How AI Generates Art and Music, and Why "True Creativity" Remains Disputed

Generative AI tools now produce images, songs and full tracks from a short text prompt, and the technology behind them has become a live legal and cultural fight. In October 2026, music publisher Reservoir filed lawsuits against Suno and Anthropic in the U.S. District Court for the Northern District of California, accusing both companies of training AI systems on thousands of copyrighted compositions without a license. Alongside that fight, researchers continue to debate a separate question: does any of this output qualify as creativity, or is it something else entirely?

How AI-generated images are built

Tools such as Stable Diffusion, Midjourney and DALL-E rely on what are called diffusion models. During training, the system is shown real images while random noise is added to them step by step until the original picture is unrecognizable. The model then learns the reverse process — how to strip that noise away again, according to a technical guide published by Cloudinary. Generation reverses this: the system starts from pure random noise and removes it in stages, with early steps producing broad shapes and composition and later steps refining texture, lighting and detail until a coherent image appears. A text prompt steers each denoising step, telling the model what subject, style and setting to move toward.

How AI composes music from a text prompt

Music systems such as Suno and Udio work differently in detail but similarly in principle. Because raw audio is too dense — millions of numbers per second — to feed directly into a language-style model, these systems first compress sound into a small number of discrete "tokens" using neural audio codecs, achieving roughly 600 times compression, according to a technical overview by AI-TLDR. Once audio is turned into tokens, a transformer model — the same type of architecture behind large language models — predicts the next token, then the next, building a full song piece by piece. A text prompt (and any lyrics) is first expanded into a rough plan covering genre, tempo and song structure before the audio tokens are generated. The models are trained on large music libraries and function, in the description used by that overview, as "composition engines" that absorb chord progressions and song structures rather than retrieving specific existing songs.

The copyright fight behind the technology

Training these systems requires enormous amounts of existing creative work, which is now the subject of active litigation. Reservoir's October 21, 2026 lawsuits allege that Suno used 1,050 copyrighted musical compositions and that Anthropic used lyrics from 853 compositions — including works by Madonna, Johnny Cash and Joni Mitchell — without authorization, according to a summary of the filings reviewed by Law Commentary. Reservoir is seeking roughly $186.4 million in statutory damages from Suno and about $151.4 million from Anthropic, plus claims under the Digital Millennium Copyright Act. Suno has argued that training AI on copyrighted material can qualify as fair use and has separately signed licensing agreements with Warner, BMG and Believe; Anthropic said it intends to defend the case. Reservoir's complaint cites those licensing deals as evidence that a legitimate market for using the music already exists.

Is any of this "creative"? Researchers disagree

Separately from the legal question, scholars disagree on whether the output of these systems meets any reasonable definition of creativity. Computational linguist Emily Bender and colleagues describe large models as "stochastic parrots" that statistically echo patterns in training data rather than originate ideas, according to the Heidelberg Laureate Foundation's review of the debate. Computer vision researcher Alexei Efros draws a distinction between closed, rule-bound domains such as chess or Go — where AI systems can be said to innovate within fixed rules — and open-ended domains such as art, where no fixed target exists.

Other researchers push back from a different angle. AI researcher Selmer Bringsjord argues that current AI output amounts to "remixing" training data rather than genuine creation, and he proposed the Lovelace Test, under which a system would only count as creative if it produced output its own designers could not explain — a bar he says today's models do not clear, according to a summary published by BrainFacts.org. Psychologist Mark Runco adds that intentionality is central to human creativity: people consciously choose to create, while AI systems respond to prompts that reflect the user's intent rather than their own. On testing grounds, the same BrainFacts.org summary notes that some models, including GPT-4, scored higher than the average human on divergent-thinking tests that measure the ability to generate varied ideas — though top human creative performers still outscored the AI systems tested.

Where the debate leaves artists and listeners

The Heidelberg Laureate Foundation's review notes that, given how contested the definition of human creativity already is among researchers, a definitive answer on AI may remain elusive. Several commentators cited in that review frame the more practical path forward as treating AI as a tool that augments human creative work rather than a separate creative agent in its own right — a framing that sits alongside, but does not resolve, the ongoing copyright disputes over how these systems were trained in the first place.