power in the hands of a single tech corporation. Algorithmic creative monopoly.

Scenario 3: The artists who provided training data are co-authors The models do not create from nothing. They synthesize patterns learned from existing music. That music is the work of thousands of human musicians. In a sense, every output is a collective derivative work from all the training inputs.

Should they receive compensation, credit, rights? How to calculate individual contribution when the algorithm has mixed millions of songs? How to distribute royalties when the model has "learned" from the Beatles, Beyoncé, and Sofia the independent composer with 200 listens?

Scenario 4: No one is the author – automatic public domain If there is no clearly identifiable human authorship, the generated work is not covered by copyright. It automatically enters the public domain. Anyone can use, modify, resell it.

Interesting scenario for creative commons but devastating for anyone wanting to monetize. Marco could not protect "his" song from unauthorized commercial uses. But neither can Sofia protect hers – if Marco copies the melody by generating it with Suno, is it technically public domain?

A European thesis addresses intellectual property problems in the music industry: the EU regulatory framework is even more fragmented than the US. Some member states recognize rights for "computer-generated works" to those who made the "necessary arrangements." Others do not. Paralyzing legal uncertainty.

As discussed in the article on AI and Insurance, when algorithms make decisions impacting individual rights, a clear legal framework is needed. In music, this framework still does not exist.

The Ethics of Training: Consent, Compensation, Control

But let's set aside the legal complexity of authorship. Let's return to the more immediate question: is it ethical to train models on others' work without permission or compensation?

Over 10,000 music professionals have signed an appeal against the unlicensed use of works to train generative models. The position is clear: training on copyright without explicit consent is industrial intellectual theft masked as technological innovation.

Main arguments:

1. Violation of the author's moral right Even if legally debatable, using a work of art without informing the author violates the moral right (recognized in many European jurisdictions) to control how one's work is used. Sofia deserves to know that her music is training a direct competitor.

2. Uncompensated economic exploitation AI platforms build multi-billion dollar business models on others' creative work. Suno has raised tens of millions in venture capital investments. It will be valued at hundreds of millions. All based on musical datasets collected without paying a cent to the original authors.

It's like building a real estate empire on expropriated land without compensation. You could technically argue that you "transformed" the land by building buildings. But it's still expropriation.

3. Creation of substitutes that erode the market for originals Musical AI does not create a "new category" that expands the market. It creates direct economic substitutes that compete with human artists for the same dollars from the same consumers.

Marco generates a "melancholic ballad" with Suno instead of listening to Sofia on Spotify. Sofia loses streams, loses royalties, loses algorithmic visibility. AI has not expanded the music market. It has redistributed existing market shares from humans to algorithms.

A global economic study by CISAC documents the impact: generative AI represents an existential risk to authors' and composers' incomes. Catastrophic value imbalance between tech companies and human creatives.

4. Absence of meaningful opt-out Artists cannot easily prevent their music from being included in training datasets. Streaming platforms sell access to data brokers. Once online, music is potentially scrapable. There is no robust technical mechanism to signal "do not use this for AI training."

Even when nominal opt-outs exist, they are hidden, complex, ineffective. The burden is on the artist to protect their rights instead of the burden being on platforms to respect rights by default.

The music industry AI ethics playbook proposes emerging standards: explicit licenses, name/image/voice rights, specific contracts for catalog use in training. Warner, Universal are negotiating selective deals with some AI platforms.

But these are majors with bargaining power. Sofia the independent composer has no leverage to negotiate with Suno. Either she accepts that her music is used (without compensation, without control) or she doesn't publish at all. An impossible choice in the digital age.

As highlighted in the article on AI and Language Teaching, when technology "personalizes" using user data without explicit consent, transparency and individual control are needed.

Proposed Solutions: AI-Royalty Fund and New Rights

If the problem is clear – uncompensated exploitation of human creativity to train algorithmic substitutes – what are possible solutions?

An academic proposal suggests an "AI-royalty fund": a tax on AI platforms based on revenue generated from musical output. Funds distributed proportionally to authors whose works contributed to the training dataset.

A mechanism similar to existing royalty collecting (ASCAP, BMI, SIAE). Calculating individual contribution is impossible precisely but possible approximately via statistical sampling, audio fingerprinting, stylistic similarity analysis.

Advantages:

• Compensates artists for training data use even without precise identification of contributions
• Preserves AI innovation (does not ban, only requires fair compensation)
• Scalable: more AI platforms, more revenue, larger distribution pool
• Precedent: blank media levy to compensate for private copying

Disadvantages:

• Complex distribution bureaucracy
• Risk of capture by majors who already have collecting infrastructure
• Independent artists might receive crumbs
• Does not solve the authorship of output problem (only training)

Analysis proposes new legislative tools: the No AI FRAUD Act (USA) would create specific rights over "voice clones" – synthetic voices that imitate a real artist without consent. Extension of image/name rights to vocal characteristics.

Other countries are exploring:

• Mandatory watermarking for AI-generated music (origin transparency)
• Opt-in instead of opt-out for training datasets (explicit consent necessary)
• Inalienable rights over an artist's style/voice (non-transferable even contractually)
• Public audits of commercial platforms' training datasets

But all proposals face massive resistance from the tech lobby. The argument is always the same: "innovation slowed, American competitiveness compromised, technological future in danger." The same rhetoric used to oppose every tech regulation ever.

As discussed in the article on AI and Disability Inclusive Art, technology can democratize creativity BUT not at the cost of exploiting those who built the cultural foundations on which algorithms are based.

Public Perception: Same Song, Different Value

But even if we solved the legal and ethical problems, a deep cultural question would remain: do people perceive music differently when they know it's generated by AI?

Empirical studies show a disturbing pattern: the same song is rated differently depending on whether it is presented as "composed by a human" or "generated by AI." When people believe it's human, they appreciate it more – they find it more emotional, more authentic, more "real."

But the interesting twist: the effect is dramatically amplified for professional musicians. La Bussola dell'IA · Articoli · Rubriche