How AI Music Works—and Who Owns the Songs

From Text Prompt to Chart-Ready Track

Type a sentence—"upbeat 1980s synth-pop about driving at midnight"—press generate, and within seconds you have a complete song: lyrics, vocals, melody, drums, and a mixed master. AI music generators like Suno and Udio have made this routine, and their output is sophisticated enough that an AI-assisted track was disqualified from Sweden's official pop charts in early 2026 only after millions of legitimate streams. The technology is no longer a novelty. Understanding how it works—and what legal battles surround it—matters for anyone who listens to music.

The Technology Under the Hood

AI music generation combines two families of machine learning that have reshaped other creative fields: transformer models and diffusion models.

Transformers: Learning Musical Language

Transformers—the same architecture behind large language models like ChatGPT—are trained on vast libraries of audio and text. The model learns statistical relationships: which chord tends to follow which, how a verse structure differs from a chorus, how a particular genre handles rhythm and tempo. When a user enters a text prompt, the transformer converts it into a numerical embedding that guides what kind of musical sequences the model generates next, token by token, much like a language model predicts the next word in a sentence.

Diffusion: Sculpting Sound From Noise

Diffusion models work differently. During training, the system learns to add random noise to real audio recordings step by step until only static remains—then learns to reverse that process and reconstruct clean audio. At generation time, the model starts with pure noise and progressively "denoises" it, guided by the text prompt, until coherent music emerges. Recent architectures such as AudioX, described in a 2026 paper published in Scientific Reports, fuse both approaches into a single Diffusion Transformer (DiT) that handles text, audio, and even video inputs together, enabling richer and more controllable outputs.

Putting It Together

In practice, platforms like Suno compress audio into compact discrete tokens that the transformer can process, then decompress the generated tokens back into audible waveforms. Lyrics are generated separately and their rhythm is matched probabilistically to the melody, while automated mixing balances vocals and instrumentation. The result, as WBUR reported in its profile of Cambridge-based Suno, is a pipeline that can simulate the human songwriting and production process end-to-end in a matter of seconds.

What AI Music Can and Cannot Do

Research from Carnegie Mellon University published in January 2026 found that AI-generated compositions used fewer notes and were rated by listeners as significantly less creative than human-made pieces. AI excels at generating competent, genre-consistent backgrounds for videos, games, and apps, but still struggles with the narrative arc and emotional surprise that defines memorable songwriting. The dominant view in the industry, as CMU's researchers put it, is not "AI replacing artists" but "AI amplifying artists"—handling technical production tasks so humans can focus on creative vision.

The Copyright War

The legal fight over AI music is as significant as the technology itself. In 2024, Universal Music Group, Sony Music Entertainment, and Warner Music Group filed landmark lawsuits against Suno and Udio, alleging that the platforms trained their models on copyrighted recordings without permission or payment. The central legal question—whether training on unlicensed material counts as "transformative" fair use—remains unresolved in court.

By late 2025, the industry began pivoting from pure litigation to negotiated coexistence. Warner Music Group settled with Udio, signing a licensing deal that lets WMG artists opt in to have their work used in Udio's new subscription service. Warner also settled with Suno, requiring the startup to launch entirely new, fully licensed models in 2026. Universal Music Group reached a similar agreement with Udio. Sony, however, has not settled either case, keeping the core copyright questions in play, according to Bloomberg Law.

Meanwhile, streaming platforms are writing their own rules. Bandcamp banned AI-generated music entirely in January 2026. Spotify and others require disclosure of AI content but currently allow it, though Digital Music News reported that policies differ widely and are still evolving.

Why It Matters

AI music generation compresses what once required a studio, a producer, and thousands of dollars into a free web tool. For independent creators, game developers, and advertisers, that is transformative. For session musicians, composers, and vocalists whose livelihoods depend on production work, it represents a direct economic threat. The legal settlements of 2025 suggest the industry is moving toward a licensed, royalty-sharing model—but the terms, and who benefits, are still being negotiated. The answer will shape how music is made, distributed, and paid for for decades to come.