The Remix Experiment – Testing Artificial Intelligence
I built and released four remixes using FADR because I wanted to find out for myself how distributors, streaming platforms, copyright systems, playlist curators and listeners would react to music that was created in part with artificial intelligence. The experiment consisted of four individual releases: "Blueberries (WUUYUH Remix)" "Rearview (WUUYUH Remix)" "Working (Bear Cole Remix)" "Forget Me (Turntable Kachina Remix)" All four remix versions have been taken down. The original songs were not part of the artificial intelligence experiment. The four remix versions were.
What we were testing
The vocals and lyrics used in the remixes came from our existing human-made songs. FADR created new musical backgrounds and I edited the rhythms it provided.
Therefore, the remixes were partially created with artificial intelligence; our source performances were maintained, however, the tool was responsible for a significant portion of the new arrangement and composition. How would a distributor respond if they identified reused or copyrighted material? How would streaming services respond if a release was identified as being partially created with artificial intelligence? Would listeners realize? Since many DJs and producers use the same tools, would the material created with artificial intelligence conflict with music that has already been uploaded by someone else?
Easy-to-use - and that was part of the problem
FADR made the process incredibly easy. All you had to do was drop in a song, select a direction for the remix, and it would provide you with multiple options for remixing.
However, as I tried it with different songs, I started hearing the exact same loops repeatedly. It felt less like it was generating completely unique material and more like it was selecting pieces from a large library of pre-existing musical elements in various keys and styles. I cannot demonstrate exactly how FADR generates each result from the outside. I can only provide an example of what I experienced; distinct inputs sometimes produced clearly recognizable loops. This raises an important question - how much is generated specifically for the user and how much is merely selected from a common library?
Distribution of the remixes
When distributing the remixes via LANDR, I indicated that they were partially created with artificial intelligence. I anticipated at least one of the remixes would be flagged. LANDR's review systems appear to be very sensitive to copyrighted works, stock loops, and possible conflicts. I assumed that another producer may have previously released music utilizing some of the same FADR-produced material. None of the four remixes were rejected due to a perceived conflict. This could mean that none of the material has appeared in another scanned release, that LANDR's systems did not perceive any overlap between the two releases, or that releases indicating partial use of artificial intelligence are evaluated differently. This experiment cannot answer which reason is accurate.
However, it does indicate that indicating that a release was "partially A.I." did not create any visible barriers in this case.
Streaming treated the songs normally
Streaming platforms appeared to treat the remixes similarly to other releases. I submitted several remixes to playlists and "Blueberries (WUUYUH Remix)" was picked up by a playlist relatively quickly. From my perspective, indicating that a release contained artificial intelligence felt similar to providing additional metadata during upload rather than classifying the release in a manner that impacted how the music was presented or evaluated. Several listeners noticed and questioned whether I was now incorporating artificial intelligence into my music. I was pleased that they recognized it. Their questioning demonstrated that transparency can be valuable when the information is accessible.
However, a vague partial-A.I. label leaves listeners speculating regarding what the machine actually did.
"Partially made with artificial intelligence" is too broad
A majority of the producers I am acquainted with employ tools that incorporate machine learning or artificial intelligence-assisted processing. Mastering software, intelligent EQ, low-end balancing, compression, noise reduction, vocal clean-up, source separation, and frequency analysis can all employ artificial intelligence. These tools may identify an issue or recommend an initial approach; however, they cannot miraculously complete a record. An experienced ear ultimately determines what the song requires and accepts accountability for producing the desired result. Sound design represents another ambiguous area. A producer may request a tool linked to a synthesizer for a deep growling bass or a lush pad, and subsequently modify that starting point. This process can resemble navigating a massive preset library.
Nevertheless, the musician continues to develop melody lines, perform instruments, program drums, arrange the song, record vocals, and make thousands of creative decisions. This process differs substantially from placing a completed song into a service, requesting a remix, and allowing the service to establish the creative foundation for that remix. In our experiment, artificial intelligence was not simply identifying frequencies or eliminating noise. Artificial intelligence performed a significant amount of the compositional and arrangement work.
Better transparency is required
Distributors and streaming services must categorize fully artificial intelligence-generated music, artificial intelligence-generated vocals, composition, background music or remixes, artificial intelligence-assisted sound design, and artificial intelligence-assisted editing/restoration/mixing/mastering. Categorizing these areas will prevent a producer utilizing an intelligent EQ from being grouped together with someone who typed a prompt and accepted a completed song.
Why I conducted this experiment
I am insatiably curious. I refuse to develop my opinions solely based upon headlines/marketing/or individuals who have never utilized the tools they discuss. I function within technology and employ artificial intelligence in my professional capacity. I recognize how beneficial artificial intelligence can be; however, I also comprehend why we must be truthful concerning its hazards. I desired direct experience with respect to the process/disclosure/checks/platform response/listener response.
Why we removed the remixes
I have spent decades acquiring skills related to producing/recording/DJing/playing instruments/writing songs/engineering audio. The aspect that I love is creating music. Allowing a generator to assume responsibility for the core creative work did not feel like an extension of my practice; rather, it felt like abandoning my practice. That is why "Blueberries (WUUYUH Remix)," "Rearview (WUUYUH Remix)," "Working (Bear Cole Remix)," and "Forget Me (Turntable Kachina Remix)" are no longer accessible. We obtained answers to our inquiries; therefore, retaining them online no longer reflected what 1st Drop Music represents. Our stance is straightforward: we do not employ artificial intelligence to generate our music. We may encounter artificial intelligence-assisted technologies embedded within contemporary production tools; however, we do not entrust the creative core of a song to a generator and present the product as if we developed it.
The artwork became part of the lesson
It seems fitting that the artwork for these remixes were created using artificial intelligence as well. I have employed artificial intelligence-generated imagery in certain segments of the 1st Drop Music initiative; however, I do not assert that employing generated imagery qualifies me as a visual artist or illustrator. The tool created the images. This experiment caused me to pause at this juncture as well. Artificial intelligence enables an abundance of content generation; however, an abundance of content generation is not necessarily meaningful. My conclusion is to utilize artificial intelligence sparingly/disclose it accurately/protect portions of the process where human expression is central to the endeavor. If a generated image facilitates storytelling/generate your next creation independently. Prioritize intent over quantity.
The Trojan horse
The risk associated with artificial intelligence is not one novelty song/one experimental remix. The risk associated with artificial intelligence is permitting an excessive amount of artificially generated material onto platforms where human-produced music becomes increasingly difficult to locate/value/sustain. An audio file can be generated from a prompt. This does not equate to writing lyrics/developing harmonies/rhythms/playing instruments/singing/arranging/producing/revising songs until they convey precisely what the artist intends them to convey.
What remains
Following this project/we published our stance regarding artificial intelligence across the 1st Drop Music web sites. We desire music to exist eternally as human-produced emotion conveyed through sound - not as an oceanic expanse of generated noise. Creating remains the most effective response. If you are a musician/continue creating music. If you are an artist/continue creating art. Learn about tools/evaluate claims/evaluate your needs/be forthright concerning what you employed. Do not relinquish control over the creative endeavors that initially motivated you to become involved in this field. This is precisely what this experiment clarified for me - I do not want artificial intelligence to create music for me - I enjoy creating music myself.