Where Does the Creativity Live?

What Sampling Can—and Cannot—Teach Us About Making Music With AI

Years ago, my composing partner and I were working on a television series with a recurring location that needed its own musical identity. It was the town’s pickup joint and slow-dance bar, a dim lounge where people met, secrets crossed paths and trouble never seemed far away. The music needed to feel smoky and intimate, retro and nostalgic, while also slightly dangerous and seductive.

We found a full-band loop in a licensed collection of Mellotron samples. We slowed it down so dramatically that it became something else entirely. What may once have sounded breezy and fun became eerie, woozy and velvety, as though it were drifting in from another decade. We built around it with haunting saxophone, vibes and other textures that deepened the atmosphere without disturbing its strange, hypnotic quality.

The cue worked so well that we wrote two more pieces in the same musical language for later scenes in that location. The producers loved what the music added to the mystery and intrigue, and the three cues collectively gave the lounge an identity within the series. We did not originate every sound, but we recognized a possibility in existing material, transformed it and gave it a dramatic function it had not previously possessed.

As our careers continued, we largely stopped using distinctive commercial samples unless we planned to alter them so radically that they became unrecognizable. Inevitably, we would hear an unaltered or lightly modified sample from one of our libraries in another composer’s cue, and its familiarity greatly cheapened the piece. The experience taught me that permission to use existing material and the ability to give it an identity are entirely different matters.

Sampling offers a useful precedent for thinking about creativity because it separated the origin of musical material from the creativity exercised upon it. Selection, juxtaposition, transformation and context could turn an existing sound into part of a work with a completely different identity and meaning. The sample did not have to be original to the artist, but the decisions made with it had to matter.

Generative AI complicates that precedent because a sample contains a fixed set of musical decisions. It can be cut, stretched, re-pitched, rearranged or placed in a new setting, but the source itself supplies nothing beyond what was recorded. A generative AI system, on the other hand, responds to direction by making new choices about melody, harmony, rhythm, arrangement, instrumentation, performance and production, often delivering nearly all of them at once. Its responsiveness matters because the person is no longer transforming only fixed material; the system is also determining much of what the new material will be.

AI represents a rupture in how much musical realization a tool can supply, while many of the human judgments exercised around its output remain connected to older practices such as sampling, producing, arranging and editing. The difference is not simply whether human creativity is present, but how creative influence is distributed between the person and the system. That distribution can vary enormously from one work to another, even when both are described with the same phrase: “made with AI.”

I encountered that question more directly while creating a song as a gift for my daughter. She had completed graduate school after living with us for two years and was moving out to begin the next phase of her life with Maple, the dog who had been beside her for the previous two years. The moment sat on two divergent emotional lines: the loss we felt as they left and the possibilities opening in front of them. I wanted the song to be addressed from her to Maple, to sound like the kind of band she loved, and to carry sadness, affection and optimism without allowing any one emotion to overwhelm the others.

The song was private. I did not create it for broadcast, release, monetization or even public sharing on the music platform. Because this was a private gift, I was able to examine creativity in a vacuum without having to settle questions of authorship, ownership, copyright, disclosure, credit and compensation.

I developed the lyrics through an extended exchange with a language model, beginning with the people, events and conflicting emotions the song needed to hold. The model proposed lines, I selected and revised them, and each draft exposed details that were plausible but untrue. One version offered: “Dad gave me some money / And Mom gave me her tears.” Neither of us gave our daughter money, but my wife was more practical while I was more emotional about the event, so I changed it to: “Mom gave me practical advice / And Dad gave me his tears.”

Other lines underwent the same scrutiny. “You helped me get through high school / When the world was feeling blue” became “You helped me after college / When the world was feeling new.” A “tank full of gas” became a “full battery,” and “all those good, sweet years” became “all those rough and tumble / Fun and bittersweet years.” The earlier lines functioned as lyrics, but they were generic and wrong for the person the song was meant for. The model could produce emotionally plausible language; it could not know whether that language was true.

Once the lyrics were finished, I began generating the music. I specified the emotional arc, tempo range, instrumentation, vocal configuration, quiet-to-loud structure and a style connected to a band my daughter loves. Across roughly twenty-five generations, I rejected versions because the tempo felt wrong, the phrasing sounded unnatural, the melodies were predictable or the performance missed the emotional balance I had in mind. The version I chose finally captured the stylistic character, melodic distinctiveness and natural phrasing I had been listening for, after which I made additional changes at the stem level.

Some of the later prompts were nearly identical even though the resulting songs were not. By the twenty-fifth generation, the prompt was not necessarily much better; my understanding of what I was listening for was. The number of attempts does not establish that I composed everything the system generated, but the process was not creatively empty merely because some of its most successful musical decisions emerged unpredictably.

A producer may not play any instrument or originate any musical detail. Their work can still be creative: recognizing what a song needs, rejecting performances that miss it, changing an arrangement, preserving an accident that improves the record or helping an artist find the clearest expression of the work. Those decisions may not make the producer the composer, but they can decisively shape what the finished recording becomes.

Rick Rubin has described his value as a producer in terms of “the confidence that I have in my taste and my ability to express what I feel.” That judgment reflects decades spent listening, producing and helping artists discover the strongest form of their work. His example demonstrates that creative influence can be substantial even when technical execution and instrumental performance belong to others.

The comparison has limits. Musicians understand direction, interpret it intentionally and bring their own conscious creative agency to a recording. A generator does not understand why a daughter leaving home matters, why one line is true to a family and another is false, or why a seemingly capable melody feels generic. It can supply material that expresses those things, but it cannot know what they mean.

I began this essay with an observation about sampling and a suspicion that it might offer a useful precedent for understanding AI-generated music. During an extended conversation with a language model, I was asked for a personal sampling experience, challenged to identify where my decisions had mattered and pressed to distinguish creative contribution from composition. The comparison expanded from sampling to producing, then into questions about candor, professional clearance and the essay’s own construction.

The system also introduced things that were wrong. It attributed a statement to me that I had never made when describing how I presented the song to my daughter. It repeatedly broke connected ideas into short, declarative paragraphs, and I reassembled them in my own syntax. It also proposed a distinction between revising this essay and regenerating music that did not withstand scrutiny, so I rejected it. Each revision depended on judgments about what was accurate, what sounded like me and what genuinely advanced the argument.

I created this essay with AI, just as I created the song with AI, but that description does not imply equal contributions or interchangeable roles. The model asked productive questions, proposed connections, challenged parts of my reasoning and supplied language that sometimes survived into the draft. I supplied the experiences, corrected the record, determined which connections held, rejected language and structures that did not sound like me and decided what the essay ultimately meant.

AI-assisted work covers an enormous range, from accepting the first plausible result to sustaining an extended process of interrogation, redirection and transformation. A music editor receiving a finished cue and its stems cannot tell whether AI replaced one texture or generated nearly the entire recording, much less how deliberately the composer directed and revised the result. In professional production, that uncertainty has practical consequences: studios and networks need clearly documented ownership and licensing rights, along with accurate writer and publisher information for cue sheets. Ambiguity can delay clearance or delivery, complicate errors-and-omissions coverage or force the music to be replaced. Candor is therefore essential not only to weighing the creative contributions, but to determining whether the music can be cleared for use at all.

Any attempt to weigh those contributions depends on an honest account of the process. When AI use is concealed or a person’s role is overstated, the finished work offers no reliable way to reconstruct how influence was divided. Prompts, drafts and generation histories can document parts of the process, but they cannot determine the creative weight of the decisions within it.

In the lounge cue, the sample supplied the foundational sound while we determined its transformation, context and dramatic purpose. In the song, I supplied the lived experience, emotional intention, lyrical judgment and final selection, while the systems contributed language, composition, arrangement and performance. In this essay, I supplied the experiences, standards and final position, while the language model questioned, connected, challenged and drafted. Human creativity was present in each case, but its location and relative weight were different.

No reliable formula may ever turn those contributions into percentages, but the inability to measure them precisely does not make every contribution equivalent. Creative weight should follow consequence: how decisively each contribution shaped the identity, form and meaning of the finished work. “AI-made” names the tool; it cannot tell us where the creativity lived.