Once again, ChatGPT's recommendation (from Part 1) for my songs on YouTube: "I could put together a 'Suno Style Prompting Guide for Shadow Light Project' for you—basically your own personal reference guide."

Using “Oh, my Johnny” (https://www.youtube.com/watch?v=2z4OSzX38xw) as an example, the bot wanted to develop a Version 2 of the song—without, as it put it, losing the basic idea. The goal was then to turn this into a precise “Shadow-Light prompt for Suno V5.5”—essentially a basic concept for my entire song development process within the Shadow Light project.

For anyone trying to establish themselves on the market with their music as a distinct brand—with a recognizable concept that runs through all their productions—this seems like a sensible approach. But as helpful as this may be for commercial success, it also poses a major limitation for the development of an experimental approach to AI music. This echoes the assessment of the AI bot “Gemini,” which critically noted that ChatGPT “very quickly falls into the mode of a methodical/didactic system architect who would prefer to develop an entire toolkit or workflow/framework right away, rather than pragmatically helping a user build a specific track or vibe.”

The System Architecture of ChatGPT

But first, I decided to take ChatGPT up on its offer to develop such a “Shadow Light Prompt” based on the song “Oh My Johnny.” After I sent the lyrics and the song to the AI assistant, I received an evaluation of the song. Of course, the bot couldn’t actually listen to the song; instead, it relied on the data made available to it through technology.

However, a detailed description of the chatbot’s assessment and next steps would be too lengthy in this context, so I’ll limit myself to the most important aspects. These include the general assessment of the song, which allowed ChatGPT to immediately take the lead on my project. It stated:

"My initial strategic assessment: I would currently classify *Oh, My Johnny* as Category B: Good material—but probably not yet fully refined."

As an overall assessment, I received a set of grades.

My Overall Review of “Oh, My Johnny”

Area	My assessment
Basic Idea	8/10
Chorus Hook	7.5/10
History	7/10
Lyrics	5.5/10
linguistic unit	5/10
Celtic/Folk Potential	8.5/10
Potential for a V5.5 Remake	High: 8.5/10

ChatGPT offered a critical comment on the overall concept: “It’s not entirely clear: Is this a modern love story? A tongue-in-cheek folk song? A ballad? A rebellious pub song?”

The bot then offers the following recommendation: More “Irish folk storytelling”: I wouldn’t turn the song into a smooth pop ballad. On the contrary: I’d emphasize its story more. The song could work very well as a modern, sassy Celtic folk-rock song with narrative verses and a big, sing-along chorus.

This criticism is understandable, even though I'm not sure whether more “Irish folk storytelling” would have significantly improved the quality of the song. I was rather skeptical of the suggestions ChatGPT included right away for improving the chorus:

"I would consider making the chorus more rhythmic and collaborative. For example, something like this:

"Oh, my Johnny, oh my Johnny, stay with me
Oh, my Johnny, come and stay with me tonight
Even if the whole world says you don't belong here –
Oh, my Johnny, without you everything would be empty."

That sounds a bit clunky (what exactly would be “empty” here?). As things progressed, ChatGPT then developed its own “Shadow Light” musical style and used it to produce an entirely new version of the song. Here, too, I didn’t find the new version to be an improvement over the original lyrics in every respect (for example, in the last line of the first verse, where Johnny has something to say about everything):

"On the banks of the Rhine, my Johnny and I sat down,
and I played for him on my violin,
the music that resonates deep in my heart. And the melody I played,
echoed throughout the whole city,
my Johnny sat next to me and sang, because he always has something to say."

I wondered to what extent ChatGPT succeeded in living up to its claim of elevating the text to a new, higher level. What struck me as far more serious, however, was that the AI assistant’s prompts severely limited the musical style to a relatively traditional folk framework. Yet this was precisely what I had been trying to overcome in the course of my work on this and similar songs. All in all, I wasn’t convinced by the experiment with this new version, and I would find it difficult to define a coherent concept for a new “Shadow-Light Prompt System” based on this. Songs like “Das gute Leben” (https://youtu.be/O30ByI_-ZFI) would likely fit poorly into this concept.

To what extent does working on the prompt contribute to a song's success?

There is also the broader question of whether such a remake would be capable, on its own, of making its own songs successful in the commercial world of music. Statistically speaking, there is evidence that the number of songs produced daily using AI—without generating any significant response—is steadily increasing. The vast majority of these productions face the problem that they are rarely listened to or played by radio stations. A recent study by the University of Chicago estimates that the vast majority (93%) of AI-generated music is played rarely or not at all and is seldom recommended. AI musicians released large quantities of music across various genres indiscriminately in the hope of landing a hit—but without success (https://doi.org/10.48550/arXiv.2606.18052). Reservations about AI music—such as it being “soulless,” devaluing the work of “real artists,” being a monotonous blend of templates generated by AI models, or constituting a general violation of copyright, etc.—can hardly be dispelled by practical advice on how to improve songs.

I therefore generally doubt whether working on songs can overcome the lack of success of AI music (even if this feeds the hopes of AI songwriters during consulting sessions). While the quality of productions can be improved—and AI can serve as something of a sparring partner in this regard—this has little to do with the commercial success of such products. In Part 4 of this blog post, I’ll explain what’s important for the future of AI music.